6502/llvm-6502
1
0
Fork 0
mirror of https://github.com/c64scene-ar/llvm-6502.git synced 2024-12-13 04:30:23 +00:00
Code Issues Projects Releases Wiki Activity
b11d8102cf
llvm-6502/test/CodeGen/AArch64/arm64-big-endian-bitconverts.ll

1102 lines
30 KiB
LLVM
Raw Normal View History

AArch64: remove "arm64_be" support in favour of "aarch64_be". There really is no arm64_be: it was a useful fiction to test big-endian support while both backends existed in parallel, but now the only platform that uses the name (iOS) doesn't have a big-endian variant, let alone one called "arm64_be". git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@213748 91177308-0d34-0410-b5e6-96231b3b80d8
2014-07-23 12:58:11 +00:00
; RUN: llc -mtriple aarch64_be < %s -aarch64-load-store-opt=false -O1 -o - | FileCheck %s
; RUN: llc -mtriple aarch64_be < %s -aarch64-load-store-opt=false -O0 -fast-isel=true -o - | FileCheck %s
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
; CHECK-LABEL: test_i64_f64:
define void @test_i64_f64(double* %p, i64* %q) {
; CHECK: ldr
; CHECK: str
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load double, double* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd double %1, %1
%3 = bitcast double %2 to i64
%4 = add i64 %3, %3
store i64 %4, i64* %q
ret void
}
; CHECK-LABEL: test_i64_v1i64:
define void @test_i64_v1i64(<1 x i64>* %p, i64* %q) {
; CHECK: ldr
; CHECK: str
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <1 x i64>, <1 x i64>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <1 x i64> %1, %1
%3 = bitcast <1 x i64> %2 to i64
%4 = add i64 %3, %3
store i64 %4, i64* %q
ret void
}
; CHECK-LABEL: test_i64_v2f32:
define void @test_i64_v2f32(<2 x float>* %p, i64* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2s }
; CHECK: rev64 v{{[0-9]+}}.2s
; CHECK: str
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <2 x float>, <2 x float>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd <2 x float> %1, %1
%3 = bitcast <2 x float> %2 to i64
%4 = add i64 %3, %3
store i64 %4, i64* %q
ret void
}
; CHECK-LABEL: test_i64_v2i32:
define void @test_i64_v2i32(<2 x i32>* %p, i64* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2s }
; CHECK: rev64 v{{[0-9]+}}.2s
; CHECK: str
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <2 x i32>, <2 x i32>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <2 x i32> %1, %1
%3 = bitcast <2 x i32> %2 to i64
%4 = add i64 %3, %3
store i64 %4, i64* %q
ret void
}
; CHECK-LABEL: test_i64_v4i16:
define void @test_i64_v4i16(<4 x i16>* %p, i64* %q) {
; CHECK: ld1 { v{{[0-9]+}}.4h }
; CHECK: rev64 v{{[0-9]+}}.4h
; CHECK: str
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <4 x i16>, <4 x i16>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <4 x i16> %1, %1
%3 = bitcast <4 x i16> %2 to i64
%4 = add i64 %3, %3
store i64 %4, i64* %q
ret void
}
; CHECK-LABEL: test_i64_v8i8:
define void @test_i64_v8i8(<8 x i8>* %p, i64* %q) {
; CHECK: ld1 { v{{[0-9]+}}.8b }
; CHECK: rev64 v{{[0-9]+}}.8b
; CHECK: str
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <8 x i8>, <8 x i8>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <8 x i8> %1, %1
%3 = bitcast <8 x i8> %2 to i64
%4 = add i64 %3, %3
store i64 %4, i64* %q
ret void
}
; CHECK-LABEL: test_f64_i64:
define void @test_f64_i64(i64* %p, double* %q) {
; CHECK: ldr
; CHECK: str
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load i64, i64* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add i64 %1, %1
%3 = bitcast i64 %2 to double
%4 = fadd double %3, %3
store double %4, double* %q
ret void
}
; CHECK-LABEL: test_f64_v1i64:
define void @test_f64_v1i64(<1 x i64>* %p, double* %q) {
; CHECK: ldr
; CHECK: str
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <1 x i64>, <1 x i64>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <1 x i64> %1, %1
%3 = bitcast <1 x i64> %2 to double
%4 = fadd double %3, %3
store double %4, double* %q
ret void
}
; CHECK-LABEL: test_f64_v2f32:
define void @test_f64_v2f32(<2 x float>* %p, double* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2s }
; CHECK: rev64 v{{[0-9]+}}.2s
; CHECK: str
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <2 x float>, <2 x float>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd <2 x float> %1, %1
%3 = bitcast <2 x float> %2 to double
%4 = fadd double %3, %3
store double %4, double* %q
ret void
}
; CHECK-LABEL: test_f64_v2i32:
define void @test_f64_v2i32(<2 x i32>* %p, double* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2s }
; CHECK: rev64 v{{[0-9]+}}.2s
; CHECK: str
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <2 x i32>, <2 x i32>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <2 x i32> %1, %1
%3 = bitcast <2 x i32> %2 to double
%4 = fadd double %3, %3
store double %4, double* %q
ret void
}
; CHECK-LABEL: test_f64_v4i16:
define void @test_f64_v4i16(<4 x i16>* %p, double* %q) {
; CHECK: ld1 { v{{[0-9]+}}.4h }
; CHECK: rev64 v{{[0-9]+}}.4h
; CHECK: str
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <4 x i16>, <4 x i16>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <4 x i16> %1, %1
%3 = bitcast <4 x i16> %2 to double
%4 = fadd double %3, %3
store double %4, double* %q
ret void
}
; CHECK-LABEL: test_f64_v8i8:
define void @test_f64_v8i8(<8 x i8>* %p, double* %q) {
; CHECK: ld1 { v{{[0-9]+}}.8b }
; CHECK: rev64 v{{[0-9]+}}.8b
; CHECK: str
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <8 x i8>, <8 x i8>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <8 x i8> %1, %1
%3 = bitcast <8 x i8> %2 to double
%4 = fadd double %3, %3
store double %4, double* %q
ret void
}
; CHECK-LABEL: test_v1i64_i64:
define void @test_v1i64_i64(i64* %p, <1 x i64>* %q) {
; CHECK: ldr
; CHECK: str
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load i64, i64* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add i64 %1, %1
%3 = bitcast i64 %2 to <1 x i64>
%4 = add <1 x i64> %3, %3
store <1 x i64> %4, <1 x i64>* %q
ret void
}
; CHECK-LABEL: test_v1i64_f64:
define void @test_v1i64_f64(double* %p, <1 x i64>* %q) {
; CHECK: ldr
; CHECK: str
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load double, double* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd double %1, %1
%3 = bitcast double %2 to <1 x i64>
%4 = add <1 x i64> %3, %3
store <1 x i64> %4, <1 x i64>* %q
ret void
}
; CHECK-LABEL: test_v1i64_v2f32:
define void @test_v1i64_v2f32(<2 x float>* %p, <1 x i64>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2s }
; CHECK: rev64 v{{[0-9]+}}.2s
; CHECK: str
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <2 x float>, <2 x float>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd <2 x float> %1, %1
%3 = bitcast <2 x float> %2 to <1 x i64>
%4 = add <1 x i64> %3, %3
store <1 x i64> %4, <1 x i64>* %q
ret void
}
; CHECK-LABEL: test_v1i64_v2i32:
define void @test_v1i64_v2i32(<2 x i32>* %p, <1 x i64>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2s }
; CHECK: rev64 v{{[0-9]+}}.2s
; CHECK: str
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <2 x i32>, <2 x i32>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <2 x i32> %1, %1
%3 = bitcast <2 x i32> %2 to <1 x i64>
%4 = add <1 x i64> %3, %3
store <1 x i64> %4, <1 x i64>* %q
ret void
}
; CHECK-LABEL: test_v1i64_v4i16:
define void @test_v1i64_v4i16(<4 x i16>* %p, <1 x i64>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.4h }
; CHECK: rev64 v{{[0-9]+}}.4h
; CHECK: str
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <4 x i16>, <4 x i16>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <4 x i16> %1, %1
%3 = bitcast <4 x i16> %2 to <1 x i64>
%4 = add <1 x i64> %3, %3
store <1 x i64> %4, <1 x i64>* %q
ret void
}
; CHECK-LABEL: test_v1i64_v8i8:
define void @test_v1i64_v8i8(<8 x i8>* %p, <1 x i64>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.8b }
; CHECK: rev64 v{{[0-9]+}}.8b
; CHECK: str
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <8 x i8>, <8 x i8>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <8 x i8> %1, %1
%3 = bitcast <8 x i8> %2 to <1 x i64>
%4 = add <1 x i64> %3, %3
store <1 x i64> %4, <1 x i64>* %q
ret void
}
; CHECK-LABEL: test_v2f32_i64:
define void @test_v2f32_i64(i64* %p, <2 x float>* %q) {
; CHECK: ldr
; CHECK: rev64 v{{[0-9]+}}.2s
; CHECK: st1 { v{{[0-9]+}}.2s }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load i64, i64* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add i64 %1, %1
%3 = bitcast i64 %2 to <2 x float>
%4 = fadd <2 x float> %3, %3
store <2 x float> %4, <2 x float>* %q
ret void
}
; CHECK-LABEL: test_v2f32_f64:
define void @test_v2f32_f64(double* %p, <2 x float>* %q) {
; CHECK: ldr
; CHECK: rev64 v{{[0-9]+}}.2s
; CHECK: st1 { v{{[0-9]+}}.2s }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load double, double* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd double %1, %1
%3 = bitcast double %2 to <2 x float>
%4 = fadd <2 x float> %3, %3
store <2 x float> %4, <2 x float>* %q
ret void
}
; CHECK-LABEL: test_v2f32_v1i64:
define void @test_v2f32_v1i64(<1 x i64>* %p, <2 x float>* %q) {
; CHECK: ldr
; CHECK: rev64 v{{[0-9]+}}.2s
; CHECK: st1 { v{{[0-9]+}}.2s }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <1 x i64>, <1 x i64>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <1 x i64> %1, %1
%3 = bitcast <1 x i64> %2 to <2 x float>
%4 = fadd <2 x float> %3, %3
store <2 x float> %4, <2 x float>* %q
ret void
}
; CHECK-LABEL: test_v2f32_v2i32:
define void @test_v2f32_v2i32(<2 x i32>* %p, <2 x float>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2s }
; CHECK: st1 { v{{[0-9]+}}.2s }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <2 x i32>, <2 x i32>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <2 x i32> %1, %1
%3 = bitcast <2 x i32> %2 to <2 x float>
%4 = fadd <2 x float> %3, %3
store <2 x float> %4, <2 x float>* %q
ret void
}
; CHECK-LABEL: test_v2f32_v4i16:
define void @test_v2f32_v4i16(<4 x i16>* %p, <2 x float>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.4h }
; CHECK: rev32 v{{[0-9]+}}.4h
; CHECK: st1 { v{{[0-9]+}}.2s }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <4 x i16>, <4 x i16>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <4 x i16> %1, %1
%3 = bitcast <4 x i16> %2 to <2 x float>
%4 = fadd <2 x float> %3, %3
store <2 x float> %4, <2 x float>* %q
ret void
}
; CHECK-LABEL: test_v2f32_v8i8:
define void @test_v2f32_v8i8(<8 x i8>* %p, <2 x float>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.8b }
; CHECK: rev32 v{{[0-9]+}}.8b
; CHECK: st1 { v{{[0-9]+}}.2s }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <8 x i8>, <8 x i8>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <8 x i8> %1, %1
%3 = bitcast <8 x i8> %2 to <2 x float>
%4 = fadd <2 x float> %3, %3
store <2 x float> %4, <2 x float>* %q
ret void
}
; CHECK-LABEL: test_v2i32_i64:
define void @test_v2i32_i64(i64* %p, <2 x i32>* %q) {
; CHECK: ldr
; CHECK: rev64 v{{[0-9]+}}.2s
; CHECK: st1 { v{{[0-9]+}}.2s }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load i64, i64* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add i64 %1, %1
%3 = bitcast i64 %2 to <2 x i32>
%4 = add <2 x i32> %3, %3
store <2 x i32> %4, <2 x i32>* %q
ret void
}
; CHECK-LABEL: test_v2i32_f64:
define void @test_v2i32_f64(double* %p, <2 x i32>* %q) {
; CHECK: ldr
; CHECK: rev64 v{{[0-9]+}}.2s
; CHECK: st1 { v{{[0-9]+}}.2s }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load double, double* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd double %1, %1
%3 = bitcast double %2 to <2 x i32>
%4 = add <2 x i32> %3, %3
store <2 x i32> %4, <2 x i32>* %q
ret void
}
; CHECK-LABEL: test_v2i32_v1i64:
define void @test_v2i32_v1i64(<1 x i64>* %p, <2 x i32>* %q) {
; CHECK: ldr
; CHECK: rev64 v{{[0-9]+}}.2s
; CHECK: st1 { v{{[0-9]+}}.2s }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <1 x i64>, <1 x i64>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <1 x i64> %1, %1
%3 = bitcast <1 x i64> %2 to <2 x i32>
%4 = add <2 x i32> %3, %3
store <2 x i32> %4, <2 x i32>* %q
ret void
}
; CHECK-LABEL: test_v2i32_v2f32:
define void @test_v2i32_v2f32(<2 x float>* %p, <2 x i32>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2s }
; CHECK: st1 { v{{[0-9]+}}.2s }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <2 x float>, <2 x float>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd <2 x float> %1, %1
%3 = bitcast <2 x float> %2 to <2 x i32>
%4 = add <2 x i32> %3, %3
store <2 x i32> %4, <2 x i32>* %q
ret void
}
; CHECK-LABEL: test_v2i32_v4i16:
define void @test_v2i32_v4i16(<4 x i16>* %p, <2 x i32>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.4h }
; CHECK: rev32 v{{[0-9]+}}.4h
; CHECK: st1 { v{{[0-9]+}}.2s }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <4 x i16>, <4 x i16>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <4 x i16> %1, %1
%3 = bitcast <4 x i16> %2 to <2 x i32>
%4 = add <2 x i32> %3, %3
store <2 x i32> %4, <2 x i32>* %q
ret void
}
; CHECK-LABEL: test_v2i32_v8i8:
define void @test_v2i32_v8i8(<8 x i8>* %p, <2 x i32>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.8b }
; CHECK: rev32 v{{[0-9]+}}.8b
; CHECK: st1 { v{{[0-9]+}}.2s }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <8 x i8>, <8 x i8>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <8 x i8> %1, %1
%3 = bitcast <8 x i8> %2 to <2 x i32>
%4 = add <2 x i32> %3, %3
store <2 x i32> %4, <2 x i32>* %q
ret void
}
; CHECK-LABEL: test_v4i16_i64:
define void @test_v4i16_i64(i64* %p, <4 x i16>* %q) {
; CHECK: ldr
; CHECK: rev64 v{{[0-9]+}}.4h
; CHECK: st1 { v{{[0-9]+}}.4h }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load i64, i64* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add i64 %1, %1
%3 = bitcast i64 %2 to <4 x i16>
%4 = add <4 x i16> %3, %3
store <4 x i16> %4, <4 x i16>* %q
ret void
}
; CHECK-LABEL: test_v4i16_f64:
define void @test_v4i16_f64(double* %p, <4 x i16>* %q) {
; CHECK: ldr
; CHECK: rev64 v{{[0-9]+}}.4h
; CHECK: st1 { v{{[0-9]+}}.4h }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load double, double* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd double %1, %1
%3 = bitcast double %2 to <4 x i16>
%4 = add <4 x i16> %3, %3
store <4 x i16> %4, <4 x i16>* %q
ret void
}
; CHECK-LABEL: test_v4i16_v1i64:
define void @test_v4i16_v1i64(<1 x i64>* %p, <4 x i16>* %q) {
; CHECK: ldr
; CHECK: rev64 v{{[0-9]+}}.4h
; CHECK: st1 { v{{[0-9]+}}.4h }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <1 x i64>, <1 x i64>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <1 x i64> %1, %1
%3 = bitcast <1 x i64> %2 to <4 x i16>
%4 = add <4 x i16> %3, %3
store <4 x i16> %4, <4 x i16>* %q
ret void
}
; CHECK-LABEL: test_v4i16_v2f32:
define void @test_v4i16_v2f32(<2 x float>* %p, <4 x i16>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2s }
; CHECK: rev32 v{{[0-9]+}}.4h
; CHECK: st1 { v{{[0-9]+}}.4h }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <2 x float>, <2 x float>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd <2 x float> %1, %1
%3 = bitcast <2 x float> %2 to <4 x i16>
%4 = add <4 x i16> %3, %3
store <4 x i16> %4, <4 x i16>* %q
ret void
}
; CHECK-LABEL: test_v4i16_v2i32:
define void @test_v4i16_v2i32(<2 x i32>* %p, <4 x i16>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2s }
; CHECK: rev32 v{{[0-9]+}}.4h
; CHECK: st1 { v{{[0-9]+}}.4h }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <2 x i32>, <2 x i32>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <2 x i32> %1, %1
%3 = bitcast <2 x i32> %2 to <4 x i16>
%4 = add <4 x i16> %3, %3
store <4 x i16> %4, <4 x i16>* %q
ret void
}
; CHECK-LABEL: test_v4i16_v8i8:
define void @test_v4i16_v8i8(<8 x i8>* %p, <4 x i16>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.8b }
; CHECK: rev16 v{{[0-9]+}}.8b
; CHECK: st1 { v{{[0-9]+}}.4h }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <8 x i8>, <8 x i8>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <8 x i8> %1, %1
%3 = bitcast <8 x i8> %2 to <4 x i16>
%4 = add <4 x i16> %3, %3
store <4 x i16> %4, <4 x i16>* %q
ret void
}
; CHECK-LABEL: test_v8i8_i64:
define void @test_v8i8_i64(i64* %p, <8 x i8>* %q) {
; CHECK: ldr
; CHECK: rev64 v{{[0-9]+}}.8b
; CHECK: st1 { v{{[0-9]+}}.8b }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load i64, i64* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add i64 %1, %1
%3 = bitcast i64 %2 to <8 x i8>
%4 = add <8 x i8> %3, %3
store <8 x i8> %4, <8 x i8>* %q
ret void
}
; CHECK-LABEL: test_v8i8_f64:
define void @test_v8i8_f64(double* %p, <8 x i8>* %q) {
; CHECK: ldr
; CHECK: rev64 v{{[0-9]+}}.8b
; CHECK: st1 { v{{[0-9]+}}.8b }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load double, double* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd double %1, %1
%3 = bitcast double %2 to <8 x i8>
%4 = add <8 x i8> %3, %3
store <8 x i8> %4, <8 x i8>* %q
ret void
}
; CHECK-LABEL: test_v8i8_v1i64:
define void @test_v8i8_v1i64(<1 x i64>* %p, <8 x i8>* %q) {
; CHECK: ldr
; CHECK: rev64 v{{[0-9]+}}.8b
; CHECK: st1 { v{{[0-9]+}}.8b }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <1 x i64>, <1 x i64>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <1 x i64> %1, %1
%3 = bitcast <1 x i64> %2 to <8 x i8>
%4 = add <8 x i8> %3, %3
store <8 x i8> %4, <8 x i8>* %q
ret void
}
; CHECK-LABEL: test_v8i8_v2f32:
define void @test_v8i8_v2f32(<2 x float>* %p, <8 x i8>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2s }
; CHECK: rev32 v{{[0-9]+}}.8b
; CHECK: st1 { v{{[0-9]+}}.8b }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <2 x float>, <2 x float>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd <2 x float> %1, %1
%3 = bitcast <2 x float> %2 to <8 x i8>
%4 = add <8 x i8> %3, %3
store <8 x i8> %4, <8 x i8>* %q
ret void
}
; CHECK-LABEL: test_v8i8_v2i32:
define void @test_v8i8_v2i32(<2 x i32>* %p, <8 x i8>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2s }
; CHECK: rev32 v{{[0-9]+}}.8b
; CHECK: st1 { v{{[0-9]+}}.8b }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <2 x i32>, <2 x i32>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <2 x i32> %1, %1
%3 = bitcast <2 x i32> %2 to <8 x i8>
%4 = add <8 x i8> %3, %3
store <8 x i8> %4, <8 x i8>* %q
ret void
}
; CHECK-LABEL: test_v8i8_v4i16:
define void @test_v8i8_v4i16(<4 x i16>* %p, <8 x i8>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.4h }
; CHECK: rev16 v{{[0-9]+}}.8b
; CHECK: st1 { v{{[0-9]+}}.8b }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <4 x i16>, <4 x i16>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <4 x i16> %1, %1
%3 = bitcast <4 x i16> %2 to <8 x i8>
%4 = add <8 x i8> %3, %3
store <8 x i8> %4, <8 x i8>* %q
ret void
}
; CHECK-LABEL: test_f128_v2f64:
define void @test_f128_v2f64(<2 x double>* %p, fp128* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2d }
; CHECK: ext
; CHECK: str
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <2 x double>, <2 x double>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd <2 x double> %1, %1
%3 = bitcast <2 x double> %2 to fp128
%4 = fadd fp128 %3, %3
store fp128 %4, fp128* %q
ret void
}
; CHECK-LABEL: test_f128_v2i64:
define void @test_f128_v2i64(<2 x i64>* %p, fp128* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2d }
; CHECK: ext
; CHECK: str
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <2 x i64>, <2 x i64>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <2 x i64> %1, %1
%3 = bitcast <2 x i64> %2 to fp128
%4 = fadd fp128 %3, %3
store fp128 %4, fp128* %q
ret void
}
; CHECK-LABEL: test_f128_v4f32:
define void @test_f128_v4f32(<4 x float>* %p, fp128* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2d }
; CHECK: rev64 v{{[0-9]+}}.4s
; CHECK: rev64 v{{[0-9]+}}.4s
; CHECK: ext
; CHECK: str q
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <4 x float>, <4 x float>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd <4 x float> %1, %1
%3 = bitcast <4 x float> %2 to fp128
%4 = fadd fp128 %3, %3
store fp128 %4, fp128* %q
ret void
}
; CHECK-LABEL: test_f128_v4i32:
define void @test_f128_v4i32(<4 x i32>* %p, fp128* %q) {
; CHECK: ld1 { v{{[0-9]+}}.4s }
; CHECK: rev64 v{{[0-9]+}}.4s
; CHECK: ext
; CHECK: str
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <4 x i32>, <4 x i32>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <4 x i32> %1, %1
%3 = bitcast <4 x i32> %2 to fp128
%4 = fadd fp128 %3, %3
store fp128 %4, fp128* %q
ret void
}
; CHECK-LABEL: test_f128_v8i16:
define void @test_f128_v8i16(<8 x i16>* %p, fp128* %q) {
; CHECK: ld1 { v{{[0-9]+}}.8h }
; CHECK: rev64 v{{[0-9]+}}.8h
; CHECK: ext
; CHECK: str
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <8 x i16>, <8 x i16>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <8 x i16> %1, %1
%3 = bitcast <8 x i16> %2 to fp128
%4 = fadd fp128 %3, %3
store fp128 %4, fp128* %q
ret void
}
; CHECK-LABEL: test_f128_v16i8:
define void @test_f128_v16i8(<16 x i8>* %p, fp128* %q) {
; CHECK: ld1 { v{{[0-9]+}}.16b }
; CHECK: ext
; CHECK: str q
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <16 x i8>, <16 x i8>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <16 x i8> %1, %1
%3 = bitcast <16 x i8> %2 to fp128
%4 = fadd fp128 %3, %3
store fp128 %4, fp128* %q
ret void
}
; CHECK-LABEL: test_v2f64_f128:
define void @test_v2f64_f128(fp128* %p, <2 x double>* %q) {
; CHECK: ldr
; CHECK: ext
; CHECK: st1 { v{{[0-9]+}}.2d }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load fp128, fp128* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd fp128 %1, %1
%3 = bitcast fp128 %2 to <2 x double>
%4 = fadd <2 x double> %3, %3
store <2 x double> %4, <2 x double>* %q
ret void
}
; CHECK-LABEL: test_v2f64_v2i64:
define void @test_v2f64_v2i64(<2 x i64>* %p, <2 x double>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2d }
; CHECK: st1 { v{{[0-9]+}}.2d }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <2 x i64>, <2 x i64>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <2 x i64> %1, %1
%3 = bitcast <2 x i64> %2 to <2 x double>
%4 = fadd <2 x double> %3, %3
store <2 x double> %4, <2 x double>* %q
ret void
}
; CHECK-LABEL: test_v2f64_v4f32:
define void @test_v2f64_v4f32(<4 x float>* %p, <2 x double>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2d }
; CHECK: rev64 v{{[0-9]+}}.4s
; CHECK: rev64 v{{[0-9]+}}.4s
; CHECK: st1 { v{{[0-9]+}}.2d }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <4 x float>, <4 x float>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd <4 x float> %1, %1
%3 = bitcast <4 x float> %2 to <2 x double>
%4 = fadd <2 x double> %3, %3
store <2 x double> %4, <2 x double>* %q
ret void
}
; CHECK-LABEL: test_v2f64_v4i32:
define void @test_v2f64_v4i32(<4 x i32>* %p, <2 x double>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.4s }
; CHECK: rev64 v{{[0-9]+}}.4s
; CHECK: st1 { v{{[0-9]+}}.2d }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <4 x i32>, <4 x i32>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <4 x i32> %1, %1
%3 = bitcast <4 x i32> %2 to <2 x double>
%4 = fadd <2 x double> %3, %3
store <2 x double> %4, <2 x double>* %q
ret void
}
; CHECK-LABEL: test_v2f64_v8i16:
define void @test_v2f64_v8i16(<8 x i16>* %p, <2 x double>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.8h }
; CHECK: rev64 v{{[0-9]+}}.8h
; CHECK: st1 { v{{[0-9]+}}.2d }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <8 x i16>, <8 x i16>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <8 x i16> %1, %1
%3 = bitcast <8 x i16> %2 to <2 x double>
%4 = fadd <2 x double> %3, %3
store <2 x double> %4, <2 x double>* %q
ret void
}
; CHECK-LABEL: test_v2f64_v16i8:
define void @test_v2f64_v16i8(<16 x i8>* %p, <2 x double>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.16b }
; CHECK: rev64 v{{[0-9]+}}.16b
; CHECK: st1 { v{{[0-9]+}}.2d }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <16 x i8>, <16 x i8>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <16 x i8> %1, %1
%3 = bitcast <16 x i8> %2 to <2 x double>
%4 = fadd <2 x double> %3, %3
store <2 x double> %4, <2 x double>* %q
ret void
}
; CHECK-LABEL: test_v2i64_f128:
define void @test_v2i64_f128(fp128* %p, <2 x i64>* %q) {
; CHECK: ldr
; CHECK: ext
; CHECK: st1 { v{{[0-9]+}}.2d }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load fp128, fp128* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd fp128 %1, %1
%3 = bitcast fp128 %2 to <2 x i64>
%4 = add <2 x i64> %3, %3
store <2 x i64> %4, <2 x i64>* %q
ret void
}
; CHECK-LABEL: test_v2i64_v2f64:
define void @test_v2i64_v2f64(<2 x double>* %p, <2 x i64>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2d }
; CHECK: st1 { v{{[0-9]+}}.2d }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <2 x double>, <2 x double>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd <2 x double> %1, %1
%3 = bitcast <2 x double> %2 to <2 x i64>
%4 = add <2 x i64> %3, %3
store <2 x i64> %4, <2 x i64>* %q
ret void
}
; CHECK-LABEL: test_v2i64_v4f32:
define void @test_v2i64_v4f32(<4 x float>* %p, <2 x i64>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2d }
; CHECK: rev64 v{{[0-9]+}}.4s
; CHECK: rev64 v{{[0-9]+}}.4s
; CHECK: st1 { v{{[0-9]+}}.2d }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <4 x float>, <4 x float>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd <4 x float> %1, %1
%3 = bitcast <4 x float> %2 to <2 x i64>
%4 = add <2 x i64> %3, %3
store <2 x i64> %4, <2 x i64>* %q
ret void
}
; CHECK-LABEL: test_v2i64_v4i32:
define void @test_v2i64_v4i32(<4 x i32>* %p, <2 x i64>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.4s }
; CHECK: rev64 v{{[0-9]+}}.4s
; CHECK: st1 { v{{[0-9]+}}.2d }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <4 x i32>, <4 x i32>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <4 x i32> %1, %1
%3 = bitcast <4 x i32> %2 to <2 x i64>
%4 = add <2 x i64> %3, %3
store <2 x i64> %4, <2 x i64>* %q
ret void
}
; CHECK-LABEL: test_v2i64_v8i16:
define void @test_v2i64_v8i16(<8 x i16>* %p, <2 x i64>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.8h }
; CHECK: rev64 v{{[0-9]+}}.8h
; CHECK: st1 { v{{[0-9]+}}.2d }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <8 x i16>, <8 x i16>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <8 x i16> %1, %1
%3 = bitcast <8 x i16> %2 to <2 x i64>
%4 = add <2 x i64> %3, %3
store <2 x i64> %4, <2 x i64>* %q
ret void
}
; CHECK-LABEL: test_v2i64_v16i8:
define void @test_v2i64_v16i8(<16 x i8>* %p, <2 x i64>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.16b }
; CHECK: rev64 v{{[0-9]+}}.16b
; CHECK: st1 { v{{[0-9]+}}.2d }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <16 x i8>, <16 x i8>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <16 x i8> %1, %1
%3 = bitcast <16 x i8> %2 to <2 x i64>
%4 = add <2 x i64> %3, %3
store <2 x i64> %4, <2 x i64>* %q
ret void
}
; CHECK-LABEL: test_v4f32_f128:
define void @test_v4f32_f128(fp128* %p, <4 x float>* %q) {
; CHECK: ldr q
; CHECK: rev64 v{{[0-9]+}}.4s
; CHECK: ext
; CHECK: rev64 v{{[0-9]+}}.4s
; CHECK: st1 { v{{[0-9]+}}.2d }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load fp128, fp128* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd fp128 %1, %1
%3 = bitcast fp128 %2 to <4 x float>
%4 = fadd <4 x float> %3, %3
store <4 x float> %4, <4 x float>* %q
ret void
}
; CHECK-LABEL: test_v4f32_v2f64:
define void @test_v4f32_v2f64(<2 x double>* %p, <4 x float>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2d }
; CHECK: rev64 v{{[0-9]+}}.4s
; CHECK: rev64 v{{[0-9]+}}.4s
; CHECK: st1 { v{{[0-9]+}}.2d }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <2 x double>, <2 x double>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd <2 x double> %1, %1
%3 = bitcast <2 x double> %2 to <4 x float>
%4 = fadd <4 x float> %3, %3
store <4 x float> %4, <4 x float>* %q
ret void
}
; CHECK-LABEL: test_v4f32_v2i64:
define void @test_v4f32_v2i64(<2 x i64>* %p, <4 x float>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2d }
; CHECK: rev64 v{{[0-9]+}}.4s
; CHECK: rev64 v{{[0-9]+}}.4s
; CHECK: st1 { v{{[0-9]+}}.2d }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <2 x i64>, <2 x i64>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <2 x i64> %1, %1
%3 = bitcast <2 x i64> %2 to <4 x float>
%4 = fadd <4 x float> %3, %3
store <4 x float> %4, <4 x float>* %q
ret void
}
; CHECK-LABEL: test_v4f32_v4i32:
define void @test_v4f32_v4i32(<4 x i32>* %p, <4 x float>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.4s }
; CHECK: rev64 v{{[0-9]+}}.4s
; CHECK: st1 { v{{[0-9]+}}.2d }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <4 x i32>, <4 x i32>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <4 x i32> %1, %1
%3 = bitcast <4 x i32> %2 to <4 x float>
%4 = fadd <4 x float> %3, %3
store <4 x float> %4, <4 x float>* %q
ret void
}
; CHECK-LABEL: test_v4f32_v8i16:
define void @test_v4f32_v8i16(<8 x i16>* %p, <4 x float>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.8h }
; CHECK: rev32 v{{[0-9]+}}.8h
; CHECK: rev64 v{{[0-9]+}}.4s
; CHECK: st1 { v{{[0-9]+}}.2d }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <8 x i16>, <8 x i16>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <8 x i16> %1, %1
%3 = bitcast <8 x i16> %2 to <4 x float>
%4 = fadd <4 x float> %3, %3
store <4 x float> %4, <4 x float>* %q
ret void
}
; CHECK-LABEL: test_v4f32_v16i8:
define void @test_v4f32_v16i8(<16 x i8>* %p, <4 x float>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.16b }
; CHECK: rev32 v{{[0-9]+}}.16b
; CHECK: rev64 v{{[0-9]+}}.4s
; CHECK: st1 { v{{[0-9]+}}.2d }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <16 x i8>, <16 x i8>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <16 x i8> %1, %1
%3 = bitcast <16 x i8> %2 to <4 x float>
%4 = fadd <4 x float> %3, %3
store <4 x float> %4, <4 x float>* %q
ret void
}
; CHECK-LABEL: test_v4i32_f128:
define void @test_v4i32_f128(fp128* %p, <4 x i32>* %q) {
; CHECK: ldr
; CHECK: rev64 v{{[0-9]+}}.4s
; CHECK: ext
; CHECK: st1 { v{{[0-9]+}}.4s }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load fp128, fp128* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd fp128 %1, %1
%3 = bitcast fp128 %2 to <4 x i32>
%4 = add <4 x i32> %3, %3
store <4 x i32> %4, <4 x i32>* %q
ret void
}
; CHECK-LABEL: test_v4i32_v2f64:
define void @test_v4i32_v2f64(<2 x double>* %p, <4 x i32>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2d }
; CHECK: rev64 v{{[0-9]+}}.4s
; CHECK: st1 { v{{[0-9]+}}.4s }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <2 x double>, <2 x double>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd <2 x double> %1, %1
%3 = bitcast <2 x double> %2 to <4 x i32>
%4 = add <4 x i32> %3, %3
store <4 x i32> %4, <4 x i32>* %q
ret void
}
; CHECK-LABEL: test_v4i32_v2i64:
define void @test_v4i32_v2i64(<2 x i64>* %p, <4 x i32>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2d }
; CHECK: rev64 v{{[0-9]+}}.4s
; CHECK: st1 { v{{[0-9]+}}.4s }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <2 x i64>, <2 x i64>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <2 x i64> %1, %1
%3 = bitcast <2 x i64> %2 to <4 x i32>
%4 = add <4 x i32> %3, %3
store <4 x i32> %4, <4 x i32>* %q
ret void
}
; CHECK-LABEL: test_v4i32_v4f32:
define void @test_v4i32_v4f32(<4 x float>* %p, <4 x i32>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2d }
; CHECK: rev64 v{{[0-9]+}}.4s
; CHECK: st1 { v{{[0-9]+}}.4s }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <4 x float>, <4 x float>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd <4 x float> %1, %1
%3 = bitcast <4 x float> %2 to <4 x i32>
%4 = add <4 x i32> %3, %3
store <4 x i32> %4, <4 x i32>* %q
ret void
}
; CHECK-LABEL: test_v4i32_v8i16:
define void @test_v4i32_v8i16(<8 x i16>* %p, <4 x i32>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.8h }
; CHECK: rev32 v{{[0-9]+}}.8h
; CHECK: st1 { v{{[0-9]+}}.4s }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <8 x i16>, <8 x i16>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <8 x i16> %1, %1
%3 = bitcast <8 x i16> %2 to <4 x i32>
%4 = add <4 x i32> %3, %3
store <4 x i32> %4, <4 x i32>* %q
ret void
}
; CHECK-LABEL: test_v4i32_v16i8:
define void @test_v4i32_v16i8(<16 x i8>* %p, <4 x i32>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.16b }
; CHECK: rev32 v{{[0-9]+}}.16b
; CHECK: st1 { v{{[0-9]+}}.4s }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <16 x i8>, <16 x i8>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <16 x i8> %1, %1
%3 = bitcast <16 x i8> %2 to <4 x i32>
%4 = add <4 x i32> %3, %3
store <4 x i32> %4, <4 x i32>* %q
ret void
}
; CHECK-LABEL: test_v8i16_f128:
define void @test_v8i16_f128(fp128* %p, <8 x i16>* %q) {
; CHECK: ldr
; CHECK: rev64 v{{[0-9]+}}.8h
; CHECK: ext
; CHECK: st1 { v{{[0-9]+}}.8h }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load fp128, fp128* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd fp128 %1, %1
%3 = bitcast fp128 %2 to <8 x i16>
%4 = add <8 x i16> %3, %3
store <8 x i16> %4, <8 x i16>* %q
ret void
}
; CHECK-LABEL: test_v8i16_v2f64:
define void @test_v8i16_v2f64(<2 x double>* %p, <8 x i16>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2d }
; CHECK: rev64 v{{[0-9]+}}.8h
; CHECK: st1 { v{{[0-9]+}}.8h }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <2 x double>, <2 x double>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd <2 x double> %1, %1
%3 = bitcast <2 x double> %2 to <8 x i16>
%4 = add <8 x i16> %3, %3
store <8 x i16> %4, <8 x i16>* %q
ret void
}
; CHECK-LABEL: test_v8i16_v2i64:
define void @test_v8i16_v2i64(<2 x i64>* %p, <8 x i16>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2d }
; CHECK: rev64 v{{[0-9]+}}.8h
; CHECK: st1 { v{{[0-9]+}}.8h }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <2 x i64>, <2 x i64>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <2 x i64> %1, %1
%3 = bitcast <2 x i64> %2 to <8 x i16>
%4 = add <8 x i16> %3, %3
store <8 x i16> %4, <8 x i16>* %q
ret void
}
; CHECK-LABEL: test_v8i16_v4f32:
define void @test_v8i16_v4f32(<4 x float>* %p, <8 x i16>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2d }
; CHECK: rev64 v{{[0-9]+}}.4s
; CHECK: rev32 v{{[0-9]+}}.8h
; CHECK: st1 { v{{[0-9]+}}.8h }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <4 x float>, <4 x float>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd <4 x float> %1, %1
%3 = bitcast <4 x float> %2 to <8 x i16>
%4 = add <8 x i16> %3, %3
store <8 x i16> %4, <8 x i16>* %q
ret void
}
; CHECK-LABEL: test_v8i16_v4i32:
define void @test_v8i16_v4i32(<4 x i32>* %p, <8 x i16>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.4s }
; CHECK: rev32 v{{[0-9]+}}.8h
; CHECK: st1 { v{{[0-9]+}}.8h }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <4 x i32>, <4 x i32>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <4 x i32> %1, %1
%3 = bitcast <4 x i32> %2 to <8 x i16>
%4 = add <8 x i16> %3, %3
store <8 x i16> %4, <8 x i16>* %q
ret void
}
; CHECK-LABEL: test_v8i16_v16i8:
define void @test_v8i16_v16i8(<16 x i8>* %p, <8 x i16>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.16b }
; CHECK: rev16 v{{[0-9]+}}.16b
; CHECK: st1 { v{{[0-9]+}}.8h }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <16 x i8>, <16 x i8>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <16 x i8> %1, %1
%3 = bitcast <16 x i8> %2 to <8 x i16>
%4 = add <8 x i16> %3, %3
store <8 x i16> %4, <8 x i16>* %q
ret void
}
; CHECK-LABEL: test_v16i8_f128:
define void @test_v16i8_f128(fp128* %p, <16 x i8>* %q) {
; CHECK: ldr q
; CHECK: rev64 v{{[0-9]+}}.16b
; CHECK: ext
; CHECK: st1 { v{{[0-9]+}}.16b }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load fp128, fp128* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd fp128 %1, %1
%3 = bitcast fp128 %2 to <16 x i8>
%4 = add <16 x i8> %3, %3
store <16 x i8> %4, <16 x i8>* %q
ret void
}
; CHECK-LABEL: test_v16i8_v2f64:
define void @test_v16i8_v2f64(<2 x double>* %p, <16 x i8>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2d }
; CHECK: rev64 v{{[0-9]+}}.16b
; CHECK: st1 { v{{[0-9]+}}.16b }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <2 x double>, <2 x double>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd <2 x double> %1, %1
%3 = bitcast <2 x double> %2 to <16 x i8>
%4 = add <16 x i8> %3, %3
store <16 x i8> %4, <16 x i8>* %q
ret void
}
; CHECK-LABEL: test_v16i8_v2i64:
define void @test_v16i8_v2i64(<2 x i64>* %p, <16 x i8>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2d }
; CHECK: rev64 v{{[0-9]+}}.16b
; CHECK: st1 { v{{[0-9]+}}.16b }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <2 x i64>, <2 x i64>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <2 x i64> %1, %1
%3 = bitcast <2 x i64> %2 to <16 x i8>
%4 = add <16 x i8> %3, %3
store <16 x i8> %4, <16 x i8>* %q
ret void
}
; CHECK-LABEL: test_v16i8_v4f32:
define void @test_v16i8_v4f32(<4 x float>* %p, <16 x i8>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.2d }
; CHECK: rev64 v{{[0-9]+}}.4s
; CHECK: rev32 v{{[0-9]+}}.16b
; CHECK: st1 { v{{[0-9]+}}.16b }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <4 x float>, <4 x float>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = fadd <4 x float> %1, %1
%3 = bitcast <4 x float> %2 to <16 x i8>
%4 = add <16 x i8> %3, %3
store <16 x i8> %4, <16 x i8>* %q
ret void
}
; CHECK-LABEL: test_v16i8_v4i32:
define void @test_v16i8_v4i32(<4 x i32>* %p, <16 x i8>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.4s }
; CHECK: rev32 v{{[0-9]+}}.16b
; CHECK: st1 { v{{[0-9]+}}.16b }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <4 x i32>, <4 x i32>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <4 x i32> %1, %1
%3 = bitcast <4 x i32> %2 to <16 x i8>
%4 = add <16 x i8> %3, %3
store <16 x i8> %4, <16 x i8>* %q
ret void
}
; CHECK-LABEL: test_v16i8_v8i16:
define void @test_v16i8_v8i16(<8 x i16>* %p, <16 x i8>* %q) {
; CHECK: ld1 { v{{[0-9]+}}.8h }
; CHECK: rev16 v{{[0-9]+}}.16b
; CHECK: st1 { v{{[0-9]+}}.16b }
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230794 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 21:17:42 +00:00
%1 = load <8 x i16>, <8 x i16>* %p
[ARM64-BE] Implement the crazy bitcast handling for big endian vectors. Because we've canonicalised on using LD1/ST1, every time we do a bitcast between vector types we must do an equivalent lane reversal. Consider a simple memory load followed by a bitconvert then a store. v0 = load v2i32 v1 = BITCAST v2i32 v0 to v4i16 store v4i16 v2 In big endian mode every memory access has an implicit byte swap. LDR and STR do a 64-bit byte swap, whereas LD1/ST1 do a byte swap per lane - that is, they treat the vector as a sequence of elements to be byte-swapped. The two pairs of instructions are fundamentally incompatible. We've decided to use LD1/ST1 only to simplify compiler implementation. LD1/ST1 perform the equivalent of a sequence of LDR/STR + REV. This makes the original code sequence: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = BITCAST v2i32 v1 to v4i16 v3 = REV v4i16 v2 (implicit) store v4i16 v3 But this is now broken - the value stored is different to the value loaded due to lane reordering. To fix this, on every BITCAST we must perform two other REVs: v0 = load v2i32 v1 = REV v2i32 (implicit) v2 = REV v2i32 v3 = BITCAST v2i32 v2 to v4i16 v4 = REV v4i16 v5 = REV v4i16 v4 (implicit) store v4i16 v5 This means an extra two instructions, but actually in most cases the two REV instructions can be combined into one. For example: (REV64_2s (REV64_4h X)) === (REV32_4h X) There is also no 128-bit REV instruction. This must be synthesized with an EXT instruction. Most bitconverts require some sort of conversion. The only exceptions are: a) Identity conversions - vNfX <-> vNiX b) Single-lane-to-scalar - v1fX <-> fX or v1iX <-> iX Even though there are hundreds of changed lines, I have a fairly high confidence that they are somewhat correct. The changes to add two REV instructions per bitcast were pretty mechanical, and once I'd done that I threw the resulting .td at a script I wrote which combined the two REVs together (and added an EXT instruction, for f128) based on an instruction description I gave it. This was much less prone to error than doing it all manually, plus my brain would not just have melted but would have vapourised. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208194 91177308-0d34-0410-b5e6-96231b3b80d8
2014-05-07 11:28:53 +00:00
%2 = add <8 x i16> %1, %1
%3 = bitcast <8 x i16> %2 to <16 x i8>
%4 = add <16 x i8> %3, %3
store <16 x i8> %4, <16 x i8>* %q
ret void
}
Reference in New Issue Copy Permalink