6502/llvm-6502
1
0
Fork 0
mirror of https://github.com/c64scene-ar/llvm-6502.git synced 2024-12-21 00:32:23 +00:00
Code Issues Projects Releases Wiki Activity
bc051faae1
llvm-6502/test/CodeGen/PowerPC/vsx-fma-m.ll

353 lines
13 KiB
LLVM
Raw Normal View History

[PowerPC] Select between VSX A-type and M-type FMA instructions just before RA The VSX instruction set has two types of FMA instructions: A-type (where the addend is taken from the output register) and M-type (where one of the product operands is taken from the output register). This adds a small pass that runs just after MI scheduling (and, thus, just before register allocation) that mutates A-type instructions (that are created during isel) into M-type instructions when: 1. This will eliminate an otherwise-necessary copy of the addend 2. One of the product operands is killed by the instruction The "right" moment to make this decision is in between scheduling and register allocation, because only there do we know whether or not one of the product operands is killed by any particular instruction. Unfortunately, this also makes the implementation somewhat complicated, because the MIs are not in SSA form and we need to preserve the LiveIntervals analysis. As a simple example, if we have: %vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9 %vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16, %RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16 ... %vreg9<def,tied1> = XSMADDADP %vreg9<tied0>, %vreg17, %vreg19, %RM<imp-use>; VSLRC:%vreg9,%vreg17,%vreg19 ... We can eliminate the copy by changing from the A-type to the M-type instruction. This means: %vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16, %RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16 is replaced by: %vreg16<def,tied1> = XSMADDMDP %vreg16<tied0>, %vreg18, %vreg9, %RM<imp-use>; VSLRC:%vreg16,%vreg18,%vreg9 and we remove: %vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204768 91177308-0d34-0410-b5e6-96231b3b80d8
2014-03-25 23:29:21 +00:00
; RUN: llc < %s -mcpu=pwr7 -mattr=+vsx | FileCheck %s
[PowerPC]Update Power VSX test cases to also test fast-isel Update of some of the VSX test cases for Power to check fast-isel codegen as well as the regular codegen. http://reviews.llvm.org/D6357 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@223509 91177308-0d34-0410-b5e6-96231b3b80d8
2014-12-05 20:32:05 +00:00
; RUN: llc < %s -mcpu=pwr7 -mattr=+vsx -fast-isel -O0 | FileCheck -check-prefix=CHECK-FISL %s
[PowerPC] Select between VSX A-type and M-type FMA instructions just before RA The VSX instruction set has two types of FMA instructions: A-type (where the addend is taken from the output register) and M-type (where one of the product operands is taken from the output register). This adds a small pass that runs just after MI scheduling (and, thus, just before register allocation) that mutates A-type instructions (that are created during isel) into M-type instructions when: 1. This will eliminate an otherwise-necessary copy of the addend 2. One of the product operands is killed by the instruction The "right" moment to make this decision is in between scheduling and register allocation, because only there do we know whether or not one of the product operands is killed by any particular instruction. Unfortunately, this also makes the implementation somewhat complicated, because the MIs are not in SSA form and we need to preserve the LiveIntervals analysis. As a simple example, if we have: %vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9 %vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16, %RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16 ... %vreg9<def,tied1> = XSMADDADP %vreg9<tied0>, %vreg17, %vreg19, %RM<imp-use>; VSLRC:%vreg9,%vreg17,%vreg19 ... We can eliminate the copy by changing from the A-type to the M-type instruction. This means: %vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16, %RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16 is replaced by: %vreg16<def,tied1> = XSMADDMDP %vreg16<tied0>, %vreg18, %vreg9, %RM<imp-use>; VSLRC:%vreg16,%vreg18,%vreg9 and we remove: %vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204768 91177308-0d34-0410-b5e6-96231b3b80d8
2014-03-25 23:29:21 +00:00
; Also run with -schedule-ppc-vsx-fma-mutation-early as a stress test for the
; live-interval-updating logic.
; RUN: llc < %s -mcpu=pwr7 -mattr=+vsx -schedule-ppc-vsx-fma-mutation-early
target datalayout = "E-m:e-i64:64-n32:64"
target triple = "powerpc64-unknown-linux-gnu"
define void @test1(double %a, double %b, double %c, double %e, double* nocapture %d) #0 {
entry:
%0 = tail call double @llvm.fma.f64(double %b, double %c, double %a)
store double %0, double* %d, align 8
%1 = tail call double @llvm.fma.f64(double %b, double %e, double %a)
%arrayidx1 = getelementptr inbounds double* %d, i64 1
store double %1, double* %arrayidx1, align 8
ret void
; CHECK-LABEL: @test1
; CHECK-DAG: li [[C1:[0-9]+]], 8
; CHECK-DAG: xsmaddmdp 3, 2, 1
; CHECK-DAG: xsmaddadp 1, 2, 4
; CHECK-DAG: stxsdx 3, 0, 7
; CHECK-DAG: stxsdx 1, 7, [[C1]]
; CHECK: blr
[PowerPC]Update Power VSX test cases to also test fast-isel Update of some of the VSX test cases for Power to check fast-isel codegen as well as the regular codegen. http://reviews.llvm.org/D6357 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@223509 91177308-0d34-0410-b5e6-96231b3b80d8
2014-12-05 20:32:05 +00:00
; CHECK-FISL-LABEL: @test1
; CHECK-FISL-DAG: fmr 0, 1
; CHECK-FISL-DAG: xsmaddadp 0, 2, 3
; CHECK-FISL-DAG: stxsdx 0, 0, 7
; CHECK-FISL-DAG: xsmaddadp 1, 2, 4
; CHECK-FISL-DAG: li [[C1:[0-9]+]], 8
; CHECK-FISL-DAG: stxsdx 1, 7, [[C1]]
; CHECK-FISL: blr
[PowerPC] Select between VSX A-type and M-type FMA instructions just before RA The VSX instruction set has two types of FMA instructions: A-type (where the addend is taken from the output register) and M-type (where one of the product operands is taken from the output register). This adds a small pass that runs just after MI scheduling (and, thus, just before register allocation) that mutates A-type instructions (that are created during isel) into M-type instructions when: 1. This will eliminate an otherwise-necessary copy of the addend 2. One of the product operands is killed by the instruction The "right" moment to make this decision is in between scheduling and register allocation, because only there do we know whether or not one of the product operands is killed by any particular instruction. Unfortunately, this also makes the implementation somewhat complicated, because the MIs are not in SSA form and we need to preserve the LiveIntervals analysis. As a simple example, if we have: %vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9 %vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16, %RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16 ... %vreg9<def,tied1> = XSMADDADP %vreg9<tied0>, %vreg17, %vreg19, %RM<imp-use>; VSLRC:%vreg9,%vreg17,%vreg19 ... We can eliminate the copy by changing from the A-type to the M-type instruction. This means: %vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16, %RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16 is replaced by: %vreg16<def,tied1> = XSMADDMDP %vreg16<tied0>, %vreg18, %vreg9, %RM<imp-use>; VSLRC:%vreg16,%vreg18,%vreg9 and we remove: %vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204768 91177308-0d34-0410-b5e6-96231b3b80d8
2014-03-25 23:29:21 +00:00
}
define void @test2(double %a, double %b, double %c, double %e, double %f, double* nocapture %d) #0 {
entry:
%0 = tail call double @llvm.fma.f64(double %b, double %c, double %a)
store double %0, double* %d, align 8
%1 = tail call double @llvm.fma.f64(double %b, double %e, double %a)
%arrayidx1 = getelementptr inbounds double* %d, i64 1
store double %1, double* %arrayidx1, align 8
%2 = tail call double @llvm.fma.f64(double %b, double %f, double %a)
%arrayidx2 = getelementptr inbounds double* %d, i64 2
store double %2, double* %arrayidx2, align 8
ret void
; CHECK-LABEL: @test2
; CHECK-DAG: li [[C1:[0-9]+]], 8
; CHECK-DAG: li [[C2:[0-9]+]], 16
; CHECK-DAG: xsmaddmdp 3, 2, 1
; CHECK-DAG: xsmaddmdp 4, 2, 1
; CHECK-DAG: xsmaddadp 1, 2, 5
; CHECK-DAG: stxsdx 3, 0, 8
; CHECK-DAG: stxsdx 4, 8, [[C1]]
; CHECK-DAG: stxsdx 1, 8, [[C2]]
; CHECK: blr
[PowerPC]Update Power VSX test cases to also test fast-isel Update of some of the VSX test cases for Power to check fast-isel codegen as well as the regular codegen. http://reviews.llvm.org/D6357 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@223509 91177308-0d34-0410-b5e6-96231b3b80d8
2014-12-05 20:32:05 +00:00
; CHECK-FISL-LABEL: @test2
; CHECK-FISL-DAG: fmr 0, 1
; CHECK-FISL-DAG: xsmaddadp 0, 2, 3
; CHECK-FISL-DAG: stxsdx 0, 0, 8
; CHECK-FISL-DAG: fmr 0, 1
; CHECK-FISL-DAG: xsmaddadp 0, 2, 4
; CHECK-FISL-DAG: li [[C1:[0-9]+]], 8
; CHECK-FISL-DAG: stxsdx 0, 8, [[C1]]
; CHECK-FISL-DAG: xsmaddadp 1, 2, 5
; CHECK-FISL-DAG: li [[C2:[0-9]+]], 16
; CHECK-FISL-DAG: stxsdx 1, 8, [[C2]]
; CHECK-FISL: blr
[PowerPC] Select between VSX A-type and M-type FMA instructions just before RA The VSX instruction set has two types of FMA instructions: A-type (where the addend is taken from the output register) and M-type (where one of the product operands is taken from the output register). This adds a small pass that runs just after MI scheduling (and, thus, just before register allocation) that mutates A-type instructions (that are created during isel) into M-type instructions when: 1. This will eliminate an otherwise-necessary copy of the addend 2. One of the product operands is killed by the instruction The "right" moment to make this decision is in between scheduling and register allocation, because only there do we know whether or not one of the product operands is killed by any particular instruction. Unfortunately, this also makes the implementation somewhat complicated, because the MIs are not in SSA form and we need to preserve the LiveIntervals analysis. As a simple example, if we have: %vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9 %vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16, %RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16 ... %vreg9<def,tied1> = XSMADDADP %vreg9<tied0>, %vreg17, %vreg19, %RM<imp-use>; VSLRC:%vreg9,%vreg17,%vreg19 ... We can eliminate the copy by changing from the A-type to the M-type instruction. This means: %vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16, %RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16 is replaced by: %vreg16<def,tied1> = XSMADDMDP %vreg16<tied0>, %vreg18, %vreg9, %RM<imp-use>; VSLRC:%vreg16,%vreg18,%vreg9 and we remove: %vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204768 91177308-0d34-0410-b5e6-96231b3b80d8
2014-03-25 23:29:21 +00:00
}
define void @test3(double %a, double %b, double %c, double %e, double %f, double* nocapture %d) #0 {
entry:
%0 = tail call double @llvm.fma.f64(double %b, double %c, double %a)
store double %0, double* %d, align 8
%1 = tail call double @llvm.fma.f64(double %b, double %e, double %a)
%2 = tail call double @llvm.fma.f64(double %b, double %c, double %1)
%arrayidx1 = getelementptr inbounds double* %d, i64 3
store double %2, double* %arrayidx1, align 8
%3 = tail call double @llvm.fma.f64(double %b, double %f, double %a)
%arrayidx2 = getelementptr inbounds double* %d, i64 2
store double %3, double* %arrayidx2, align 8
%arrayidx3 = getelementptr inbounds double* %d, i64 1
store double %1, double* %arrayidx3, align 8
ret void
; CHECK-LABEL: @test3
[PowerPC] Add subregister classes for f64 VSX values We had stored both f64 values and v2f64, etc. values in the VSX registers. This worked, but was suboptimal because we would always spill 16-byte values even through we almost always had scalar 8-byte values. This resulted in an increase in stack-size use, extra memory bandwidth, etc. To fix this, I've added 64-bit subregisters of the Altivec registers, and combined those with the existing scalar floating-point registers to form a class of VSX scalar floating-point registers. The ABI code has also been enhanced to use this register class and some other necessary improvements have been made. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205075 91177308-0d34-0410-b5e6-96231b3b80d8
2014-03-29 05:29:01 +00:00
; CHECK-DAG: fmr [[F1:[0-9]+]], 1
[PowerPC] Select between VSX A-type and M-type FMA instructions just before RA The VSX instruction set has two types of FMA instructions: A-type (where the addend is taken from the output register) and M-type (where one of the product operands is taken from the output register). This adds a small pass that runs just after MI scheduling (and, thus, just before register allocation) that mutates A-type instructions (that are created during isel) into M-type instructions when: 1. This will eliminate an otherwise-necessary copy of the addend 2. One of the product operands is killed by the instruction The "right" moment to make this decision is in between scheduling and register allocation, because only there do we know whether or not one of the product operands is killed by any particular instruction. Unfortunately, this also makes the implementation somewhat complicated, because the MIs are not in SSA form and we need to preserve the LiveIntervals analysis. As a simple example, if we have: %vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9 %vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16, %RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16 ... %vreg9<def,tied1> = XSMADDADP %vreg9<tied0>, %vreg17, %vreg19, %RM<imp-use>; VSLRC:%vreg9,%vreg17,%vreg19 ... We can eliminate the copy by changing from the A-type to the M-type instruction. This means: %vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16, %RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16 is replaced by: %vreg16<def,tied1> = XSMADDMDP %vreg16<tied0>, %vreg18, %vreg9, %RM<imp-use>; VSLRC:%vreg16,%vreg18,%vreg9 and we remove: %vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204768 91177308-0d34-0410-b5e6-96231b3b80d8
2014-03-25 23:29:21 +00:00
; CHECK-DAG: li [[C1:[0-9]+]], 24
; CHECK-DAG: li [[C2:[0-9]+]], 16
; CHECK-DAG: li [[C3:[0-9]+]], 8
; CHECK-DAG: xsmaddmdp 4, 2, 1
; CHECK-DAG: xsmaddadp 1, 2, 5
; Note: We could convert this next FMA to M-type as well, but it would require
; re-ordering the instructions.
; CHECK-DAG: xsmaddadp [[F1]], 2, 3
; CHECK-DAG: xsmaddmdp 2, 3, 4
; CHECK-DAG: stxsdx [[F1]], 0, 8
; CHECK-DAG: stxsdx 2, 8, [[C1]]
; CHECK-DAG: stxsdx 1, 8, [[C2]]
; CHECK-DAG: stxsdx 4, 8, [[C3]]
[PowerPC] Add subregister classes for f64 VSX values We had stored both f64 values and v2f64, etc. values in the VSX registers. This worked, but was suboptimal because we would always spill 16-byte values even through we almost always had scalar 8-byte values. This resulted in an increase in stack-size use, extra memory bandwidth, etc. To fix this, I've added 64-bit subregisters of the Altivec registers, and combined those with the existing scalar floating-point registers to form a class of VSX scalar floating-point registers. The ABI code has also been enhanced to use this register class and some other necessary improvements have been made. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205075 91177308-0d34-0410-b5e6-96231b3b80d8
2014-03-29 05:29:01 +00:00
; CHECK: blr
[PowerPC]Update Power VSX test cases to also test fast-isel Update of some of the VSX test cases for Power to check fast-isel codegen as well as the regular codegen. http://reviews.llvm.org/D6357 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@223509 91177308-0d34-0410-b5e6-96231b3b80d8
2014-12-05 20:32:05 +00:00
; CHECK-FISL-LABEL: @test3
; CHECK-FISL-DAG: fmr [[F1:[0-9]+]], 1
; CHECK-FISL-DAG: xsmaddadp [[F1]], 2, 4
; CHECK-FISL-DAG: fmr 4, [[F1]]
; CHECK-FISL-DAG: xsmaddadp 4, 2, 3
; CHECK-FISL-DAG: li [[C1:[0-9]+]], 24
; CHECK-FISL-DAG: stxsdx 4, 8, [[C1]]
; CHECK-FISL-DAG: xsmaddadp 1, 2, 5
; CHECK-FISL-DAG: li [[C2:[0-9]+]], 16
; CHECK-FISL-DAG: stxsdx 1, 8, [[C2]]
; CHECK-FISL-DAG: li [[C3:[0-9]+]], 8
; CHECK-FISL-DAG: stxsdx 0, 8, [[C3]]
; CHECK-FISL: blr
[PowerPC] Select between VSX A-type and M-type FMA instructions just before RA The VSX instruction set has two types of FMA instructions: A-type (where the addend is taken from the output register) and M-type (where one of the product operands is taken from the output register). This adds a small pass that runs just after MI scheduling (and, thus, just before register allocation) that mutates A-type instructions (that are created during isel) into M-type instructions when: 1. This will eliminate an otherwise-necessary copy of the addend 2. One of the product operands is killed by the instruction The "right" moment to make this decision is in between scheduling and register allocation, because only there do we know whether or not one of the product operands is killed by any particular instruction. Unfortunately, this also makes the implementation somewhat complicated, because the MIs are not in SSA form and we need to preserve the LiveIntervals analysis. As a simple example, if we have: %vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9 %vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16, %RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16 ... %vreg9<def,tied1> = XSMADDADP %vreg9<tied0>, %vreg17, %vreg19, %RM<imp-use>; VSLRC:%vreg9,%vreg17,%vreg19 ... We can eliminate the copy by changing from the A-type to the M-type instruction. This means: %vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16, %RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16 is replaced by: %vreg16<def,tied1> = XSMADDMDP %vreg16<tied0>, %vreg18, %vreg9, %RM<imp-use>; VSLRC:%vreg16,%vreg18,%vreg9 and we remove: %vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204768 91177308-0d34-0410-b5e6-96231b3b80d8
2014-03-25 23:29:21 +00:00
}
define void @test4(double %a, double %b, double %c, double %e, double %f, double* nocapture %d) #0 {
entry:
%0 = tail call double @llvm.fma.f64(double %b, double %c, double %a)
store double %0, double* %d, align 8
%1 = tail call double @llvm.fma.f64(double %b, double %e, double %a)
%arrayidx1 = getelementptr inbounds double* %d, i64 1
store double %1, double* %arrayidx1, align 8
%2 = tail call double @llvm.fma.f64(double %b, double %c, double %1)
%arrayidx3 = getelementptr inbounds double* %d, i64 3
store double %2, double* %arrayidx3, align 8
%3 = tail call double @llvm.fma.f64(double %b, double %f, double %a)
%arrayidx4 = getelementptr inbounds double* %d, i64 2
store double %3, double* %arrayidx4, align 8
ret void
; CHECK-LABEL: @test4
[PowerPC] Add subregister classes for f64 VSX values We had stored both f64 values and v2f64, etc. values in the VSX registers. This worked, but was suboptimal because we would always spill 16-byte values even through we almost always had scalar 8-byte values. This resulted in an increase in stack-size use, extra memory bandwidth, etc. To fix this, I've added 64-bit subregisters of the Altivec registers, and combined those with the existing scalar floating-point registers to form a class of VSX scalar floating-point registers. The ABI code has also been enhanced to use this register class and some other necessary improvements have been made. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205075 91177308-0d34-0410-b5e6-96231b3b80d8
2014-03-29 05:29:01 +00:00
; CHECK-DAG: fmr [[F1:[0-9]+]], 1
[PowerPC] Select between VSX A-type and M-type FMA instructions just before RA The VSX instruction set has two types of FMA instructions: A-type (where the addend is taken from the output register) and M-type (where one of the product operands is taken from the output register). This adds a small pass that runs just after MI scheduling (and, thus, just before register allocation) that mutates A-type instructions (that are created during isel) into M-type instructions when: 1. This will eliminate an otherwise-necessary copy of the addend 2. One of the product operands is killed by the instruction The "right" moment to make this decision is in between scheduling and register allocation, because only there do we know whether or not one of the product operands is killed by any particular instruction. Unfortunately, this also makes the implementation somewhat complicated, because the MIs are not in SSA form and we need to preserve the LiveIntervals analysis. As a simple example, if we have: %vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9 %vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16, %RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16 ... %vreg9<def,tied1> = XSMADDADP %vreg9<tied0>, %vreg17, %vreg19, %RM<imp-use>; VSLRC:%vreg9,%vreg17,%vreg19 ... We can eliminate the copy by changing from the A-type to the M-type instruction. This means: %vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16, %RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16 is replaced by: %vreg16<def,tied1> = XSMADDMDP %vreg16<tied0>, %vreg18, %vreg9, %RM<imp-use>; VSLRC:%vreg16,%vreg18,%vreg9 and we remove: %vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204768 91177308-0d34-0410-b5e6-96231b3b80d8
2014-03-25 23:29:21 +00:00
; CHECK-DAG: li [[C1:[0-9]+]], 8
; CHECK-DAG: li [[C2:[0-9]+]], 16
; CHECK-DAG: xsmaddmdp 4, 2, 1
; Note: We could convert this next FMA to M-type as well, but it would require
; re-ordering the instructions.
; CHECK-DAG: xsmaddadp 1, 2, 5
; CHECK-DAG: xsmaddadp [[F1]], 2, 3
; CHECK-DAG: stxsdx [[F1]], 0, 8
; CHECK-DAG: stxsdx 4, 8, [[C1]]
; CHECK-DAG: li [[C3:[0-9]+]], 24
; CHECK-DAG: xsmaddadp 4, 2, 3
; CHECK-DAG: stxsdx 4, 8, [[C3]]
; CHECK-DAG: stxsdx 1, 8, [[C2]]
; CHECK: blr
[PowerPC]Update Power VSX test cases to also test fast-isel Update of some of the VSX test cases for Power to check fast-isel codegen as well as the regular codegen. http://reviews.llvm.org/D6357 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@223509 91177308-0d34-0410-b5e6-96231b3b80d8
2014-12-05 20:32:05 +00:00
; CHECK-FISL-LABEL: @test4
; CHECK-FISL-DAG: fmr [[F1:[0-9]+]], 1
; CHECK-FISL-DAG: xsmaddadp [[F1]], 2, 3
; CHECK-FISL-DAG: stxsdx 0, 0, 8
; CHECK-FISL-DAG: fmr [[F1]], 1
; CHECK-FISL-DAG: xsmaddadp [[F1]], 2, 4
; CHECK-FISL-DAG: li [[C3:[0-9]+]], 8
; CHECK-FISL-DAG: stxsdx 0, 8, [[C3]]
; CHECK-FISL-DAG: xsmaddadp 0, 2, 3
; CHECK-FISL-DAG: li [[C1:[0-9]+]], 24
; CHECK-FISL-DAG: stxsdx 0, 8, [[C1]]
; CHECK-FISL-DAG: xsmaddadp 1, 2, 5
; CHECK-FISL-DAG: li [[C2:[0-9]+]], 16
; CHECK-FISL-DAG: stxsdx 1, 8, [[C2]]
; CHECK-FISL: blr
[PowerPC] Select between VSX A-type and M-type FMA instructions just before RA The VSX instruction set has two types of FMA instructions: A-type (where the addend is taken from the output register) and M-type (where one of the product operands is taken from the output register). This adds a small pass that runs just after MI scheduling (and, thus, just before register allocation) that mutates A-type instructions (that are created during isel) into M-type instructions when: 1. This will eliminate an otherwise-necessary copy of the addend 2. One of the product operands is killed by the instruction The "right" moment to make this decision is in between scheduling and register allocation, because only there do we know whether or not one of the product operands is killed by any particular instruction. Unfortunately, this also makes the implementation somewhat complicated, because the MIs are not in SSA form and we need to preserve the LiveIntervals analysis. As a simple example, if we have: %vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9 %vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16, %RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16 ... %vreg9<def,tied1> = XSMADDADP %vreg9<tied0>, %vreg17, %vreg19, %RM<imp-use>; VSLRC:%vreg9,%vreg17,%vreg19 ... We can eliminate the copy by changing from the A-type to the M-type instruction. This means: %vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16, %RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16 is replaced by: %vreg16<def,tied1> = XSMADDMDP %vreg16<tied0>, %vreg18, %vreg9, %RM<imp-use>; VSLRC:%vreg16,%vreg18,%vreg9 and we remove: %vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204768 91177308-0d34-0410-b5e6-96231b3b80d8
2014-03-25 23:29:21 +00:00
}
declare double @llvm.fma.f64(double, double, double) #0
[PowerPC] Add subregister classes for f64 VSX values We had stored both f64 values and v2f64, etc. values in the VSX registers. This worked, but was suboptimal because we would always spill 16-byte values even through we almost always had scalar 8-byte values. This resulted in an increase in stack-size use, extra memory bandwidth, etc. To fix this, I've added 64-bit subregisters of the Altivec registers, and combined those with the existing scalar floating-point registers to form a class of VSX scalar floating-point registers. The ABI code has also been enhanced to use this register class and some other necessary improvements have been made. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205075 91177308-0d34-0410-b5e6-96231b3b80d8
2014-03-29 05:29:01 +00:00
define void @testv1(<2 x double> %a, <2 x double> %b, <2 x double> %c, <2 x double> %e, <2 x double>* nocapture %d) #0 {
entry:
%0 = tail call <2 x double> @llvm.fma.v2f64(<2 x double> %b, <2 x double> %c, <2 x double> %a)
store <2 x double> %0, <2 x double>* %d, align 8
%1 = tail call <2 x double> @llvm.fma.v2f64(<2 x double> %b, <2 x double> %e, <2 x double> %a)
%arrayidx1 = getelementptr inbounds <2 x double>* %d, i64 1
store <2 x double> %1, <2 x double>* %arrayidx1, align 8
ret void
; CHECK-LABEL: @testv1
; CHECK-DAG: xvmaddmdp 36, 35, 34
; CHECK-DAG: xvmaddadp 34, 35, 37
; CHECK-DAG: li [[C1:[0-9]+]], 16
; CHECK-DAG: stxvd2x 36, 0, 3
; CHECK-DAG: stxvd2x 34, 3, [[C1:[0-9]+]]
; CHECK: blr
[PowerPC]Update Power VSX test cases to also test fast-isel Update of some of the VSX test cases for Power to check fast-isel codegen as well as the regular codegen. http://reviews.llvm.org/D6357 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@223509 91177308-0d34-0410-b5e6-96231b3b80d8
2014-12-05 20:32:05 +00:00
; CHECK-FISL-LABEL: @testv1
; CHECK-FISL-DAG: xxlor 0, 34, 34
; CHECK-FISL-DAG: xvmaddadp 0, 35, 36
; CHECK-FISL-DAG: stxvd2x 0, 0, 3
; CHECK-FISL-DAG: xvmaddadp 34, 35, 37
; CHECK-FISL-DAG: li [[C1:[0-9]+]], 16
; CHECK-FISL-DAG: stxvd2x 34, 3, [[C1:[0-9]+]]
; CHECK-FISL: blr
[PowerPC] Add subregister classes for f64 VSX values We had stored both f64 values and v2f64, etc. values in the VSX registers. This worked, but was suboptimal because we would always spill 16-byte values even through we almost always had scalar 8-byte values. This resulted in an increase in stack-size use, extra memory bandwidth, etc. To fix this, I've added 64-bit subregisters of the Altivec registers, and combined those with the existing scalar floating-point registers to form a class of VSX scalar floating-point registers. The ABI code has also been enhanced to use this register class and some other necessary improvements have been made. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205075 91177308-0d34-0410-b5e6-96231b3b80d8
2014-03-29 05:29:01 +00:00
}
define void @testv2(<2 x double> %a, <2 x double> %b, <2 x double> %c, <2 x double> %e, <2 x double> %f, <2 x double>* nocapture %d) #0 {
entry:
%0 = tail call <2 x double> @llvm.fma.v2f64(<2 x double> %b, <2 x double> %c, <2 x double> %a)
store <2 x double> %0, <2 x double>* %d, align 8
%1 = tail call <2 x double> @llvm.fma.v2f64(<2 x double> %b, <2 x double> %e, <2 x double> %a)
%arrayidx1 = getelementptr inbounds <2 x double>* %d, i64 1
store <2 x double> %1, <2 x double>* %arrayidx1, align 8
%2 = tail call <2 x double> @llvm.fma.v2f64(<2 x double> %b, <2 x double> %f, <2 x double> %a)
%arrayidx2 = getelementptr inbounds <2 x double>* %d, i64 2
store <2 x double> %2, <2 x double>* %arrayidx2, align 8
ret void
; CHECK-LABEL: @testv2
; CHECK-DAG: xvmaddmdp 36, 35, 34
; CHECK-DAG: xvmaddmdp 37, 35, 34
; CHECK-DAG: li [[C1:[0-9]+]], 16
; CHECK-DAG: li [[C2:[0-9]+]], 32
; CHECK-DAG: xvmaddadp 34, 35, 38
; CHECK-DAG: stxvd2x 36, 0, 3
; CHECK-DAG: stxvd2x 37, 3, [[C1:[0-9]+]]
; CHECK-DAG: stxvd2x 34, 3, [[C2:[0-9]+]]
; CHECK: blr
[PowerPC]Update Power VSX test cases to also test fast-isel Update of some of the VSX test cases for Power to check fast-isel codegen as well as the regular codegen. http://reviews.llvm.org/D6357 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@223509 91177308-0d34-0410-b5e6-96231b3b80d8
2014-12-05 20:32:05 +00:00
; CHECK-FISL-LABEL: @testv2
; CHECK-FISL-DAG: xxlor 0, 34, 34
; CHECK-FISL-DAG: xvmaddadp 0, 35, 36
; CHECK-FISL-DAG: stxvd2x 0, 0, 3
; CHECK-FISL-DAG: xxlor 0, 34, 34
; CHECK-FISL-DAG: xvmaddadp 0, 35, 37
; CHECK-FISL-DAG: li [[C1:[0-9]+]], 16
; CHECK-FISL-DAG: stxvd2x 0, 3, [[C1:[0-9]+]]
; CHECK-FISL-DAG: xvmaddadp 34, 35, 38
; CHECK-FISL-DAG: li [[C2:[0-9]+]], 32
; CHECK-FISL-DAG: stxvd2x 34, 3, [[C2:[0-9]+]]
; CHECK-FISL: blr
[PowerPC] Add subregister classes for f64 VSX values We had stored both f64 values and v2f64, etc. values in the VSX registers. This worked, but was suboptimal because we would always spill 16-byte values even through we almost always had scalar 8-byte values. This resulted in an increase in stack-size use, extra memory bandwidth, etc. To fix this, I've added 64-bit subregisters of the Altivec registers, and combined those with the existing scalar floating-point registers to form a class of VSX scalar floating-point registers. The ABI code has also been enhanced to use this register class and some other necessary improvements have been made. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205075 91177308-0d34-0410-b5e6-96231b3b80d8
2014-03-29 05:29:01 +00:00
}
define void @testv3(<2 x double> %a, <2 x double> %b, <2 x double> %c, <2 x double> %e, <2 x double> %f, <2 x double>* nocapture %d) #0 {
entry:
%0 = tail call <2 x double> @llvm.fma.v2f64(<2 x double> %b, <2 x double> %c, <2 x double> %a)
store <2 x double> %0, <2 x double>* %d, align 8
%1 = tail call <2 x double> @llvm.fma.v2f64(<2 x double> %b, <2 x double> %e, <2 x double> %a)
%2 = tail call <2 x double> @llvm.fma.v2f64(<2 x double> %b, <2 x double> %c, <2 x double> %1)
%arrayidx1 = getelementptr inbounds <2 x double>* %d, i64 3
store <2 x double> %2, <2 x double>* %arrayidx1, align 8
%3 = tail call <2 x double> @llvm.fma.v2f64(<2 x double> %b, <2 x double> %f, <2 x double> %a)
%arrayidx2 = getelementptr inbounds <2 x double>* %d, i64 2
store <2 x double> %3, <2 x double>* %arrayidx2, align 8
%arrayidx3 = getelementptr inbounds <2 x double>* %d, i64 1
store <2 x double> %1, <2 x double>* %arrayidx3, align 8
ret void
[PowerPC] Add vec_vsx_ld and vec_vsx_st intrinsics This patch enables the vec_vsx_ld and vec_vsx_st intrinsics for PowerPC, which provide programmer access to the lxvd2x, lxvw4x, stxvd2x, and stxvw4x instructions. New LLVM intrinsics are provided to represent these four instructions in IntrinsicsPowerPC.td. These are patterned after the similar intrinsics for lvx and stvx (Altivec). In PPCInstrVSX.td, these intrinsics are tied to the code gen patterns, with additional patterns to allow plain vanilla loads and stores to still generate these instructions. At -O1 and higher the intrinsics are immediately converted to loads and stores in InstCombineCalls.cpp. This will open up more optimization opportunities while still allowing the correct instructions to be generated. (Similar code exists for aligned Altivec loads and stores.) The new intrinsics are added to the code that checks for consecutive loads and stores in PPCISelLowering.cpp, as well as to PPCTargetLowering::getTgtMemIntrinsic(). There's a new test to verify the correct instructions are generated. The loads and stores tend to be reordered, so the test just counts their number. It runs at -O2, as it's not very effective to test this at -O0, when many unnecessary loads and stores are generated. I ended up having to modify vsx-fma-m.ll. It turns out this test case is slightly unreliable, but I don't know a good way to prevent problems with it. The xvmaddmdp instructions read and write the same register, which is one of the multiplicands. Commutativity allows either to be chosen. If the FMAs are reordered differently than expected by the test, the register assignment can be different as a result. Hopefully this doesn't change often. There is a companion patch for Clang. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@221767 91177308-0d34-0410-b5e6-96231b3b80d8
2014-11-12 04:19:40 +00:00
; Note: There is some unavoidable changeability in this variant. If the
; FMAs are reordered differently, the algorithm can pick a different
; multiplicand to destroy, changing the register assignment. There isn't
; a good way to express this possibility, so hopefully this doesn't change
; too often.
[PowerPC] Add subregister classes for f64 VSX values We had stored both f64 values and v2f64, etc. values in the VSX registers. This worked, but was suboptimal because we would always spill 16-byte values even through we almost always had scalar 8-byte values. This resulted in an increase in stack-size use, extra memory bandwidth, etc. To fix this, I've added 64-bit subregisters of the Altivec registers, and combined those with the existing scalar floating-point registers to form a class of VSX scalar floating-point registers. The ABI code has also been enhanced to use this register class and some other necessary improvements have been made. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205075 91177308-0d34-0410-b5e6-96231b3b80d8
2014-03-29 05:29:01 +00:00
; CHECK-LABEL: @testv3
; CHECK-DAG: xxlor [[V1:[0-9]+]], 34, 34
; CHECK-DAG: li [[C1:[0-9]+]], 48
; CHECK-DAG: li [[C2:[0-9]+]], 32
[PowerPC] Add vec_vsx_ld and vec_vsx_st intrinsics This patch enables the vec_vsx_ld and vec_vsx_st intrinsics for PowerPC, which provide programmer access to the lxvd2x, lxvw4x, stxvd2x, and stxvw4x instructions. New LLVM intrinsics are provided to represent these four instructions in IntrinsicsPowerPC.td. These are patterned after the similar intrinsics for lvx and stvx (Altivec). In PPCInstrVSX.td, these intrinsics are tied to the code gen patterns, with additional patterns to allow plain vanilla loads and stores to still generate these instructions. At -O1 and higher the intrinsics are immediately converted to loads and stores in InstCombineCalls.cpp. This will open up more optimization opportunities while still allowing the correct instructions to be generated. (Similar code exists for aligned Altivec loads and stores.) The new intrinsics are added to the code that checks for consecutive loads and stores in PPCISelLowering.cpp, as well as to PPCTargetLowering::getTgtMemIntrinsic(). There's a new test to verify the correct instructions are generated. The loads and stores tend to be reordered, so the test just counts their number. It runs at -O2, as it's not very effective to test this at -O0, when many unnecessary loads and stores are generated. I ended up having to modify vsx-fma-m.ll. It turns out this test case is slightly unreliable, but I don't know a good way to prevent problems with it. The xvmaddmdp instructions read and write the same register, which is one of the multiplicands. Commutativity allows either to be chosen. If the FMAs are reordered differently than expected by the test, the register assignment can be different as a result. Hopefully this doesn't change often. There is a companion patch for Clang. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@221767 91177308-0d34-0410-b5e6-96231b3b80d8
2014-11-12 04:19:40 +00:00
; CHECK-DAG: xvmaddmdp 37, 35, 34
[PowerPC] Add subregister classes for f64 VSX values We had stored both f64 values and v2f64, etc. values in the VSX registers. This worked, but was suboptimal because we would always spill 16-byte values even through we almost always had scalar 8-byte values. This resulted in an increase in stack-size use, extra memory bandwidth, etc. To fix this, I've added 64-bit subregisters of the Altivec registers, and combined those with the existing scalar floating-point registers to form a class of VSX scalar floating-point registers. The ABI code has also been enhanced to use this register class and some other necessary improvements have been made. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205075 91177308-0d34-0410-b5e6-96231b3b80d8
2014-03-29 05:29:01 +00:00
; CHECK-DAG: li [[C3:[0-9]+]], 16
; Note: We could convert this next FMA to M-type as well, but it would require
; re-ordering the instructions.
; CHECK-DAG: xvmaddadp [[V1]], 35, 36
[PowerPC] VSX stores don't also read The VSX store instructions were also picking up an implicit "may read" from the default pattern, which was an intrinsic (and we don't currently have a way of specifying write-only intrinsics). This was causing MI verification to fail for VSX spill restores. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@227759 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-01 19:07:41 +00:00
; CHECK-DAG: xvmaddmdp 35, 36, 37
[PowerPC] Add vec_vsx_ld and vec_vsx_st intrinsics This patch enables the vec_vsx_ld and vec_vsx_st intrinsics for PowerPC, which provide programmer access to the lxvd2x, lxvw4x, stxvd2x, and stxvw4x instructions. New LLVM intrinsics are provided to represent these four instructions in IntrinsicsPowerPC.td. These are patterned after the similar intrinsics for lvx and stvx (Altivec). In PPCInstrVSX.td, these intrinsics are tied to the code gen patterns, with additional patterns to allow plain vanilla loads and stores to still generate these instructions. At -O1 and higher the intrinsics are immediately converted to loads and stores in InstCombineCalls.cpp. This will open up more optimization opportunities while still allowing the correct instructions to be generated. (Similar code exists for aligned Altivec loads and stores.) The new intrinsics are added to the code that checks for consecutive loads and stores in PPCISelLowering.cpp, as well as to PPCTargetLowering::getTgtMemIntrinsic(). There's a new test to verify the correct instructions are generated. The loads and stores tend to be reordered, so the test just counts their number. It runs at -O2, as it's not very effective to test this at -O0, when many unnecessary loads and stores are generated. I ended up having to modify vsx-fma-m.ll. It turns out this test case is slightly unreliable, but I don't know a good way to prevent problems with it. The xvmaddmdp instructions read and write the same register, which is one of the multiplicands. Commutativity allows either to be chosen. If the FMAs are reordered differently than expected by the test, the register assignment can be different as a result. Hopefully this doesn't change often. There is a companion patch for Clang. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@221767 91177308-0d34-0410-b5e6-96231b3b80d8
2014-11-12 04:19:40 +00:00
; CHECK-DAG: xvmaddadp 34, 35, 38
[PowerPC] Add subregister classes for f64 VSX values We had stored both f64 values and v2f64, etc. values in the VSX registers. This worked, but was suboptimal because we would always spill 16-byte values even through we almost always had scalar 8-byte values. This resulted in an increase in stack-size use, extra memory bandwidth, etc. To fix this, I've added 64-bit subregisters of the Altivec registers, and combined those with the existing scalar floating-point registers to form a class of VSX scalar floating-point registers. The ABI code has also been enhanced to use this register class and some other necessary improvements have been made. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205075 91177308-0d34-0410-b5e6-96231b3b80d8
2014-03-29 05:29:01 +00:00
; CHECK-DAG: stxvd2x 32, 0, 3
[PowerPC] VSX stores don't also read The VSX store instructions were also picking up an implicit "may read" from the default pattern, which was an intrinsic (and we don't currently have a way of specifying write-only intrinsics). This was causing MI verification to fail for VSX spill restores. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@227759 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-01 19:07:41 +00:00
; CHECK-DAG: stxvd2x 35, 3, [[C1]]
[PowerPC] Add subregister classes for f64 VSX values We had stored both f64 values and v2f64, etc. values in the VSX registers. This worked, but was suboptimal because we would always spill 16-byte values even through we almost always had scalar 8-byte values. This resulted in an increase in stack-size use, extra memory bandwidth, etc. To fix this, I've added 64-bit subregisters of the Altivec registers, and combined those with the existing scalar floating-point registers to form a class of VSX scalar floating-point registers. The ABI code has also been enhanced to use this register class and some other necessary improvements have been made. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205075 91177308-0d34-0410-b5e6-96231b3b80d8
2014-03-29 05:29:01 +00:00
; CHECK-DAG: stxvd2x 34, 3, [[C2]]
; CHECK-DAG: stxvd2x 37, 3, [[C3]]
; CHECK: blr
[PowerPC]Update Power VSX test cases to also test fast-isel Update of some of the VSX test cases for Power to check fast-isel codegen as well as the regular codegen. http://reviews.llvm.org/D6357 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@223509 91177308-0d34-0410-b5e6-96231b3b80d8
2014-12-05 20:32:05 +00:00
; CHECK-FISL-LABEL: @testv3
; CHECK-FISL-DAG: xxlor [[V1:[0-9]+]], 34, 34
; CHECK-FISL-DAG: xvmaddadp [[V1]], 35, 36
; CHECK-FISL-DAG: stxvd2x [[V1]], 0, 3
; CHECK-FISL-DAG: xxlor [[V2:[0-9]+]], 34, 34
; CHECK-FISL-DAG: xvmaddadp [[V2]], 35, 37
; CHECK-FISL-DAG: xxlor [[V3:[0-9]+]], 0, 0
; CHECK-FISL-DAG: xvmaddadp [[V3]], 35, 36
; CHECK-FISL-DAG: li [[C1:[0-9]+]], 48
; CHECK-FISL-DAG: stxvd2x [[V3]], 3, [[C1]]
; CHECK-FISL-DAG: xvmaddadp 34, 35, 38
; CHECK-FISL-DAG: li [[C2:[0-9]+]], 32
; CHECK-FISL-DAG: stxvd2x 34, 3, [[C2]]
; CHECK-FISL-DAG: li [[C3:[0-9]+]], 16
; CHECK-FISL-DAG: stxvd2x 0, 3, [[C3]]
; CHECK-FISL: blr
[PowerPC] Add subregister classes for f64 VSX values We had stored both f64 values and v2f64, etc. values in the VSX registers. This worked, but was suboptimal because we would always spill 16-byte values even through we almost always had scalar 8-byte values. This resulted in an increase in stack-size use, extra memory bandwidth, etc. To fix this, I've added 64-bit subregisters of the Altivec registers, and combined those with the existing scalar floating-point registers to form a class of VSX scalar floating-point registers. The ABI code has also been enhanced to use this register class and some other necessary improvements have been made. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205075 91177308-0d34-0410-b5e6-96231b3b80d8
2014-03-29 05:29:01 +00:00
}
define void @testv4(<2 x double> %a, <2 x double> %b, <2 x double> %c, <2 x double> %e, <2 x double> %f, <2 x double>* nocapture %d) #0 {
entry:
%0 = tail call <2 x double> @llvm.fma.v2f64(<2 x double> %b, <2 x double> %c, <2 x double> %a)
store <2 x double> %0, <2 x double>* %d, align 8
%1 = tail call <2 x double> @llvm.fma.v2f64(<2 x double> %b, <2 x double> %e, <2 x double> %a)
%arrayidx1 = getelementptr inbounds <2 x double>* %d, i64 1
store <2 x double> %1, <2 x double>* %arrayidx1, align 8
%2 = tail call <2 x double> @llvm.fma.v2f64(<2 x double> %b, <2 x double> %c, <2 x double> %1)
%arrayidx3 = getelementptr inbounds <2 x double>* %d, i64 3
store <2 x double> %2, <2 x double>* %arrayidx3, align 8
%3 = tail call <2 x double> @llvm.fma.v2f64(<2 x double> %b, <2 x double> %f, <2 x double> %a)
%arrayidx4 = getelementptr inbounds <2 x double>* %d, i64 2
store <2 x double> %3, <2 x double>* %arrayidx4, align 8
ret void
; CHECK-LABEL: @testv4
; CHECK-DAG: xxlor [[V1:[0-9]+]], 34, 34
; CHECK-DAG: xvmaddmdp 37, 35, 34
; CHECK-DAG: li [[C1:[0-9]+]], 16
; CHECK-DAG: li [[C2:[0-9]+]], 32
; CHECK-DAG: xvmaddadp 34, 35, 38
; Note: We could convert this next FMA to M-type as well, but it would require
; re-ordering the instructions.
; CHECK-DAG: xvmaddadp [[V1]], 35, 36
; CHECK-DAG: stxvd2x 32, 0, 3
; CHECK-DAG: stxvd2x 37, 3, [[C1]]
; CHECK-DAG: li [[C3:[0-9]+]], 48
; CHECK-DAG: xvmaddadp 37, 35, 36
; CHECK-DAG: stxvd2x 37, 3, [[C3]]
; CHECK-DAG: stxvd2x 34, 3, [[C2]]
; CHECK: blr
[PowerPC]Update Power VSX test cases to also test fast-isel Update of some of the VSX test cases for Power to check fast-isel codegen as well as the regular codegen. http://reviews.llvm.org/D6357 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@223509 91177308-0d34-0410-b5e6-96231b3b80d8
2014-12-05 20:32:05 +00:00
; CHECK-FISL-LABEL: @testv4
; CHECK-FISL-DAG: xxlor [[V1:[0-9]+]], 34, 34
; CHECK-FISL-DAG: xvmaddadp [[V1]], 35, 36
; CHECK-FISL-DAG: stxvd2x 0, 0, 3
; CHECK-FISL-DAG: xxlor [[V2:[0-9]+]], 34, 34
; CHECK-FISL-DAG: xvmaddadp [[V2]], 35, 37
; CHECK-FISL-DAG: li [[C1:[0-9]+]], 16
; CHECK-FISL-DAG: stxvd2x 0, 3, [[C1]]
; CHECK-FISL-DAG: xvmaddadp 0, 35, 37
; CHECK-FISL-DAG: li [[C3:[0-9]+]], 48
; CHECK-FISL-DAG: stxvd2x 0, 3, [[C3]]
; CHECK-FISL-DAG: xvmaddadp 0, 35, 36
; CHECK-FISL-DAG: li [[C2:[0-9]+]], 32
; CHECK-FISL-DAG: stxvd2x 34, 3, [[C2]]
; CHECK-FISL: blr
[PowerPC] Add subregister classes for f64 VSX values We had stored both f64 values and v2f64, etc. values in the VSX registers. This worked, but was suboptimal because we would always spill 16-byte values even through we almost always had scalar 8-byte values. This resulted in an increase in stack-size use, extra memory bandwidth, etc. To fix this, I've added 64-bit subregisters of the Altivec registers, and combined those with the existing scalar floating-point registers to form a class of VSX scalar floating-point registers. The ABI code has also been enhanced to use this register class and some other necessary improvements have been made. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205075 91177308-0d34-0410-b5e6-96231b3b80d8
2014-03-29 05:29:01 +00:00
}
declare <2 x double> @llvm.fma.v2f64(<2 x double>, <2 x double>, <2 x double>) #0
[PowerPC] Select between VSX A-type and M-type FMA instructions just before RA The VSX instruction set has two types of FMA instructions: A-type (where the addend is taken from the output register) and M-type (where one of the product operands is taken from the output register). This adds a small pass that runs just after MI scheduling (and, thus, just before register allocation) that mutates A-type instructions (that are created during isel) into M-type instructions when: 1. This will eliminate an otherwise-necessary copy of the addend 2. One of the product operands is killed by the instruction The "right" moment to make this decision is in between scheduling and register allocation, because only there do we know whether or not one of the product operands is killed by any particular instruction. Unfortunately, this also makes the implementation somewhat complicated, because the MIs are not in SSA form and we need to preserve the LiveIntervals analysis. As a simple example, if we have: %vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9 %vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16, %RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16 ... %vreg9<def,tied1> = XSMADDADP %vreg9<tied0>, %vreg17, %vreg19, %RM<imp-use>; VSLRC:%vreg9,%vreg17,%vreg19 ... We can eliminate the copy by changing from the A-type to the M-type instruction. This means: %vreg5<def,tied1> = XSMADDADP %vreg5<tied0>, %vreg17, %vreg16, %RM<imp-use>; VSLRC:%vreg5,%vreg17,%vreg16 is replaced by: %vreg16<def,tied1> = XSMADDMDP %vreg16<tied0>, %vreg18, %vreg9, %RM<imp-use>; VSLRC:%vreg16,%vreg18,%vreg9 and we remove: %vreg5<def> = COPY %vreg9; VSLRC:%vreg5,%vreg9 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204768 91177308-0d34-0410-b5e6-96231b3b80d8
2014-03-25 23:29:21 +00:00
attributes #0 = { nounwind readnone }
Reference in New Issue Copy Permalink