2013-10-30 19:05:41 +00:00
|
|
|
; RUN: opt < %s -default-data-layout="e-p:32:32:32-p1:16:16:16-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-n8:16:32" -basicaa -gvn -S -die | FileCheck %s
|
|
|
|
|
; RUN: opt < %s -default-data-layout="E-p:32:32:32-p1:16:16:16-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:64:64-n32" -basicaa -gvn -S -die | FileCheck %s
|
2009-09-20 19:03:47 +00:00
|
|
|
|
|
|
|
|
;; Trivial RLE test.
|
|
|
|
|
define i32 @test0(i32 %V, i32* %P) {
|
|
|
|
|
store i32 %V, i32* %P
|
|
|
|
|
|
2015-02-27 21:17:42 +00:00
|
|
|
%A = load i32, i32* %P
|
2009-09-20 19:03:47 +00:00
|
|
|
ret i32 %A
|
2013-07-14 01:42:54 +00:00
|
|
|
; CHECK-LABEL: @test0(
|
2009-09-20 19:03:47 +00:00
|
|
|
; CHECK: ret i32 %V
|
|
|
|
|
}
|
|
|
|
|
|
2009-09-21 17:24:04 +00:00
|
|
|
|
|
|
|
|
;;===----------------------------------------------------------------------===;;
|
|
|
|
|
;; Tests for crashers
|
|
|
|
|
;;===----------------------------------------------------------------------===;;
|
|
|
|
|
|
|
|
|
|
;; PR5016
|
|
|
|
|
define i8 @crash0({i32, i32} %A, {i32, i32}* %P) {
|
|
|
|
|
store {i32, i32} %A, {i32, i32}* %P
|
|
|
|
|
%X = bitcast {i32, i32}* %P to i8*
|
2015-02-27 21:17:42 +00:00
|
|
|
%Y = load i8, i8* %X
|
2009-09-21 17:24:04 +00:00
|
|
|
ret i8 %Y
|
|
|
|
|
}
|
|
|
|
|
|
2011-11-21 19:42:56 +00:00
|
|
|
;; No PR filed, crashed in CaptureTracker.
|
|
|
|
|
declare void @helper()
|
|
|
|
|
define void @crash1() {
|
|
|
|
|
tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* undef, i8* undef, i64 undef, i32 1, i1 false) nounwind
|
2015-02-27 21:17:42 +00:00
|
|
|
%tmp = load i8, i8* bitcast (void ()* @helper to i8*)
|
2011-11-21 19:42:56 +00:00
|
|
|
%x = icmp eq i8 %tmp, 15
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
2009-09-21 17:24:04 +00:00
|
|
|
|
2009-09-20 19:03:47 +00:00
|
|
|
;;===----------------------------------------------------------------------===;;
|
|
|
|
|
;; Store -> Load and Load -> Load forwarding where src and dst are different
|
|
|
|
|
;; types, but where the base pointer is a must alias.
|
|
|
|
|
;;===----------------------------------------------------------------------===;;
|
|
|
|
|
|
|
|
|
|
;; i32 -> f32 forwarding.
|
|
|
|
|
define float @coerce_mustalias1(i32 %V, i32* %P) {
|
|
|
|
|
store i32 %V, i32* %P
|
|
|
|
|
|
|
|
|
|
%P2 = bitcast i32* %P to float*
|
|
|
|
|
|
2015-02-27 21:17:42 +00:00
|
|
|
%A = load float, float* %P2
|
2009-09-20 19:03:47 +00:00
|
|
|
ret float %A
|
2013-07-14 01:42:54 +00:00
|
|
|
; CHECK-LABEL: @coerce_mustalias1(
|
2009-09-20 19:03:47 +00:00
|
|
|
; CHECK-NOT: load
|
|
|
|
|
; CHECK: ret float
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
;; i32* -> float forwarding.
|
|
|
|
|
define float @coerce_mustalias2(i32* %V, i32** %P) {
|
|
|
|
|
store i32* %V, i32** %P
|
|
|
|
|
|
|
|
|
|
%P2 = bitcast i32** %P to float*
|
|
|
|
|
|
2015-02-27 21:17:42 +00:00
|
|
|
%A = load float, float* %P2
|
2009-09-20 19:03:47 +00:00
|
|
|
ret float %A
|
2013-07-14 01:42:54 +00:00
|
|
|
; CHECK-LABEL: @coerce_mustalias2(
|
2009-09-20 19:03:47 +00:00
|
|
|
; CHECK-NOT: load
|
|
|
|
|
; CHECK: ret float
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
;; float -> i32* forwarding.
|
|
|
|
|
define i32* @coerce_mustalias3(float %V, float* %P) {
|
|
|
|
|
store float %V, float* %P
|
|
|
|
|
|
|
|
|
|
%P2 = bitcast float* %P to i32**
|
|
|
|
|
|
2015-02-27 21:17:42 +00:00
|
|
|
%A = load i32*, i32** %P2
|
2009-09-20 19:03:47 +00:00
|
|
|
ret i32* %A
|
2013-07-14 01:42:54 +00:00
|
|
|
; CHECK-LABEL: @coerce_mustalias3(
|
2009-09-20 19:03:47 +00:00
|
|
|
; CHECK-NOT: load
|
|
|
|
|
; CHECK: ret i32*
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
;; i32 -> f32 load forwarding.
|
|
|
|
|
define float @coerce_mustalias4(i32* %P, i1 %cond) {
|
2015-02-27 21:17:42 +00:00
|
|
|
%A = load i32, i32* %P
|
2009-09-20 20:09:34 +00:00
|
|
|
|
2009-09-20 19:03:47 +00:00
|
|
|
%P2 = bitcast i32* %P to float*
|
2015-02-27 21:17:42 +00:00
|
|
|
%B = load float, float* %P2
|
2009-09-20 20:09:34 +00:00
|
|
|
br i1 %cond, label %T, label %F
|
|
|
|
|
T:
|
2009-09-20 19:03:47 +00:00
|
|
|
ret float %B
|
|
|
|
|
|
|
|
|
|
F:
|
|
|
|
|
%X = bitcast i32 %A to float
|
|
|
|
|
ret float %X
|
|
|
|
|
|
2013-07-14 01:42:54 +00:00
|
|
|
; CHECK-LABEL: @coerce_mustalias4(
|
2015-02-27 21:17:42 +00:00
|
|
|
; CHECK: %A = load i32, i32* %P
|
2009-09-20 19:03:47 +00:00
|
|
|
; CHECK-NOT: load
|
|
|
|
|
; CHECK: ret float
|
|
|
|
|
; CHECK: F:
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
;; i32 -> i8 forwarding
|
|
|
|
|
define i8 @coerce_mustalias5(i32 %V, i32* %P) {
|
|
|
|
|
store i32 %V, i32* %P
|
|
|
|
|
|
|
|
|
|
%P2 = bitcast i32* %P to i8*
|
|
|
|
|
|
2015-02-27 21:17:42 +00:00
|
|
|
%A = load i8, i8* %P2
|
2009-09-20 19:03:47 +00:00
|
|
|
ret i8 %A
|
2013-07-14 01:42:54 +00:00
|
|
|
; CHECK-LABEL: @coerce_mustalias5(
|
2009-09-20 19:03:47 +00:00
|
|
|
; CHECK-NOT: load
|
|
|
|
|
; CHECK: ret i8
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
;; i64 -> float forwarding
|
|
|
|
|
define float @coerce_mustalias6(i64 %V, i64* %P) {
|
|
|
|
|
store i64 %V, i64* %P
|
|
|
|
|
|
|
|
|
|
%P2 = bitcast i64* %P to float*
|
|
|
|
|
|
2015-02-27 21:17:42 +00:00
|
|
|
%A = load float, float* %P2
|
2009-09-20 19:03:47 +00:00
|
|
|
ret float %A
|
2013-07-14 01:42:54 +00:00
|
|
|
; CHECK-LABEL: @coerce_mustalias6(
|
2009-09-20 19:03:47 +00:00
|
|
|
; CHECK-NOT: load
|
|
|
|
|
; CHECK: ret float
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
;; i64 -> i8* (32-bit) forwarding
|
|
|
|
|
define i8* @coerce_mustalias7(i64 %V, i64* %P) {
|
|
|
|
|
store i64 %V, i64* %P
|
|
|
|
|
|
|
|
|
|
%P2 = bitcast i64* %P to i8**
|
|
|
|
|
|
2015-02-27 21:17:42 +00:00
|
|
|
%A = load i8*, i8** %P2
|
2009-09-20 19:03:47 +00:00
|
|
|
ret i8* %A
|
2013-07-14 01:42:54 +00:00
|
|
|
; CHECK-LABEL: @coerce_mustalias7(
|
2009-09-20 19:03:47 +00:00
|
|
|
; CHECK-NOT: load
|
|
|
|
|
; CHECK: ret i8*
|
|
|
|
|
}
|
|
|
|
|
|
2009-12-06 01:57:02 +00:00
|
|
|
; memset -> i16 forwarding.
|
|
|
|
|
define signext i16 @memset_to_i16_local(i16* %A) nounwind ssp {
|
|
|
|
|
entry:
|
|
|
|
|
%conv = bitcast i16* %A to i8*
|
2011-06-18 06:05:24 +00:00
|
|
|
tail call void @llvm.memset.p0i8.i64(i8* %conv, i8 1, i64 200, i32 1, i1 false)
|
[opaque pointer type] Add textual IR support for explicit type parameter to getelementptr instruction
One of several parallel first steps to remove the target type of pointers,
replacing them with a single opaque pointer type.
This adds an explicit type parameter to the gep instruction so that when the
first parameter becomes an opaque pointer type, the type to gep through is
still available to the instructions.
* This doesn't modify gep operators, only instructions (operators will be
handled separately)
* Textual IR changes only. Bitcode (including upgrade) and changing the
in-memory representation will be in separate changes.
* geps of vectors are transformed as:
getelementptr <4 x float*> %x, ...
->getelementptr float, <4 x float*> %x, ...
Then, once the opaque pointer type is introduced, this will ultimately look
like:
getelementptr float, <4 x ptr> %x
with the unambiguous interpretation that it is a vector of pointers to float.
* address spaces remain on the pointer, not the type:
getelementptr float addrspace(1)* %x
->getelementptr float, float addrspace(1)* %x
Then, eventually:
getelementptr float, ptr addrspace(1) %x
Importantly, the massive amount of test case churn has been automated by
same crappy python code. I had to manually update a few test cases that
wouldn't fit the script's model (r228970,r229196,r229197,r229198). The
python script just massages stdin and writes the result to stdout, I
then wrapped that in a shell script to handle replacing files, then
using the usual find+xargs to migrate all the files.
update.py:
import fileinput
import sys
import re
ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
def conv(match, line):
if not match:
return line
line = match.groups()[0]
if len(match.groups()[5]) == 0:
line += match.groups()[2]
line += match.groups()[3]
line += ", "
line += match.groups()[1]
line += "\n"
return line
for line in sys.stdin:
if line.find("getelementptr ") == line.find("getelementptr inbounds"):
if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("):
line = conv(re.match(ibrep, line), line)
elif line.find("getelementptr ") != line.find("getelementptr ("):
line = conv(re.match(normrep, line), line)
sys.stdout.write(line)
apply.sh:
for name in "$@"
do
python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name"
rm -f "$name.tmp"
done
The actual commands:
From llvm/src:
find test/ -name *.ll | xargs ./apply.sh
From llvm/src/tools/clang:
find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}"
From llvm/src/tools/polly:
find test/ -name *.ll | xargs ./apply.sh
After that, check-all (with llvm, clang, clang-tools-extra, lld,
compiler-rt, and polly all checked out).
The extra 'rm' in the apply.sh script is due to a few files in clang's test
suite using interesting unicode stuff that my python script was throwing
exceptions on. None of those files needed to be migrated, so it seemed
sufficient to ignore those cases.
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7636
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230786 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 19:29:02 +00:00
|
|
|
%arrayidx = getelementptr inbounds i16, i16* %A, i64 42
|
2015-02-27 21:17:42 +00:00
|
|
|
%tmp2 = load i16, i16* %arrayidx
|
2009-12-06 01:57:02 +00:00
|
|
|
ret i16 %tmp2
|
2013-07-14 01:42:54 +00:00
|
|
|
; CHECK-LABEL: @memset_to_i16_local(
|
2009-12-06 01:57:02 +00:00
|
|
|
; CHECK-NOT: load
|
|
|
|
|
; CHECK: ret i16 257
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; memset -> float forwarding.
|
|
|
|
|
define float @memset_to_float_local(float* %A, i8 %Val) nounwind ssp {
|
|
|
|
|
entry:
|
|
|
|
|
%conv = bitcast float* %A to i8* ; <i8*> [#uses=1]
|
2011-06-18 06:05:24 +00:00
|
|
|
tail call void @llvm.memset.p0i8.i64(i8* %conv, i8 %Val, i64 400, i32 1, i1 false)
|
[opaque pointer type] Add textual IR support for explicit type parameter to getelementptr instruction
One of several parallel first steps to remove the target type of pointers,
replacing them with a single opaque pointer type.
This adds an explicit type parameter to the gep instruction so that when the
first parameter becomes an opaque pointer type, the type to gep through is
still available to the instructions.
* This doesn't modify gep operators, only instructions (operators will be
handled separately)
* Textual IR changes only. Bitcode (including upgrade) and changing the
in-memory representation will be in separate changes.
* geps of vectors are transformed as:
getelementptr <4 x float*> %x, ...
->getelementptr float, <4 x float*> %x, ...
Then, once the opaque pointer type is introduced, this will ultimately look
like:
getelementptr float, <4 x ptr> %x
with the unambiguous interpretation that it is a vector of pointers to float.
* address spaces remain on the pointer, not the type:
getelementptr float addrspace(1)* %x
->getelementptr float, float addrspace(1)* %x
Then, eventually:
getelementptr float, ptr addrspace(1) %x
Importantly, the massive amount of test case churn has been automated by
same crappy python code. I had to manually update a few test cases that
wouldn't fit the script's model (r228970,r229196,r229197,r229198). The
python script just massages stdin and writes the result to stdout, I
then wrapped that in a shell script to handle replacing files, then
using the usual find+xargs to migrate all the files.
update.py:
import fileinput
import sys
import re
ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
def conv(match, line):
if not match:
return line
line = match.groups()[0]
if len(match.groups()[5]) == 0:
line += match.groups()[2]
line += match.groups()[3]
line += ", "
line += match.groups()[1]
line += "\n"
return line
for line in sys.stdin:
if line.find("getelementptr ") == line.find("getelementptr inbounds"):
if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("):
line = conv(re.match(ibrep, line), line)
elif line.find("getelementptr ") != line.find("getelementptr ("):
line = conv(re.match(normrep, line), line)
sys.stdout.write(line)
apply.sh:
for name in "$@"
do
python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name"
rm -f "$name.tmp"
done
The actual commands:
From llvm/src:
find test/ -name *.ll | xargs ./apply.sh
From llvm/src/tools/clang:
find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}"
From llvm/src/tools/polly:
find test/ -name *.ll | xargs ./apply.sh
After that, check-all (with llvm, clang, clang-tools-extra, lld,
compiler-rt, and polly all checked out).
The extra 'rm' in the apply.sh script is due to a few files in clang's test
suite using interesting unicode stuff that my python script was throwing
exceptions on. None of those files needed to be migrated, so it seemed
sufficient to ignore those cases.
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7636
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230786 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 19:29:02 +00:00
|
|
|
%arrayidx = getelementptr inbounds float, float* %A, i64 42 ; <float*> [#uses=1]
|
2015-02-27 21:17:42 +00:00
|
|
|
%tmp2 = load float, float* %arrayidx ; <float> [#uses=1]
|
2009-12-06 01:57:02 +00:00
|
|
|
ret float %tmp2
|
2013-07-14 01:42:54 +00:00
|
|
|
; CHECK-LABEL: @memset_to_float_local(
|
2009-12-06 01:57:02 +00:00
|
|
|
; CHECK-NOT: load
|
|
|
|
|
; CHECK: zext
|
|
|
|
|
; CHECK-NEXT: shl
|
|
|
|
|
; CHECK-NEXT: or
|
|
|
|
|
; CHECK-NEXT: shl
|
|
|
|
|
; CHECK-NEXT: or
|
|
|
|
|
; CHECK-NEXT: bitcast
|
|
|
|
|
; CHECK-NEXT: ret float
|
|
|
|
|
}
|
|
|
|
|
|
2009-12-06 04:54:31 +00:00
|
|
|
;; non-local memset -> i16 load forwarding.
|
|
|
|
|
define i16 @memset_to_i16_nonlocal0(i16* %P, i1 %cond) {
|
|
|
|
|
%P3 = bitcast i16* %P to i8*
|
|
|
|
|
br i1 %cond, label %T, label %F
|
|
|
|
|
T:
|
2011-06-18 06:05:24 +00:00
|
|
|
tail call void @llvm.memset.p0i8.i64(i8* %P3, i8 1, i64 400, i32 1, i1 false)
|
2009-12-06 04:54:31 +00:00
|
|
|
br label %Cont
|
|
|
|
|
|
|
|
|
|
F:
|
2011-06-18 06:05:24 +00:00
|
|
|
tail call void @llvm.memset.p0i8.i64(i8* %P3, i8 2, i64 400, i32 1, i1 false)
|
2009-12-06 04:54:31 +00:00
|
|
|
br label %Cont
|
|
|
|
|
|
|
|
|
|
Cont:
|
[opaque pointer type] Add textual IR support for explicit type parameter to getelementptr instruction
One of several parallel first steps to remove the target type of pointers,
replacing them with a single opaque pointer type.
This adds an explicit type parameter to the gep instruction so that when the
first parameter becomes an opaque pointer type, the type to gep through is
still available to the instructions.
* This doesn't modify gep operators, only instructions (operators will be
handled separately)
* Textual IR changes only. Bitcode (including upgrade) and changing the
in-memory representation will be in separate changes.
* geps of vectors are transformed as:
getelementptr <4 x float*> %x, ...
->getelementptr float, <4 x float*> %x, ...
Then, once the opaque pointer type is introduced, this will ultimately look
like:
getelementptr float, <4 x ptr> %x
with the unambiguous interpretation that it is a vector of pointers to float.
* address spaces remain on the pointer, not the type:
getelementptr float addrspace(1)* %x
->getelementptr float, float addrspace(1)* %x
Then, eventually:
getelementptr float, ptr addrspace(1) %x
Importantly, the massive amount of test case churn has been automated by
same crappy python code. I had to manually update a few test cases that
wouldn't fit the script's model (r228970,r229196,r229197,r229198). The
python script just massages stdin and writes the result to stdout, I
then wrapped that in a shell script to handle replacing files, then
using the usual find+xargs to migrate all the files.
update.py:
import fileinput
import sys
import re
ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
def conv(match, line):
if not match:
return line
line = match.groups()[0]
if len(match.groups()[5]) == 0:
line += match.groups()[2]
line += match.groups()[3]
line += ", "
line += match.groups()[1]
line += "\n"
return line
for line in sys.stdin:
if line.find("getelementptr ") == line.find("getelementptr inbounds"):
if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("):
line = conv(re.match(ibrep, line), line)
elif line.find("getelementptr ") != line.find("getelementptr ("):
line = conv(re.match(normrep, line), line)
sys.stdout.write(line)
apply.sh:
for name in "$@"
do
python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name"
rm -f "$name.tmp"
done
The actual commands:
From llvm/src:
find test/ -name *.ll | xargs ./apply.sh
From llvm/src/tools/clang:
find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}"
From llvm/src/tools/polly:
find test/ -name *.ll | xargs ./apply.sh
After that, check-all (with llvm, clang, clang-tools-extra, lld,
compiler-rt, and polly all checked out).
The extra 'rm' in the apply.sh script is due to a few files in clang's test
suite using interesting unicode stuff that my python script was throwing
exceptions on. None of those files needed to be migrated, so it seemed
sufficient to ignore those cases.
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7636
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230786 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 19:29:02 +00:00
|
|
|
%P2 = getelementptr i16, i16* %P, i32 4
|
2015-02-27 21:17:42 +00:00
|
|
|
%A = load i16, i16* %P2
|
2009-12-06 04:54:31 +00:00
|
|
|
ret i16 %A
|
|
|
|
|
|
2013-07-14 01:42:54 +00:00
|
|
|
; CHECK-LABEL: @memset_to_i16_nonlocal0(
|
2009-12-06 04:54:31 +00:00
|
|
|
; CHECK: Cont:
|
|
|
|
|
; CHECK-NEXT: %A = phi i16 [ 514, %F ], [ 257, %T ]
|
|
|
|
|
; CHECK-NOT: load
|
|
|
|
|
; CHECK: ret i16 %A
|
|
|
|
|
}
|
|
|
|
|
|
2009-12-06 05:29:56 +00:00
|
|
|
@GCst = constant {i32, float, i32 } { i32 42, float 14., i32 97 }
|
2013-10-30 19:05:41 +00:00
|
|
|
@GCst_as1 = addrspace(1) constant {i32, float, i32 } { i32 42, float 14., i32 97 }
|
2009-12-06 05:29:56 +00:00
|
|
|
|
|
|
|
|
; memset -> float forwarding.
|
|
|
|
|
define float @memcpy_to_float_local(float* %A) nounwind ssp {
|
|
|
|
|
entry:
|
|
|
|
|
%conv = bitcast float* %A to i8* ; <i8*> [#uses=1]
|
2011-06-18 06:05:24 +00:00
|
|
|
tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %conv, i8* bitcast ({i32, float, i32 }* @GCst to i8*), i64 12, i32 1, i1 false)
|
[opaque pointer type] Add textual IR support for explicit type parameter to getelementptr instruction
One of several parallel first steps to remove the target type of pointers,
replacing them with a single opaque pointer type.
This adds an explicit type parameter to the gep instruction so that when the
first parameter becomes an opaque pointer type, the type to gep through is
still available to the instructions.
* This doesn't modify gep operators, only instructions (operators will be
handled separately)
* Textual IR changes only. Bitcode (including upgrade) and changing the
in-memory representation will be in separate changes.
* geps of vectors are transformed as:
getelementptr <4 x float*> %x, ...
->getelementptr float, <4 x float*> %x, ...
Then, once the opaque pointer type is introduced, this will ultimately look
like:
getelementptr float, <4 x ptr> %x
with the unambiguous interpretation that it is a vector of pointers to float.
* address spaces remain on the pointer, not the type:
getelementptr float addrspace(1)* %x
->getelementptr float, float addrspace(1)* %x
Then, eventually:
getelementptr float, ptr addrspace(1) %x
Importantly, the massive amount of test case churn has been automated by
same crappy python code. I had to manually update a few test cases that
wouldn't fit the script's model (r228970,r229196,r229197,r229198). The
python script just massages stdin and writes the result to stdout, I
then wrapped that in a shell script to handle replacing files, then
using the usual find+xargs to migrate all the files.
update.py:
import fileinput
import sys
import re
ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
def conv(match, line):
if not match:
return line
line = match.groups()[0]
if len(match.groups()[5]) == 0:
line += match.groups()[2]
line += match.groups()[3]
line += ", "
line += match.groups()[1]
line += "\n"
return line
for line in sys.stdin:
if line.find("getelementptr ") == line.find("getelementptr inbounds"):
if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("):
line = conv(re.match(ibrep, line), line)
elif line.find("getelementptr ") != line.find("getelementptr ("):
line = conv(re.match(normrep, line), line)
sys.stdout.write(line)
apply.sh:
for name in "$@"
do
python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name"
rm -f "$name.tmp"
done
The actual commands:
From llvm/src:
find test/ -name *.ll | xargs ./apply.sh
From llvm/src/tools/clang:
find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}"
From llvm/src/tools/polly:
find test/ -name *.ll | xargs ./apply.sh
After that, check-all (with llvm, clang, clang-tools-extra, lld,
compiler-rt, and polly all checked out).
The extra 'rm' in the apply.sh script is due to a few files in clang's test
suite using interesting unicode stuff that my python script was throwing
exceptions on. None of those files needed to be migrated, so it seemed
sufficient to ignore those cases.
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7636
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230786 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 19:29:02 +00:00
|
|
|
%arrayidx = getelementptr inbounds float, float* %A, i64 1 ; <float*> [#uses=1]
|
2015-02-27 21:17:42 +00:00
|
|
|
%tmp2 = load float, float* %arrayidx ; <float> [#uses=1]
|
2009-12-06 05:29:56 +00:00
|
|
|
ret float %tmp2
|
2013-07-14 01:42:54 +00:00
|
|
|
; CHECK-LABEL: @memcpy_to_float_local(
|
2009-12-06 05:29:56 +00:00
|
|
|
; CHECK-NOT: load
|
|
|
|
|
; CHECK: ret float 1.400000e+01
|
|
|
|
|
}
|
|
|
|
|
|
2013-10-30 19:05:41 +00:00
|
|
|
; memcpy from address space 1
|
|
|
|
|
define float @memcpy_to_float_local_as1(float* %A) nounwind ssp {
|
|
|
|
|
entry:
|
|
|
|
|
%conv = bitcast float* %A to i8* ; <i8*> [#uses=1]
|
|
|
|
|
tail call void @llvm.memcpy.p0i8.p1i8.i64(i8* %conv, i8 addrspace(1)* bitcast ({i32, float, i32 } addrspace(1)* @GCst_as1 to i8 addrspace(1)*), i64 12, i32 1, i1 false)
|
[opaque pointer type] Add textual IR support for explicit type parameter to getelementptr instruction
One of several parallel first steps to remove the target type of pointers,
replacing them with a single opaque pointer type.
This adds an explicit type parameter to the gep instruction so that when the
first parameter becomes an opaque pointer type, the type to gep through is
still available to the instructions.
* This doesn't modify gep operators, only instructions (operators will be
handled separately)
* Textual IR changes only. Bitcode (including upgrade) and changing the
in-memory representation will be in separate changes.
* geps of vectors are transformed as:
getelementptr <4 x float*> %x, ...
->getelementptr float, <4 x float*> %x, ...
Then, once the opaque pointer type is introduced, this will ultimately look
like:
getelementptr float, <4 x ptr> %x
with the unambiguous interpretation that it is a vector of pointers to float.
* address spaces remain on the pointer, not the type:
getelementptr float addrspace(1)* %x
->getelementptr float, float addrspace(1)* %x
Then, eventually:
getelementptr float, ptr addrspace(1) %x
Importantly, the massive amount of test case churn has been automated by
same crappy python code. I had to manually update a few test cases that
wouldn't fit the script's model (r228970,r229196,r229197,r229198). The
python script just massages stdin and writes the result to stdout, I
then wrapped that in a shell script to handle replacing files, then
using the usual find+xargs to migrate all the files.
update.py:
import fileinput
import sys
import re
ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
def conv(match, line):
if not match:
return line
line = match.groups()[0]
if len(match.groups()[5]) == 0:
line += match.groups()[2]
line += match.groups()[3]
line += ", "
line += match.groups()[1]
line += "\n"
return line
for line in sys.stdin:
if line.find("getelementptr ") == line.find("getelementptr inbounds"):
if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("):
line = conv(re.match(ibrep, line), line)
elif line.find("getelementptr ") != line.find("getelementptr ("):
line = conv(re.match(normrep, line), line)
sys.stdout.write(line)
apply.sh:
for name in "$@"
do
python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name"
rm -f "$name.tmp"
done
The actual commands:
From llvm/src:
find test/ -name *.ll | xargs ./apply.sh
From llvm/src/tools/clang:
find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}"
From llvm/src/tools/polly:
find test/ -name *.ll | xargs ./apply.sh
After that, check-all (with llvm, clang, clang-tools-extra, lld,
compiler-rt, and polly all checked out).
The extra 'rm' in the apply.sh script is due to a few files in clang's test
suite using interesting unicode stuff that my python script was throwing
exceptions on. None of those files needed to be migrated, so it seemed
sufficient to ignore those cases.
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7636
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230786 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 19:29:02 +00:00
|
|
|
%arrayidx = getelementptr inbounds float, float* %A, i64 1 ; <float*> [#uses=1]
|
2015-02-27 21:17:42 +00:00
|
|
|
%tmp2 = load float, float* %arrayidx ; <float> [#uses=1]
|
2013-10-30 19:05:41 +00:00
|
|
|
ret float %tmp2
|
|
|
|
|
; CHECK-LABEL: @memcpy_to_float_local_as1(
|
|
|
|
|
; CHECK-NOT: load
|
|
|
|
|
; CHECK: ret float 1.400000e+01
|
|
|
|
|
}
|
2009-12-06 01:57:02 +00:00
|
|
|
|
2009-09-20 20:09:34 +00:00
|
|
|
;; non-local i32/float -> i8 load forwarding.
|
|
|
|
|
define i8 @coerce_mustalias_nonlocal0(i32* %P, i1 %cond) {
|
|
|
|
|
%P2 = bitcast i32* %P to float*
|
|
|
|
|
%P3 = bitcast i32* %P to i8*
|
|
|
|
|
br i1 %cond, label %T, label %F
|
|
|
|
|
T:
|
|
|
|
|
store i32 42, i32* %P
|
|
|
|
|
br label %Cont
|
|
|
|
|
|
|
|
|
|
F:
|
|
|
|
|
store float 1.0, float* %P2
|
|
|
|
|
br label %Cont
|
|
|
|
|
|
|
|
|
|
Cont:
|
2015-02-27 21:17:42 +00:00
|
|
|
%A = load i8, i8* %P3
|
2009-09-20 20:09:34 +00:00
|
|
|
ret i8 %A
|
|
|
|
|
|
2013-07-14 01:42:54 +00:00
|
|
|
; CHECK-LABEL: @coerce_mustalias_nonlocal0(
|
2009-09-20 20:09:34 +00:00
|
|
|
; CHECK: Cont:
|
|
|
|
|
; CHECK: %A = phi i8 [
|
|
|
|
|
; CHECK-NOT: load
|
|
|
|
|
; CHECK: ret i8 %A
|
|
|
|
|
}
|
|
|
|
|
|
2009-12-06 04:54:31 +00:00
|
|
|
|
Improve GVN to be able to forward substitute a small load
from a piece of a large store when both are in the same block.
This allows clang to compile the testcase in PR4216 to this code:
_test_bitfield:
movl 4(%esp), %eax
movl %eax, %ecx
andl $-65536, %ecx
orl $32962, %eax
andl $40186, %eax
orl %ecx, %eax
ret
This is not ideal, but is a whole lot better than the code produced
by llvm-gcc:
_test_bitfield:
movw $-32574, %ax
orw 4(%esp), %ax
andw $-25350, %ax
movw %ax, 4(%esp)
movw 7(%esp), %cx
shlw $8, %cx
movzbl 6(%esp), %edx
orw %cx, %dx
movzwl %dx, %ecx
shll $16, %ecx
movzwl %ax, %eax
orl %ecx, %eax
ret
and dramatically better than that produced by gcc 4.2:
_test_bitfield:
pushl %ebx
call L3
"L00000000001$pb":
L3:
popl %ebx
movl 8(%esp), %eax
leal 0(,%eax,4), %edx
sarb $7, %dl
movl %eax, %ecx
andl $7168, %ecx
andl $-7201, %ebx
movzbl %dl, %edx
andl $1, %edx
sall $5, %edx
orl %ecx, %ebx
orl %edx, %ebx
andl $24, %eax
andl $-58336, %ebx
orl %eax, %ebx
orl $32962, %ebx
movl %ebx, %eax
popl %ebx
ret
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82439 91177308-0d34-0410-b5e6-96231b3b80d8
2009-09-21 05:57:11 +00:00
|
|
|
;; non-local i32/float -> i8 load forwarding. This also tests that the "P3"
|
|
|
|
|
;; bitcast equivalence can be properly phi translated.
|
|
|
|
|
define i8 @coerce_mustalias_nonlocal1(i32* %P, i1 %cond) {
|
|
|
|
|
%P2 = bitcast i32* %P to float*
|
|
|
|
|
br i1 %cond, label %T, label %F
|
|
|
|
|
T:
|
|
|
|
|
store i32 42, i32* %P
|
|
|
|
|
br label %Cont
|
|
|
|
|
|
|
|
|
|
F:
|
|
|
|
|
store float 1.0, float* %P2
|
|
|
|
|
br label %Cont
|
|
|
|
|
|
|
|
|
|
Cont:
|
|
|
|
|
%P3 = bitcast i32* %P to i8*
|
2015-02-27 21:17:42 +00:00
|
|
|
%A = load i8, i8* %P3
|
Improve GVN to be able to forward substitute a small load
from a piece of a large store when both are in the same block.
This allows clang to compile the testcase in PR4216 to this code:
_test_bitfield:
movl 4(%esp), %eax
movl %eax, %ecx
andl $-65536, %ecx
orl $32962, %eax
andl $40186, %eax
orl %ecx, %eax
ret
This is not ideal, but is a whole lot better than the code produced
by llvm-gcc:
_test_bitfield:
movw $-32574, %ax
orw 4(%esp), %ax
andw $-25350, %ax
movw %ax, 4(%esp)
movw 7(%esp), %cx
shlw $8, %cx
movzbl 6(%esp), %edx
orw %cx, %dx
movzwl %dx, %ecx
shll $16, %ecx
movzwl %ax, %eax
orl %ecx, %eax
ret
and dramatically better than that produced by gcc 4.2:
_test_bitfield:
pushl %ebx
call L3
"L00000000001$pb":
L3:
popl %ebx
movl 8(%esp), %eax
leal 0(,%eax,4), %edx
sarb $7, %dl
movl %eax, %ecx
andl $7168, %ecx
andl $-7201, %ebx
movzbl %dl, %edx
andl $1, %edx
sall $5, %edx
orl %ecx, %ebx
orl %edx, %ebx
andl $24, %eax
andl $-58336, %ebx
orl %eax, %ebx
orl $32962, %ebx
movl %ebx, %eax
popl %ebx
ret
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82439 91177308-0d34-0410-b5e6-96231b3b80d8
2009-09-21 05:57:11 +00:00
|
|
|
ret i8 %A
|
|
|
|
|
|
2013-07-14 01:42:54 +00:00
|
|
|
; CHECK-LABEL: @coerce_mustalias_nonlocal1(
|
2012-12-12 20:43:00 +00:00
|
|
|
; CHECK: Cont:
|
|
|
|
|
; CHECK: %A = phi i8 [
|
|
|
|
|
; CHECK-NOT: load
|
|
|
|
|
; CHECK: ret i8 %A
|
Improve GVN to be able to forward substitute a small load
from a piece of a large store when both are in the same block.
This allows clang to compile the testcase in PR4216 to this code:
_test_bitfield:
movl 4(%esp), %eax
movl %eax, %ecx
andl $-65536, %ecx
orl $32962, %eax
andl $40186, %eax
orl %ecx, %eax
ret
This is not ideal, but is a whole lot better than the code produced
by llvm-gcc:
_test_bitfield:
movw $-32574, %ax
orw 4(%esp), %ax
andw $-25350, %ax
movw %ax, 4(%esp)
movw 7(%esp), %cx
shlw $8, %cx
movzbl 6(%esp), %edx
orw %cx, %dx
movzwl %dx, %ecx
shll $16, %ecx
movzwl %ax, %eax
orl %ecx, %eax
ret
and dramatically better than that produced by gcc 4.2:
_test_bitfield:
pushl %ebx
call L3
"L00000000001$pb":
L3:
popl %ebx
movl 8(%esp), %eax
leal 0(,%eax,4), %edx
sarb $7, %dl
movl %eax, %ecx
andl $7168, %ecx
andl $-7201, %ebx
movzbl %dl, %edx
andl $1, %edx
sall $5, %edx
orl %ecx, %ebx
orl %edx, %ebx
andl $24, %eax
andl $-58336, %ebx
orl %eax, %ebx
orl $32962, %ebx
movl %ebx, %eax
popl %ebx
ret
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82439 91177308-0d34-0410-b5e6-96231b3b80d8
2009-09-21 05:57:11 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
2009-09-20 20:09:34 +00:00
|
|
|
;; non-local i32 -> i8 partial redundancy load forwarding.
|
|
|
|
|
define i8 @coerce_mustalias_pre0(i32* %P, i1 %cond) {
|
|
|
|
|
%P3 = bitcast i32* %P to i8*
|
|
|
|
|
br i1 %cond, label %T, label %F
|
|
|
|
|
T:
|
|
|
|
|
store i32 42, i32* %P
|
|
|
|
|
br label %Cont
|
|
|
|
|
|
|
|
|
|
F:
|
|
|
|
|
br label %Cont
|
|
|
|
|
|
|
|
|
|
Cont:
|
2015-02-27 21:17:42 +00:00
|
|
|
%A = load i8, i8* %P3
|
2009-09-20 20:09:34 +00:00
|
|
|
ret i8 %A
|
|
|
|
|
|
2013-07-14 01:42:54 +00:00
|
|
|
; CHECK-LABEL: @coerce_mustalias_pre0(
|
2009-09-20 20:09:34 +00:00
|
|
|
; CHECK: F:
|
2015-02-27 21:17:42 +00:00
|
|
|
; CHECK: load i8, i8* %P3
|
2009-09-20 20:09:34 +00:00
|
|
|
; CHECK: Cont:
|
|
|
|
|
; CHECK: %A = phi i8 [
|
|
|
|
|
; CHECK-NOT: load
|
|
|
|
|
; CHECK: ret i8 %A
|
|
|
|
|
}
|
|
|
|
|
|
Improve GVN to be able to forward substitute a small load
from a piece of a large store when both are in the same block.
This allows clang to compile the testcase in PR4216 to this code:
_test_bitfield:
movl 4(%esp), %eax
movl %eax, %ecx
andl $-65536, %ecx
orl $32962, %eax
andl $40186, %eax
orl %ecx, %eax
ret
This is not ideal, but is a whole lot better than the code produced
by llvm-gcc:
_test_bitfield:
movw $-32574, %ax
orw 4(%esp), %ax
andw $-25350, %ax
movw %ax, 4(%esp)
movw 7(%esp), %cx
shlw $8, %cx
movzbl 6(%esp), %edx
orw %cx, %dx
movzwl %dx, %ecx
shll $16, %ecx
movzwl %ax, %eax
orl %ecx, %eax
ret
and dramatically better than that produced by gcc 4.2:
_test_bitfield:
pushl %ebx
call L3
"L00000000001$pb":
L3:
popl %ebx
movl 8(%esp), %eax
leal 0(,%eax,4), %edx
sarb $7, %dl
movl %eax, %ecx
andl $7168, %ecx
andl $-7201, %ebx
movzbl %dl, %edx
andl $1, %edx
sall $5, %edx
orl %ecx, %ebx
orl %edx, %ebx
andl $24, %eax
andl $-58336, %ebx
orl %eax, %ebx
orl $32962, %ebx
movl %ebx, %eax
popl %ebx
ret
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82439 91177308-0d34-0410-b5e6-96231b3b80d8
2009-09-21 05:57:11 +00:00
|
|
|
;;===----------------------------------------------------------------------===;;
|
|
|
|
|
;; Store -> Load and Load -> Load forwarding where src and dst are different
|
|
|
|
|
;; types, and the reload is an offset from the store pointer.
|
|
|
|
|
;;===----------------------------------------------------------------------===;;
|
|
|
|
|
|
2009-09-21 06:48:08 +00:00
|
|
|
;; i32 -> i8 forwarding.
|
2009-09-21 05:57:47 +00:00
|
|
|
;; PR4216
|
Improve GVN to be able to forward substitute a small load
from a piece of a large store when both are in the same block.
This allows clang to compile the testcase in PR4216 to this code:
_test_bitfield:
movl 4(%esp), %eax
movl %eax, %ecx
andl $-65536, %ecx
orl $32962, %eax
andl $40186, %eax
orl %ecx, %eax
ret
This is not ideal, but is a whole lot better than the code produced
by llvm-gcc:
_test_bitfield:
movw $-32574, %ax
orw 4(%esp), %ax
andw $-25350, %ax
movw %ax, 4(%esp)
movw 7(%esp), %cx
shlw $8, %cx
movzbl 6(%esp), %edx
orw %cx, %dx
movzwl %dx, %ecx
shll $16, %ecx
movzwl %ax, %eax
orl %ecx, %eax
ret
and dramatically better than that produced by gcc 4.2:
_test_bitfield:
pushl %ebx
call L3
"L00000000001$pb":
L3:
popl %ebx
movl 8(%esp), %eax
leal 0(,%eax,4), %edx
sarb $7, %dl
movl %eax, %ecx
andl $7168, %ecx
andl $-7201, %ebx
movzbl %dl, %edx
andl $1, %edx
sall $5, %edx
orl %ecx, %ebx
orl %edx, %ebx
andl $24, %eax
andl $-58336, %ebx
orl %eax, %ebx
orl $32962, %ebx
movl %ebx, %eax
popl %ebx
ret
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82439 91177308-0d34-0410-b5e6-96231b3b80d8
2009-09-21 05:57:11 +00:00
|
|
|
define i8 @coerce_offset0(i32 %V, i32* %P) {
|
|
|
|
|
store i32 %V, i32* %P
|
|
|
|
|
|
|
|
|
|
%P2 = bitcast i32* %P to i8*
|
[opaque pointer type] Add textual IR support for explicit type parameter to getelementptr instruction
One of several parallel first steps to remove the target type of pointers,
replacing them with a single opaque pointer type.
This adds an explicit type parameter to the gep instruction so that when the
first parameter becomes an opaque pointer type, the type to gep through is
still available to the instructions.
* This doesn't modify gep operators, only instructions (operators will be
handled separately)
* Textual IR changes only. Bitcode (including upgrade) and changing the
in-memory representation will be in separate changes.
* geps of vectors are transformed as:
getelementptr <4 x float*> %x, ...
->getelementptr float, <4 x float*> %x, ...
Then, once the opaque pointer type is introduced, this will ultimately look
like:
getelementptr float, <4 x ptr> %x
with the unambiguous interpretation that it is a vector of pointers to float.
* address spaces remain on the pointer, not the type:
getelementptr float addrspace(1)* %x
->getelementptr float, float addrspace(1)* %x
Then, eventually:
getelementptr float, ptr addrspace(1) %x
Importantly, the massive amount of test case churn has been automated by
same crappy python code. I had to manually update a few test cases that
wouldn't fit the script's model (r228970,r229196,r229197,r229198). The
python script just massages stdin and writes the result to stdout, I
then wrapped that in a shell script to handle replacing files, then
using the usual find+xargs to migrate all the files.
update.py:
import fileinput
import sys
import re
ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
def conv(match, line):
if not match:
return line
line = match.groups()[0]
if len(match.groups()[5]) == 0:
line += match.groups()[2]
line += match.groups()[3]
line += ", "
line += match.groups()[1]
line += "\n"
return line
for line in sys.stdin:
if line.find("getelementptr ") == line.find("getelementptr inbounds"):
if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("):
line = conv(re.match(ibrep, line), line)
elif line.find("getelementptr ") != line.find("getelementptr ("):
line = conv(re.match(normrep, line), line)
sys.stdout.write(line)
apply.sh:
for name in "$@"
do
python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name"
rm -f "$name.tmp"
done
The actual commands:
From llvm/src:
find test/ -name *.ll | xargs ./apply.sh
From llvm/src/tools/clang:
find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}"
From llvm/src/tools/polly:
find test/ -name *.ll | xargs ./apply.sh
After that, check-all (with llvm, clang, clang-tools-extra, lld,
compiler-rt, and polly all checked out).
The extra 'rm' in the apply.sh script is due to a few files in clang's test
suite using interesting unicode stuff that my python script was throwing
exceptions on. None of those files needed to be migrated, so it seemed
sufficient to ignore those cases.
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7636
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230786 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 19:29:02 +00:00
|
|
|
%P3 = getelementptr i8, i8* %P2, i32 2
|
Improve GVN to be able to forward substitute a small load
from a piece of a large store when both are in the same block.
This allows clang to compile the testcase in PR4216 to this code:
_test_bitfield:
movl 4(%esp), %eax
movl %eax, %ecx
andl $-65536, %ecx
orl $32962, %eax
andl $40186, %eax
orl %ecx, %eax
ret
This is not ideal, but is a whole lot better than the code produced
by llvm-gcc:
_test_bitfield:
movw $-32574, %ax
orw 4(%esp), %ax
andw $-25350, %ax
movw %ax, 4(%esp)
movw 7(%esp), %cx
shlw $8, %cx
movzbl 6(%esp), %edx
orw %cx, %dx
movzwl %dx, %ecx
shll $16, %ecx
movzwl %ax, %eax
orl %ecx, %eax
ret
and dramatically better than that produced by gcc 4.2:
_test_bitfield:
pushl %ebx
call L3
"L00000000001$pb":
L3:
popl %ebx
movl 8(%esp), %eax
leal 0(,%eax,4), %edx
sarb $7, %dl
movl %eax, %ecx
andl $7168, %ecx
andl $-7201, %ebx
movzbl %dl, %edx
andl $1, %edx
sall $5, %edx
orl %ecx, %ebx
orl %edx, %ebx
andl $24, %eax
andl $-58336, %ebx
orl %eax, %ebx
orl $32962, %ebx
movl %ebx, %eax
popl %ebx
ret
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82439 91177308-0d34-0410-b5e6-96231b3b80d8
2009-09-21 05:57:11 +00:00
|
|
|
|
2015-02-27 21:17:42 +00:00
|
|
|
%A = load i8, i8* %P3
|
Improve GVN to be able to forward substitute a small load
from a piece of a large store when both are in the same block.
This allows clang to compile the testcase in PR4216 to this code:
_test_bitfield:
movl 4(%esp), %eax
movl %eax, %ecx
andl $-65536, %ecx
orl $32962, %eax
andl $40186, %eax
orl %ecx, %eax
ret
This is not ideal, but is a whole lot better than the code produced
by llvm-gcc:
_test_bitfield:
movw $-32574, %ax
orw 4(%esp), %ax
andw $-25350, %ax
movw %ax, 4(%esp)
movw 7(%esp), %cx
shlw $8, %cx
movzbl 6(%esp), %edx
orw %cx, %dx
movzwl %dx, %ecx
shll $16, %ecx
movzwl %ax, %eax
orl %ecx, %eax
ret
and dramatically better than that produced by gcc 4.2:
_test_bitfield:
pushl %ebx
call L3
"L00000000001$pb":
L3:
popl %ebx
movl 8(%esp), %eax
leal 0(,%eax,4), %edx
sarb $7, %dl
movl %eax, %ecx
andl $7168, %ecx
andl $-7201, %ebx
movzbl %dl, %edx
andl $1, %edx
sall $5, %edx
orl %ecx, %ebx
orl %edx, %ebx
andl $24, %eax
andl $-58336, %ebx
orl %eax, %ebx
orl $32962, %ebx
movl %ebx, %eax
popl %ebx
ret
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82439 91177308-0d34-0410-b5e6-96231b3b80d8
2009-09-21 05:57:11 +00:00
|
|
|
ret i8 %A
|
2013-07-14 01:42:54 +00:00
|
|
|
; CHECK-LABEL: @coerce_offset0(
|
Improve GVN to be able to forward substitute a small load
from a piece of a large store when both are in the same block.
This allows clang to compile the testcase in PR4216 to this code:
_test_bitfield:
movl 4(%esp), %eax
movl %eax, %ecx
andl $-65536, %ecx
orl $32962, %eax
andl $40186, %eax
orl %ecx, %eax
ret
This is not ideal, but is a whole lot better than the code produced
by llvm-gcc:
_test_bitfield:
movw $-32574, %ax
orw 4(%esp), %ax
andw $-25350, %ax
movw %ax, 4(%esp)
movw 7(%esp), %cx
shlw $8, %cx
movzbl 6(%esp), %edx
orw %cx, %dx
movzwl %dx, %ecx
shll $16, %ecx
movzwl %ax, %eax
orl %ecx, %eax
ret
and dramatically better than that produced by gcc 4.2:
_test_bitfield:
pushl %ebx
call L3
"L00000000001$pb":
L3:
popl %ebx
movl 8(%esp), %eax
leal 0(,%eax,4), %edx
sarb $7, %dl
movl %eax, %ecx
andl $7168, %ecx
andl $-7201, %ebx
movzbl %dl, %edx
andl $1, %edx
sall $5, %edx
orl %ecx, %ebx
orl %edx, %ebx
andl $24, %eax
andl $-58336, %ebx
orl %eax, %ebx
orl $32962, %ebx
movl %ebx, %eax
popl %ebx
ret
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82439 91177308-0d34-0410-b5e6-96231b3b80d8
2009-09-21 05:57:11 +00:00
|
|
|
; CHECK-NOT: load
|
|
|
|
|
; CHECK: ret i8
|
|
|
|
|
}
|
|
|
|
|
|
2014-07-14 22:39:22 +00:00
|
|
|
define i8 @coerce_offset0_addrspacecast(i32 %V, i32* %P) {
|
|
|
|
|
store i32 %V, i32* %P
|
|
|
|
|
|
|
|
|
|
%P2 = addrspacecast i32* %P to i8 addrspace(1)*
|
[opaque pointer type] Add textual IR support for explicit type parameter to getelementptr instruction
One of several parallel first steps to remove the target type of pointers,
replacing them with a single opaque pointer type.
This adds an explicit type parameter to the gep instruction so that when the
first parameter becomes an opaque pointer type, the type to gep through is
still available to the instructions.
* This doesn't modify gep operators, only instructions (operators will be
handled separately)
* Textual IR changes only. Bitcode (including upgrade) and changing the
in-memory representation will be in separate changes.
* geps of vectors are transformed as:
getelementptr <4 x float*> %x, ...
->getelementptr float, <4 x float*> %x, ...
Then, once the opaque pointer type is introduced, this will ultimately look
like:
getelementptr float, <4 x ptr> %x
with the unambiguous interpretation that it is a vector of pointers to float.
* address spaces remain on the pointer, not the type:
getelementptr float addrspace(1)* %x
->getelementptr float, float addrspace(1)* %x
Then, eventually:
getelementptr float, ptr addrspace(1) %x
Importantly, the massive amount of test case churn has been automated by
same crappy python code. I had to manually update a few test cases that
wouldn't fit the script's model (r228970,r229196,r229197,r229198). The
python script just massages stdin and writes the result to stdout, I
then wrapped that in a shell script to handle replacing files, then
using the usual find+xargs to migrate all the files.
update.py:
import fileinput
import sys
import re
ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
def conv(match, line):
if not match:
return line
line = match.groups()[0]
if len(match.groups()[5]) == 0:
line += match.groups()[2]
line += match.groups()[3]
line += ", "
line += match.groups()[1]
line += "\n"
return line
for line in sys.stdin:
if line.find("getelementptr ") == line.find("getelementptr inbounds"):
if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("):
line = conv(re.match(ibrep, line), line)
elif line.find("getelementptr ") != line.find("getelementptr ("):
line = conv(re.match(normrep, line), line)
sys.stdout.write(line)
apply.sh:
for name in "$@"
do
python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name"
rm -f "$name.tmp"
done
The actual commands:
From llvm/src:
find test/ -name *.ll | xargs ./apply.sh
From llvm/src/tools/clang:
find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}"
From llvm/src/tools/polly:
find test/ -name *.ll | xargs ./apply.sh
After that, check-all (with llvm, clang, clang-tools-extra, lld,
compiler-rt, and polly all checked out).
The extra 'rm' in the apply.sh script is due to a few files in clang's test
suite using interesting unicode stuff that my python script was throwing
exceptions on. None of those files needed to be migrated, so it seemed
sufficient to ignore those cases.
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7636
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230786 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 19:29:02 +00:00
|
|
|
%P3 = getelementptr i8, i8 addrspace(1)* %P2, i32 2
|
2014-07-14 22:39:22 +00:00
|
|
|
|
2015-02-27 21:17:42 +00:00
|
|
|
%A = load i8, i8 addrspace(1)* %P3
|
2014-07-14 22:39:22 +00:00
|
|
|
ret i8 %A
|
|
|
|
|
; CHECK-LABEL: @coerce_offset0_addrspacecast(
|
|
|
|
|
; CHECK-NOT: load
|
|
|
|
|
; CHECK: ret i8
|
|
|
|
|
}
|
|
|
|
|
|
2009-09-21 06:48:08 +00:00
|
|
|
;; non-local i32/float -> i8 load forwarding.
|
|
|
|
|
define i8 @coerce_offset_nonlocal0(i32* %P, i1 %cond) {
|
|
|
|
|
%P2 = bitcast i32* %P to float*
|
|
|
|
|
%P3 = bitcast i32* %P to i8*
|
[opaque pointer type] Add textual IR support for explicit type parameter to getelementptr instruction
One of several parallel first steps to remove the target type of pointers,
replacing them with a single opaque pointer type.
This adds an explicit type parameter to the gep instruction so that when the
first parameter becomes an opaque pointer type, the type to gep through is
still available to the instructions.
* This doesn't modify gep operators, only instructions (operators will be
handled separately)
* Textual IR changes only. Bitcode (including upgrade) and changing the
in-memory representation will be in separate changes.
* geps of vectors are transformed as:
getelementptr <4 x float*> %x, ...
->getelementptr float, <4 x float*> %x, ...
Then, once the opaque pointer type is introduced, this will ultimately look
like:
getelementptr float, <4 x ptr> %x
with the unambiguous interpretation that it is a vector of pointers to float.
* address spaces remain on the pointer, not the type:
getelementptr float addrspace(1)* %x
->getelementptr float, float addrspace(1)* %x
Then, eventually:
getelementptr float, ptr addrspace(1) %x
Importantly, the massive amount of test case churn has been automated by
same crappy python code. I had to manually update a few test cases that
wouldn't fit the script's model (r228970,r229196,r229197,r229198). The
python script just massages stdin and writes the result to stdout, I
then wrapped that in a shell script to handle replacing files, then
using the usual find+xargs to migrate all the files.
update.py:
import fileinput
import sys
import re
ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
def conv(match, line):
if not match:
return line
line = match.groups()[0]
if len(match.groups()[5]) == 0:
line += match.groups()[2]
line += match.groups()[3]
line += ", "
line += match.groups()[1]
line += "\n"
return line
for line in sys.stdin:
if line.find("getelementptr ") == line.find("getelementptr inbounds"):
if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("):
line = conv(re.match(ibrep, line), line)
elif line.find("getelementptr ") != line.find("getelementptr ("):
line = conv(re.match(normrep, line), line)
sys.stdout.write(line)
apply.sh:
for name in "$@"
do
python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name"
rm -f "$name.tmp"
done
The actual commands:
From llvm/src:
find test/ -name *.ll | xargs ./apply.sh
From llvm/src/tools/clang:
find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}"
From llvm/src/tools/polly:
find test/ -name *.ll | xargs ./apply.sh
After that, check-all (with llvm, clang, clang-tools-extra, lld,
compiler-rt, and polly all checked out).
The extra 'rm' in the apply.sh script is due to a few files in clang's test
suite using interesting unicode stuff that my python script was throwing
exceptions on. None of those files needed to be migrated, so it seemed
sufficient to ignore those cases.
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7636
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230786 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 19:29:02 +00:00
|
|
|
%P4 = getelementptr i8, i8* %P3, i32 2
|
2009-09-21 06:48:08 +00:00
|
|
|
br i1 %cond, label %T, label %F
|
|
|
|
|
T:
|
2012-11-08 20:34:25 +00:00
|
|
|
store i32 57005, i32* %P
|
2009-09-21 06:48:08 +00:00
|
|
|
br label %Cont
|
|
|
|
|
|
|
|
|
|
F:
|
|
|
|
|
store float 1.0, float* %P2
|
|
|
|
|
br label %Cont
|
|
|
|
|
|
|
|
|
|
Cont:
|
2015-02-27 21:17:42 +00:00
|
|
|
%A = load i8, i8* %P4
|
2009-09-21 06:48:08 +00:00
|
|
|
ret i8 %A
|
|
|
|
|
|
2013-07-14 01:42:54 +00:00
|
|
|
; CHECK-LABEL: @coerce_offset_nonlocal0(
|
2009-09-21 06:48:08 +00:00
|
|
|
; CHECK: Cont:
|
|
|
|
|
; CHECK: %A = phi i8 [
|
|
|
|
|
; CHECK-NOT: load
|
|
|
|
|
; CHECK: ret i8 %A
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
;; non-local i32 -> i8 partial redundancy load forwarding.
|
|
|
|
|
define i8 @coerce_offset_pre0(i32* %P, i1 %cond) {
|
|
|
|
|
%P3 = bitcast i32* %P to i8*
|
[opaque pointer type] Add textual IR support for explicit type parameter to getelementptr instruction
One of several parallel first steps to remove the target type of pointers,
replacing them with a single opaque pointer type.
This adds an explicit type parameter to the gep instruction so that when the
first parameter becomes an opaque pointer type, the type to gep through is
still available to the instructions.
* This doesn't modify gep operators, only instructions (operators will be
handled separately)
* Textual IR changes only. Bitcode (including upgrade) and changing the
in-memory representation will be in separate changes.
* geps of vectors are transformed as:
getelementptr <4 x float*> %x, ...
->getelementptr float, <4 x float*> %x, ...
Then, once the opaque pointer type is introduced, this will ultimately look
like:
getelementptr float, <4 x ptr> %x
with the unambiguous interpretation that it is a vector of pointers to float.
* address spaces remain on the pointer, not the type:
getelementptr float addrspace(1)* %x
->getelementptr float, float addrspace(1)* %x
Then, eventually:
getelementptr float, ptr addrspace(1) %x
Importantly, the massive amount of test case churn has been automated by
same crappy python code. I had to manually update a few test cases that
wouldn't fit the script's model (r228970,r229196,r229197,r229198). The
python script just massages stdin and writes the result to stdout, I
then wrapped that in a shell script to handle replacing files, then
using the usual find+xargs to migrate all the files.
update.py:
import fileinput
import sys
import re
ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
def conv(match, line):
if not match:
return line
line = match.groups()[0]
if len(match.groups()[5]) == 0:
line += match.groups()[2]
line += match.groups()[3]
line += ", "
line += match.groups()[1]
line += "\n"
return line
for line in sys.stdin:
if line.find("getelementptr ") == line.find("getelementptr inbounds"):
if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("):
line = conv(re.match(ibrep, line), line)
elif line.find("getelementptr ") != line.find("getelementptr ("):
line = conv(re.match(normrep, line), line)
sys.stdout.write(line)
apply.sh:
for name in "$@"
do
python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name"
rm -f "$name.tmp"
done
The actual commands:
From llvm/src:
find test/ -name *.ll | xargs ./apply.sh
From llvm/src/tools/clang:
find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}"
From llvm/src/tools/polly:
find test/ -name *.ll | xargs ./apply.sh
After that, check-all (with llvm, clang, clang-tools-extra, lld,
compiler-rt, and polly all checked out).
The extra 'rm' in the apply.sh script is due to a few files in clang's test
suite using interesting unicode stuff that my python script was throwing
exceptions on. None of those files needed to be migrated, so it seemed
sufficient to ignore those cases.
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7636
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230786 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 19:29:02 +00:00
|
|
|
%P4 = getelementptr i8, i8* %P3, i32 2
|
2009-09-21 06:48:08 +00:00
|
|
|
br i1 %cond, label %T, label %F
|
|
|
|
|
T:
|
|
|
|
|
store i32 42, i32* %P
|
|
|
|
|
br label %Cont
|
|
|
|
|
|
|
|
|
|
F:
|
|
|
|
|
br label %Cont
|
|
|
|
|
|
|
|
|
|
Cont:
|
2015-02-27 21:17:42 +00:00
|
|
|
%A = load i8, i8* %P4
|
2009-09-21 06:48:08 +00:00
|
|
|
ret i8 %A
|
|
|
|
|
|
2013-07-14 01:42:54 +00:00
|
|
|
; CHECK-LABEL: @coerce_offset_pre0(
|
2009-09-21 06:48:08 +00:00
|
|
|
; CHECK: F:
|
2015-02-27 21:17:42 +00:00
|
|
|
; CHECK: load i8, i8* %P4
|
2009-09-21 06:48:08 +00:00
|
|
|
; CHECK: Cont:
|
|
|
|
|
; CHECK: %A = phi i8 [
|
|
|
|
|
; CHECK-NOT: load
|
|
|
|
|
; CHECK: ret i8 %A
|
|
|
|
|
}
|
Improve GVN to be able to forward substitute a small load
from a piece of a large store when both are in the same block.
This allows clang to compile the testcase in PR4216 to this code:
_test_bitfield:
movl 4(%esp), %eax
movl %eax, %ecx
andl $-65536, %ecx
orl $32962, %eax
andl $40186, %eax
orl %ecx, %eax
ret
This is not ideal, but is a whole lot better than the code produced
by llvm-gcc:
_test_bitfield:
movw $-32574, %ax
orw 4(%esp), %ax
andw $-25350, %ax
movw %ax, 4(%esp)
movw 7(%esp), %cx
shlw $8, %cx
movzbl 6(%esp), %edx
orw %cx, %dx
movzwl %dx, %ecx
shll $16, %ecx
movzwl %ax, %eax
orl %ecx, %eax
ret
and dramatically better than that produced by gcc 4.2:
_test_bitfield:
pushl %ebx
call L3
"L00000000001$pb":
L3:
popl %ebx
movl 8(%esp), %eax
leal 0(,%eax,4), %edx
sarb $7, %dl
movl %eax, %ecx
andl $7168, %ecx
andl $-7201, %ebx
movzbl %dl, %edx
andl $1, %edx
sall $5, %edx
orl %ecx, %ebx
orl %edx, %ebx
andl $24, %eax
andl $-58336, %ebx
orl %eax, %ebx
orl $32962, %ebx
movl %ebx, %eax
popl %ebx
ret
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82439 91177308-0d34-0410-b5e6-96231b3b80d8
2009-09-21 05:57:11 +00:00
|
|
|
|
2013-11-11 22:00:23 +00:00
|
|
|
define i32 @chained_load(i32** %p, i32 %x, i32 %y) {
|
2009-12-06 01:47:24 +00:00
|
|
|
block1:
|
2011-05-22 07:02:43 +00:00
|
|
|
%A = alloca i32*
|
|
|
|
|
|
2015-02-27 21:17:42 +00:00
|
|
|
%z = load i32*, i32** %p
|
2011-05-22 07:02:43 +00:00
|
|
|
store i32* %z, i32** %A
|
2013-11-11 22:00:23 +00:00
|
|
|
%cmp = icmp eq i32 %x, %y
|
|
|
|
|
br i1 %cmp, label %block2, label %block3
|
2009-12-06 01:47:24 +00:00
|
|
|
|
|
|
|
|
block2:
|
2015-02-27 21:17:42 +00:00
|
|
|
%a = load i32*, i32** %p
|
2009-12-06 01:47:24 +00:00
|
|
|
br label %block4
|
|
|
|
|
|
|
|
|
|
block3:
|
2015-02-27 21:17:42 +00:00
|
|
|
%b = load i32*, i32** %p
|
2009-12-06 01:47:24 +00:00
|
|
|
br label %block4
|
|
|
|
|
|
|
|
|
|
block4:
|
2015-02-27 21:17:42 +00:00
|
|
|
%c = load i32*, i32** %p
|
|
|
|
|
%d = load i32, i32* %c
|
2009-12-06 01:47:24 +00:00
|
|
|
ret i32 %d
|
|
|
|
|
|
2013-07-14 01:42:54 +00:00
|
|
|
; CHECK-LABEL: @chained_load(
|
2015-02-27 21:17:42 +00:00
|
|
|
; CHECK: %z = load i32*, i32** %p
|
2009-12-06 01:47:24 +00:00
|
|
|
; CHECK-NOT: load
|
2015-02-27 21:17:42 +00:00
|
|
|
; CHECK: %d = load i32, i32* %z
|
2009-12-06 01:47:24 +00:00
|
|
|
; CHECK-NEXT: ret i32 %d
|
|
|
|
|
}
|
|
|
|
|
|
Improve GVN to be able to forward substitute a small load
from a piece of a large store when both are in the same block.
This allows clang to compile the testcase in PR4216 to this code:
_test_bitfield:
movl 4(%esp), %eax
movl %eax, %ecx
andl $-65536, %ecx
orl $32962, %eax
andl $40186, %eax
orl %ecx, %eax
ret
This is not ideal, but is a whole lot better than the code produced
by llvm-gcc:
_test_bitfield:
movw $-32574, %ax
orw 4(%esp), %ax
andw $-25350, %ax
movw %ax, 4(%esp)
movw 7(%esp), %cx
shlw $8, %cx
movzbl 6(%esp), %edx
orw %cx, %dx
movzwl %dx, %ecx
shll $16, %ecx
movzwl %ax, %eax
orl %ecx, %eax
ret
and dramatically better than that produced by gcc 4.2:
_test_bitfield:
pushl %ebx
call L3
"L00000000001$pb":
L3:
popl %ebx
movl 8(%esp), %eax
leal 0(,%eax,4), %edx
sarb $7, %dl
movl %eax, %ecx
andl $7168, %ecx
andl $-7201, %ebx
movzbl %dl, %edx
andl $1, %edx
sall $5, %edx
orl %ecx, %ebx
orl %edx, %ebx
andl $24, %eax
andl $-58336, %ebx
orl %eax, %ebx
orl $32962, %ebx
movl %ebx, %eax
popl %ebx
ret
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@82439 91177308-0d34-0410-b5e6-96231b3b80d8
2009-09-21 05:57:11 +00:00
|
|
|
|
2009-12-04 02:12:12 +00:00
|
|
|
declare i1 @cond() readonly
|
|
|
|
|
declare i1 @cond2() readonly
|
|
|
|
|
|
|
|
|
|
define i32 @phi_trans2() {
|
2013-07-14 01:42:54 +00:00
|
|
|
; CHECK-LABEL: @phi_trans2(
|
2009-12-04 02:12:12 +00:00
|
|
|
entry:
|
|
|
|
|
%P = alloca i32, i32 400
|
|
|
|
|
br label %F1
|
|
|
|
|
|
|
|
|
|
F1:
|
|
|
|
|
%A = phi i32 [1, %entry], [2, %F]
|
|
|
|
|
%cond2 = call i1 @cond()
|
|
|
|
|
br i1 %cond2, label %T1, label %TY
|
|
|
|
|
|
|
|
|
|
T1:
|
[opaque pointer type] Add textual IR support for explicit type parameter to getelementptr instruction
One of several parallel first steps to remove the target type of pointers,
replacing them with a single opaque pointer type.
This adds an explicit type parameter to the gep instruction so that when the
first parameter becomes an opaque pointer type, the type to gep through is
still available to the instructions.
* This doesn't modify gep operators, only instructions (operators will be
handled separately)
* Textual IR changes only. Bitcode (including upgrade) and changing the
in-memory representation will be in separate changes.
* geps of vectors are transformed as:
getelementptr <4 x float*> %x, ...
->getelementptr float, <4 x float*> %x, ...
Then, once the opaque pointer type is introduced, this will ultimately look
like:
getelementptr float, <4 x ptr> %x
with the unambiguous interpretation that it is a vector of pointers to float.
* address spaces remain on the pointer, not the type:
getelementptr float addrspace(1)* %x
->getelementptr float, float addrspace(1)* %x
Then, eventually:
getelementptr float, ptr addrspace(1) %x
Importantly, the massive amount of test case churn has been automated by
same crappy python code. I had to manually update a few test cases that
wouldn't fit the script's model (r228970,r229196,r229197,r229198). The
python script just massages stdin and writes the result to stdout, I
then wrapped that in a shell script to handle replacing files, then
using the usual find+xargs to migrate all the files.
update.py:
import fileinput
import sys
import re
ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
def conv(match, line):
if not match:
return line
line = match.groups()[0]
if len(match.groups()[5]) == 0:
line += match.groups()[2]
line += match.groups()[3]
line += ", "
line += match.groups()[1]
line += "\n"
return line
for line in sys.stdin:
if line.find("getelementptr ") == line.find("getelementptr inbounds"):
if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("):
line = conv(re.match(ibrep, line), line)
elif line.find("getelementptr ") != line.find("getelementptr ("):
line = conv(re.match(normrep, line), line)
sys.stdout.write(line)
apply.sh:
for name in "$@"
do
python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name"
rm -f "$name.tmp"
done
The actual commands:
From llvm/src:
find test/ -name *.ll | xargs ./apply.sh
From llvm/src/tools/clang:
find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}"
From llvm/src/tools/polly:
find test/ -name *.ll | xargs ./apply.sh
After that, check-all (with llvm, clang, clang-tools-extra, lld,
compiler-rt, and polly all checked out).
The extra 'rm' in the apply.sh script is due to a few files in clang's test
suite using interesting unicode stuff that my python script was throwing
exceptions on. None of those files needed to be migrated, so it seemed
sufficient to ignore those cases.
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7636
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230786 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 19:29:02 +00:00
|
|
|
%P2 = getelementptr i32, i32* %P, i32 %A
|
2015-02-27 21:17:42 +00:00
|
|
|
%x = load i32, i32* %P2
|
2009-12-04 02:12:12 +00:00
|
|
|
%cond = call i1 @cond2()
|
|
|
|
|
br i1 %cond, label %TX, label %F
|
|
|
|
|
|
|
|
|
|
F:
|
[opaque pointer type] Add textual IR support for explicit type parameter to getelementptr instruction
One of several parallel first steps to remove the target type of pointers,
replacing them with a single opaque pointer type.
This adds an explicit type parameter to the gep instruction so that when the
first parameter becomes an opaque pointer type, the type to gep through is
still available to the instructions.
* This doesn't modify gep operators, only instructions (operators will be
handled separately)
* Textual IR changes only. Bitcode (including upgrade) and changing the
in-memory representation will be in separate changes.
* geps of vectors are transformed as:
getelementptr <4 x float*> %x, ...
->getelementptr float, <4 x float*> %x, ...
Then, once the opaque pointer type is introduced, this will ultimately look
like:
getelementptr float, <4 x ptr> %x
with the unambiguous interpretation that it is a vector of pointers to float.
* address spaces remain on the pointer, not the type:
getelementptr float addrspace(1)* %x
->getelementptr float, float addrspace(1)* %x
Then, eventually:
getelementptr float, ptr addrspace(1) %x
Importantly, the massive amount of test case churn has been automated by
same crappy python code. I had to manually update a few test cases that
wouldn't fit the script's model (r228970,r229196,r229197,r229198). The
python script just massages stdin and writes the result to stdout, I
then wrapped that in a shell script to handle replacing files, then
using the usual find+xargs to migrate all the files.
update.py:
import fileinput
import sys
import re
ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
def conv(match, line):
if not match:
return line
line = match.groups()[0]
if len(match.groups()[5]) == 0:
line += match.groups()[2]
line += match.groups()[3]
line += ", "
line += match.groups()[1]
line += "\n"
return line
for line in sys.stdin:
if line.find("getelementptr ") == line.find("getelementptr inbounds"):
if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("):
line = conv(re.match(ibrep, line), line)
elif line.find("getelementptr ") != line.find("getelementptr ("):
line = conv(re.match(normrep, line), line)
sys.stdout.write(line)
apply.sh:
for name in "$@"
do
python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name"
rm -f "$name.tmp"
done
The actual commands:
From llvm/src:
find test/ -name *.ll | xargs ./apply.sh
From llvm/src/tools/clang:
find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}"
From llvm/src/tools/polly:
find test/ -name *.ll | xargs ./apply.sh
After that, check-all (with llvm, clang, clang-tools-extra, lld,
compiler-rt, and polly all checked out).
The extra 'rm' in the apply.sh script is due to a few files in clang's test
suite using interesting unicode stuff that my python script was throwing
exceptions on. None of those files needed to be migrated, so it seemed
sufficient to ignore those cases.
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7636
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230786 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 19:29:02 +00:00
|
|
|
%P3 = getelementptr i32, i32* %P, i32 2
|
2009-12-04 02:12:12 +00:00
|
|
|
store i32 17, i32* %P3
|
|
|
|
|
|
|
|
|
|
store i32 42, i32* %P2 ; Provides "P[A]".
|
|
|
|
|
br label %F1
|
|
|
|
|
|
|
|
|
|
TX:
|
2009-12-09 01:59:31 +00:00
|
|
|
; This load should not be compiled to 'ret i32 42'. An overly clever
|
|
|
|
|
; implementation of GVN would see that we're returning 17 if the loop
|
|
|
|
|
; executes once or 42 if it executes more than that, but we'd have to do
|
|
|
|
|
; loop restructuring to expose this, and GVN shouldn't do this sort of CFG
|
|
|
|
|
; transformation.
|
|
|
|
|
|
|
|
|
|
; CHECK: TX:
|
|
|
|
|
; CHECK: ret i32 %x
|
|
|
|
|
ret i32 %x
|
2009-12-04 02:12:12 +00:00
|
|
|
TY:
|
|
|
|
|
ret i32 0
|
|
|
|
|
}
|
|
|
|
|
|
2013-11-11 22:00:23 +00:00
|
|
|
define i32 @phi_trans3(i32* %p, i32 %x, i32 %y, i32 %z) {
|
2013-07-14 01:42:54 +00:00
|
|
|
; CHECK-LABEL: @phi_trans3(
|
2009-12-09 01:59:31 +00:00
|
|
|
block1:
|
2013-11-11 22:00:23 +00:00
|
|
|
%cmpxy = icmp eq i32 %x, %y
|
|
|
|
|
br i1 %cmpxy, label %block2, label %block3
|
2009-12-09 01:59:31 +00:00
|
|
|
|
|
|
|
|
block2:
|
|
|
|
|
store i32 87, i32* %p
|
|
|
|
|
br label %block4
|
|
|
|
|
|
|
|
|
|
block3:
|
[opaque pointer type] Add textual IR support for explicit type parameter to getelementptr instruction
One of several parallel first steps to remove the target type of pointers,
replacing them with a single opaque pointer type.
This adds an explicit type parameter to the gep instruction so that when the
first parameter becomes an opaque pointer type, the type to gep through is
still available to the instructions.
* This doesn't modify gep operators, only instructions (operators will be
handled separately)
* Textual IR changes only. Bitcode (including upgrade) and changing the
in-memory representation will be in separate changes.
* geps of vectors are transformed as:
getelementptr <4 x float*> %x, ...
->getelementptr float, <4 x float*> %x, ...
Then, once the opaque pointer type is introduced, this will ultimately look
like:
getelementptr float, <4 x ptr> %x
with the unambiguous interpretation that it is a vector of pointers to float.
* address spaces remain on the pointer, not the type:
getelementptr float addrspace(1)* %x
->getelementptr float, float addrspace(1)* %x
Then, eventually:
getelementptr float, ptr addrspace(1) %x
Importantly, the massive amount of test case churn has been automated by
same crappy python code. I had to manually update a few test cases that
wouldn't fit the script's model (r228970,r229196,r229197,r229198). The
python script just massages stdin and writes the result to stdout, I
then wrapped that in a shell script to handle replacing files, then
using the usual find+xargs to migrate all the files.
update.py:
import fileinput
import sys
import re
ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
def conv(match, line):
if not match:
return line
line = match.groups()[0]
if len(match.groups()[5]) == 0:
line += match.groups()[2]
line += match.groups()[3]
line += ", "
line += match.groups()[1]
line += "\n"
return line
for line in sys.stdin:
if line.find("getelementptr ") == line.find("getelementptr inbounds"):
if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("):
line = conv(re.match(ibrep, line), line)
elif line.find("getelementptr ") != line.find("getelementptr ("):
line = conv(re.match(normrep, line), line)
sys.stdout.write(line)
apply.sh:
for name in "$@"
do
python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name"
rm -f "$name.tmp"
done
The actual commands:
From llvm/src:
find test/ -name *.ll | xargs ./apply.sh
From llvm/src/tools/clang:
find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}"
From llvm/src/tools/polly:
find test/ -name *.ll | xargs ./apply.sh
After that, check-all (with llvm, clang, clang-tools-extra, lld,
compiler-rt, and polly all checked out).
The extra 'rm' in the apply.sh script is due to a few files in clang's test
suite using interesting unicode stuff that my python script was throwing
exceptions on. None of those files needed to be migrated, so it seemed
sufficient to ignore those cases.
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7636
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230786 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 19:29:02 +00:00
|
|
|
%p2 = getelementptr i32, i32* %p, i32 43
|
2009-12-09 01:59:31 +00:00
|
|
|
store i32 97, i32* %p2
|
|
|
|
|
br label %block4
|
|
|
|
|
|
|
|
|
|
block4:
|
|
|
|
|
%A = phi i32 [-1, %block2], [42, %block3]
|
2013-11-11 22:00:23 +00:00
|
|
|
br i1 %cmpxy, label %block5, label %exit
|
2009-12-09 01:59:31 +00:00
|
|
|
|
|
|
|
|
; CHECK: block4:
|
|
|
|
|
; CHECK-NEXT: %D = phi i32 [ 87, %block2 ], [ 97, %block3 ]
|
|
|
|
|
; CHECK-NOT: load
|
|
|
|
|
|
|
|
|
|
block5:
|
|
|
|
|
%B = add i32 %A, 1
|
2013-11-11 22:00:23 +00:00
|
|
|
br i1 %cmpxy, label %block6, label %exit
|
2009-12-09 01:59:31 +00:00
|
|
|
|
|
|
|
|
block6:
|
[opaque pointer type] Add textual IR support for explicit type parameter to getelementptr instruction
One of several parallel first steps to remove the target type of pointers,
replacing them with a single opaque pointer type.
This adds an explicit type parameter to the gep instruction so that when the
first parameter becomes an opaque pointer type, the type to gep through is
still available to the instructions.
* This doesn't modify gep operators, only instructions (operators will be
handled separately)
* Textual IR changes only. Bitcode (including upgrade) and changing the
in-memory representation will be in separate changes.
* geps of vectors are transformed as:
getelementptr <4 x float*> %x, ...
->getelementptr float, <4 x float*> %x, ...
Then, once the opaque pointer type is introduced, this will ultimately look
like:
getelementptr float, <4 x ptr> %x
with the unambiguous interpretation that it is a vector of pointers to float.
* address spaces remain on the pointer, not the type:
getelementptr float addrspace(1)* %x
->getelementptr float, float addrspace(1)* %x
Then, eventually:
getelementptr float, ptr addrspace(1) %x
Importantly, the massive amount of test case churn has been automated by
same crappy python code. I had to manually update a few test cases that
wouldn't fit the script's model (r228970,r229196,r229197,r229198). The
python script just massages stdin and writes the result to stdout, I
then wrapped that in a shell script to handle replacing files, then
using the usual find+xargs to migrate all the files.
update.py:
import fileinput
import sys
import re
ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
def conv(match, line):
if not match:
return line
line = match.groups()[0]
if len(match.groups()[5]) == 0:
line += match.groups()[2]
line += match.groups()[3]
line += ", "
line += match.groups()[1]
line += "\n"
return line
for line in sys.stdin:
if line.find("getelementptr ") == line.find("getelementptr inbounds"):
if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("):
line = conv(re.match(ibrep, line), line)
elif line.find("getelementptr ") != line.find("getelementptr ("):
line = conv(re.match(normrep, line), line)
sys.stdout.write(line)
apply.sh:
for name in "$@"
do
python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name"
rm -f "$name.tmp"
done
The actual commands:
From llvm/src:
find test/ -name *.ll | xargs ./apply.sh
From llvm/src/tools/clang:
find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}"
From llvm/src/tools/polly:
find test/ -name *.ll | xargs ./apply.sh
After that, check-all (with llvm, clang, clang-tools-extra, lld,
compiler-rt, and polly all checked out).
The extra 'rm' in the apply.sh script is due to a few files in clang's test
suite using interesting unicode stuff that my python script was throwing
exceptions on. None of those files needed to be migrated, so it seemed
sufficient to ignore those cases.
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7636
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230786 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 19:29:02 +00:00
|
|
|
%C = getelementptr i32, i32* %p, i32 %B
|
2013-11-11 22:00:23 +00:00
|
|
|
br i1 %cmpxy, label %block7, label %exit
|
2009-12-09 01:59:31 +00:00
|
|
|
|
|
|
|
|
block7:
|
2015-02-27 21:17:42 +00:00
|
|
|
%D = load i32, i32* %C
|
2009-12-09 01:59:31 +00:00
|
|
|
ret i32 %D
|
|
|
|
|
|
|
|
|
|
; CHECK: block7:
|
|
|
|
|
; CHECK-NEXT: ret i32 %D
|
|
|
|
|
|
|
|
|
|
exit:
|
|
|
|
|
ret i32 -1
|
|
|
|
|
}
|
2009-12-04 02:12:12 +00:00
|
|
|
|
2009-12-09 02:43:05 +00:00
|
|
|
define i8 @phi_trans4(i8* %p) {
|
2013-07-14 01:42:54 +00:00
|
|
|
; CHECK-LABEL: @phi_trans4(
|
2009-12-09 02:43:05 +00:00
|
|
|
entry:
|
[opaque pointer type] Add textual IR support for explicit type parameter to getelementptr instruction
One of several parallel first steps to remove the target type of pointers,
replacing them with a single opaque pointer type.
This adds an explicit type parameter to the gep instruction so that when the
first parameter becomes an opaque pointer type, the type to gep through is
still available to the instructions.
* This doesn't modify gep operators, only instructions (operators will be
handled separately)
* Textual IR changes only. Bitcode (including upgrade) and changing the
in-memory representation will be in separate changes.
* geps of vectors are transformed as:
getelementptr <4 x float*> %x, ...
->getelementptr float, <4 x float*> %x, ...
Then, once the opaque pointer type is introduced, this will ultimately look
like:
getelementptr float, <4 x ptr> %x
with the unambiguous interpretation that it is a vector of pointers to float.
* address spaces remain on the pointer, not the type:
getelementptr float addrspace(1)* %x
->getelementptr float, float addrspace(1)* %x
Then, eventually:
getelementptr float, ptr addrspace(1) %x
Importantly, the massive amount of test case churn has been automated by
same crappy python code. I had to manually update a few test cases that
wouldn't fit the script's model (r228970,r229196,r229197,r229198). The
python script just massages stdin and writes the result to stdout, I
then wrapped that in a shell script to handle replacing files, then
using the usual find+xargs to migrate all the files.
update.py:
import fileinput
import sys
import re
ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
def conv(match, line):
if not match:
return line
line = match.groups()[0]
if len(match.groups()[5]) == 0:
line += match.groups()[2]
line += match.groups()[3]
line += ", "
line += match.groups()[1]
line += "\n"
return line
for line in sys.stdin:
if line.find("getelementptr ") == line.find("getelementptr inbounds"):
if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("):
line = conv(re.match(ibrep, line), line)
elif line.find("getelementptr ") != line.find("getelementptr ("):
line = conv(re.match(normrep, line), line)
sys.stdout.write(line)
apply.sh:
for name in "$@"
do
python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name"
rm -f "$name.tmp"
done
The actual commands:
From llvm/src:
find test/ -name *.ll | xargs ./apply.sh
From llvm/src/tools/clang:
find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}"
From llvm/src/tools/polly:
find test/ -name *.ll | xargs ./apply.sh
After that, check-all (with llvm, clang, clang-tools-extra, lld,
compiler-rt, and polly all checked out).
The extra 'rm' in the apply.sh script is due to a few files in clang's test
suite using interesting unicode stuff that my python script was throwing
exceptions on. None of those files needed to be migrated, so it seemed
sufficient to ignore those cases.
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7636
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230786 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 19:29:02 +00:00
|
|
|
%X3 = getelementptr i8, i8* %p, i32 192
|
2009-12-09 18:21:46 +00:00
|
|
|
store i8 192, i8* %X3
|
|
|
|
|
|
[opaque pointer type] Add textual IR support for explicit type parameter to getelementptr instruction
One of several parallel first steps to remove the target type of pointers,
replacing them with a single opaque pointer type.
This adds an explicit type parameter to the gep instruction so that when the
first parameter becomes an opaque pointer type, the type to gep through is
still available to the instructions.
* This doesn't modify gep operators, only instructions (operators will be
handled separately)
* Textual IR changes only. Bitcode (including upgrade) and changing the
in-memory representation will be in separate changes.
* geps of vectors are transformed as:
getelementptr <4 x float*> %x, ...
->getelementptr float, <4 x float*> %x, ...
Then, once the opaque pointer type is introduced, this will ultimately look
like:
getelementptr float, <4 x ptr> %x
with the unambiguous interpretation that it is a vector of pointers to float.
* address spaces remain on the pointer, not the type:
getelementptr float addrspace(1)* %x
->getelementptr float, float addrspace(1)* %x
Then, eventually:
getelementptr float, ptr addrspace(1) %x
Importantly, the massive amount of test case churn has been automated by
same crappy python code. I had to manually update a few test cases that
wouldn't fit the script's model (r228970,r229196,r229197,r229198). The
python script just massages stdin and writes the result to stdout, I
then wrapped that in a shell script to handle replacing files, then
using the usual find+xargs to migrate all the files.
update.py:
import fileinput
import sys
import re
ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
def conv(match, line):
if not match:
return line
line = match.groups()[0]
if len(match.groups()[5]) == 0:
line += match.groups()[2]
line += match.groups()[3]
line += ", "
line += match.groups()[1]
line += "\n"
return line
for line in sys.stdin:
if line.find("getelementptr ") == line.find("getelementptr inbounds"):
if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("):
line = conv(re.match(ibrep, line), line)
elif line.find("getelementptr ") != line.find("getelementptr ("):
line = conv(re.match(normrep, line), line)
sys.stdout.write(line)
apply.sh:
for name in "$@"
do
python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name"
rm -f "$name.tmp"
done
The actual commands:
From llvm/src:
find test/ -name *.ll | xargs ./apply.sh
From llvm/src/tools/clang:
find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}"
From llvm/src/tools/polly:
find test/ -name *.ll | xargs ./apply.sh
After that, check-all (with llvm, clang, clang-tools-extra, lld,
compiler-rt, and polly all checked out).
The extra 'rm' in the apply.sh script is due to a few files in clang's test
suite using interesting unicode stuff that my python script was throwing
exceptions on. None of those files needed to be migrated, so it seemed
sufficient to ignore those cases.
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7636
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230786 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 19:29:02 +00:00
|
|
|
%X = getelementptr i8, i8* %p, i32 4
|
2015-02-27 21:17:42 +00:00
|
|
|
%Y = load i8, i8* %X
|
2009-12-09 02:43:05 +00:00
|
|
|
br label %loop
|
|
|
|
|
|
|
|
|
|
loop:
|
|
|
|
|
%i = phi i32 [4, %entry], [192, %loop]
|
[opaque pointer type] Add textual IR support for explicit type parameter to getelementptr instruction
One of several parallel first steps to remove the target type of pointers,
replacing them with a single opaque pointer type.
This adds an explicit type parameter to the gep instruction so that when the
first parameter becomes an opaque pointer type, the type to gep through is
still available to the instructions.
* This doesn't modify gep operators, only instructions (operators will be
handled separately)
* Textual IR changes only. Bitcode (including upgrade) and changing the
in-memory representation will be in separate changes.
* geps of vectors are transformed as:
getelementptr <4 x float*> %x, ...
->getelementptr float, <4 x float*> %x, ...
Then, once the opaque pointer type is introduced, this will ultimately look
like:
getelementptr float, <4 x ptr> %x
with the unambiguous interpretation that it is a vector of pointers to float.
* address spaces remain on the pointer, not the type:
getelementptr float addrspace(1)* %x
->getelementptr float, float addrspace(1)* %x
Then, eventually:
getelementptr float, ptr addrspace(1) %x
Importantly, the massive amount of test case churn has been automated by
same crappy python code. I had to manually update a few test cases that
wouldn't fit the script's model (r228970,r229196,r229197,r229198). The
python script just massages stdin and writes the result to stdout, I
then wrapped that in a shell script to handle replacing files, then
using the usual find+xargs to migrate all the files.
update.py:
import fileinput
import sys
import re
ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
def conv(match, line):
if not match:
return line
line = match.groups()[0]
if len(match.groups()[5]) == 0:
line += match.groups()[2]
line += match.groups()[3]
line += ", "
line += match.groups()[1]
line += "\n"
return line
for line in sys.stdin:
if line.find("getelementptr ") == line.find("getelementptr inbounds"):
if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("):
line = conv(re.match(ibrep, line), line)
elif line.find("getelementptr ") != line.find("getelementptr ("):
line = conv(re.match(normrep, line), line)
sys.stdout.write(line)
apply.sh:
for name in "$@"
do
python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name"
rm -f "$name.tmp"
done
The actual commands:
From llvm/src:
find test/ -name *.ll | xargs ./apply.sh
From llvm/src/tools/clang:
find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}"
From llvm/src/tools/polly:
find test/ -name *.ll | xargs ./apply.sh
After that, check-all (with llvm, clang, clang-tools-extra, lld,
compiler-rt, and polly all checked out).
The extra 'rm' in the apply.sh script is due to a few files in clang's test
suite using interesting unicode stuff that my python script was throwing
exceptions on. None of those files needed to be migrated, so it seemed
sufficient to ignore those cases.
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7636
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230786 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 19:29:02 +00:00
|
|
|
%X2 = getelementptr i8, i8* %p, i32 %i
|
2015-02-27 21:17:42 +00:00
|
|
|
%Y2 = load i8, i8* %X2
|
2009-12-09 18:21:46 +00:00
|
|
|
|
|
|
|
|
; CHECK: loop:
|
|
|
|
|
; CHECK-NEXT: %Y2 = phi i8 [ %Y, %entry ], [ 0, %loop ]
|
|
|
|
|
; CHECK-NOT: load i8
|
2009-12-09 02:43:05 +00:00
|
|
|
|
|
|
|
|
%cond = call i1 @cond2()
|
|
|
|
|
|
2009-12-09 18:21:46 +00:00
|
|
|
%Z = bitcast i8 *%X3 to i32*
|
2009-12-09 02:43:05 +00:00
|
|
|
store i32 0, i32* %Z
|
|
|
|
|
br i1 %cond, label %loop, label %out
|
|
|
|
|
|
|
|
|
|
out:
|
|
|
|
|
%R = add i8 %Y, %Y2
|
|
|
|
|
ret i8 %R
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
define i8 @phi_trans5(i8* %p) {
|
2013-07-14 01:42:54 +00:00
|
|
|
; CHECK-LABEL: @phi_trans5(
|
2009-12-09 02:43:05 +00:00
|
|
|
entry:
|
2009-12-09 18:21:46 +00:00
|
|
|
|
[opaque pointer type] Add textual IR support for explicit type parameter to getelementptr instruction
One of several parallel first steps to remove the target type of pointers,
replacing them with a single opaque pointer type.
This adds an explicit type parameter to the gep instruction so that when the
first parameter becomes an opaque pointer type, the type to gep through is
still available to the instructions.
* This doesn't modify gep operators, only instructions (operators will be
handled separately)
* Textual IR changes only. Bitcode (including upgrade) and changing the
in-memory representation will be in separate changes.
* geps of vectors are transformed as:
getelementptr <4 x float*> %x, ...
->getelementptr float, <4 x float*> %x, ...
Then, once the opaque pointer type is introduced, this will ultimately look
like:
getelementptr float, <4 x ptr> %x
with the unambiguous interpretation that it is a vector of pointers to float.
* address spaces remain on the pointer, not the type:
getelementptr float addrspace(1)* %x
->getelementptr float, float addrspace(1)* %x
Then, eventually:
getelementptr float, ptr addrspace(1) %x
Importantly, the massive amount of test case churn has been automated by
same crappy python code. I had to manually update a few test cases that
wouldn't fit the script's model (r228970,r229196,r229197,r229198). The
python script just massages stdin and writes the result to stdout, I
then wrapped that in a shell script to handle replacing files, then
using the usual find+xargs to migrate all the files.
update.py:
import fileinput
import sys
import re
ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
def conv(match, line):
if not match:
return line
line = match.groups()[0]
if len(match.groups()[5]) == 0:
line += match.groups()[2]
line += match.groups()[3]
line += ", "
line += match.groups()[1]
line += "\n"
return line
for line in sys.stdin:
if line.find("getelementptr ") == line.find("getelementptr inbounds"):
if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("):
line = conv(re.match(ibrep, line), line)
elif line.find("getelementptr ") != line.find("getelementptr ("):
line = conv(re.match(normrep, line), line)
sys.stdout.write(line)
apply.sh:
for name in "$@"
do
python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name"
rm -f "$name.tmp"
done
The actual commands:
From llvm/src:
find test/ -name *.ll | xargs ./apply.sh
From llvm/src/tools/clang:
find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}"
From llvm/src/tools/polly:
find test/ -name *.ll | xargs ./apply.sh
After that, check-all (with llvm, clang, clang-tools-extra, lld,
compiler-rt, and polly all checked out).
The extra 'rm' in the apply.sh script is due to a few files in clang's test
suite using interesting unicode stuff that my python script was throwing
exceptions on. None of those files needed to be migrated, so it seemed
sufficient to ignore those cases.
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7636
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230786 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 19:29:02 +00:00
|
|
|
%X4 = getelementptr i8, i8* %p, i32 2
|
2009-12-09 02:43:05 +00:00
|
|
|
store i8 19, i8* %X4
|
|
|
|
|
|
[opaque pointer type] Add textual IR support for explicit type parameter to getelementptr instruction
One of several parallel first steps to remove the target type of pointers,
replacing them with a single opaque pointer type.
This adds an explicit type parameter to the gep instruction so that when the
first parameter becomes an opaque pointer type, the type to gep through is
still available to the instructions.
* This doesn't modify gep operators, only instructions (operators will be
handled separately)
* Textual IR changes only. Bitcode (including upgrade) and changing the
in-memory representation will be in separate changes.
* geps of vectors are transformed as:
getelementptr <4 x float*> %x, ...
->getelementptr float, <4 x float*> %x, ...
Then, once the opaque pointer type is introduced, this will ultimately look
like:
getelementptr float, <4 x ptr> %x
with the unambiguous interpretation that it is a vector of pointers to float.
* address spaces remain on the pointer, not the type:
getelementptr float addrspace(1)* %x
->getelementptr float, float addrspace(1)* %x
Then, eventually:
getelementptr float, ptr addrspace(1) %x
Importantly, the massive amount of test case churn has been automated by
same crappy python code. I had to manually update a few test cases that
wouldn't fit the script's model (r228970,r229196,r229197,r229198). The
python script just massages stdin and writes the result to stdout, I
then wrapped that in a shell script to handle replacing files, then
using the usual find+xargs to migrate all the files.
update.py:
import fileinput
import sys
import re
ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
def conv(match, line):
if not match:
return line
line = match.groups()[0]
if len(match.groups()[5]) == 0:
line += match.groups()[2]
line += match.groups()[3]
line += ", "
line += match.groups()[1]
line += "\n"
return line
for line in sys.stdin:
if line.find("getelementptr ") == line.find("getelementptr inbounds"):
if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("):
line = conv(re.match(ibrep, line), line)
elif line.find("getelementptr ") != line.find("getelementptr ("):
line = conv(re.match(normrep, line), line)
sys.stdout.write(line)
apply.sh:
for name in "$@"
do
python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name"
rm -f "$name.tmp"
done
The actual commands:
From llvm/src:
find test/ -name *.ll | xargs ./apply.sh
From llvm/src/tools/clang:
find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}"
From llvm/src/tools/polly:
find test/ -name *.ll | xargs ./apply.sh
After that, check-all (with llvm, clang, clang-tools-extra, lld,
compiler-rt, and polly all checked out).
The extra 'rm' in the apply.sh script is due to a few files in clang's test
suite using interesting unicode stuff that my python script was throwing
exceptions on. None of those files needed to be migrated, so it seemed
sufficient to ignore those cases.
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7636
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230786 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 19:29:02 +00:00
|
|
|
%X = getelementptr i8, i8* %p, i32 4
|
2015-02-27 21:17:42 +00:00
|
|
|
%Y = load i8, i8* %X
|
2009-12-09 02:43:05 +00:00
|
|
|
br label %loop
|
|
|
|
|
|
|
|
|
|
loop:
|
|
|
|
|
%i = phi i32 [4, %entry], [3, %cont]
|
[opaque pointer type] Add textual IR support for explicit type parameter to getelementptr instruction
One of several parallel first steps to remove the target type of pointers,
replacing them with a single opaque pointer type.
This adds an explicit type parameter to the gep instruction so that when the
first parameter becomes an opaque pointer type, the type to gep through is
still available to the instructions.
* This doesn't modify gep operators, only instructions (operators will be
handled separately)
* Textual IR changes only. Bitcode (including upgrade) and changing the
in-memory representation will be in separate changes.
* geps of vectors are transformed as:
getelementptr <4 x float*> %x, ...
->getelementptr float, <4 x float*> %x, ...
Then, once the opaque pointer type is introduced, this will ultimately look
like:
getelementptr float, <4 x ptr> %x
with the unambiguous interpretation that it is a vector of pointers to float.
* address spaces remain on the pointer, not the type:
getelementptr float addrspace(1)* %x
->getelementptr float, float addrspace(1)* %x
Then, eventually:
getelementptr float, ptr addrspace(1) %x
Importantly, the massive amount of test case churn has been automated by
same crappy python code. I had to manually update a few test cases that
wouldn't fit the script's model (r228970,r229196,r229197,r229198). The
python script just massages stdin and writes the result to stdout, I
then wrapped that in a shell script to handle replacing files, then
using the usual find+xargs to migrate all the files.
update.py:
import fileinput
import sys
import re
ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
def conv(match, line):
if not match:
return line
line = match.groups()[0]
if len(match.groups()[5]) == 0:
line += match.groups()[2]
line += match.groups()[3]
line += ", "
line += match.groups()[1]
line += "\n"
return line
for line in sys.stdin:
if line.find("getelementptr ") == line.find("getelementptr inbounds"):
if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("):
line = conv(re.match(ibrep, line), line)
elif line.find("getelementptr ") != line.find("getelementptr ("):
line = conv(re.match(normrep, line), line)
sys.stdout.write(line)
apply.sh:
for name in "$@"
do
python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name"
rm -f "$name.tmp"
done
The actual commands:
From llvm/src:
find test/ -name *.ll | xargs ./apply.sh
From llvm/src/tools/clang:
find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}"
From llvm/src/tools/polly:
find test/ -name *.ll | xargs ./apply.sh
After that, check-all (with llvm, clang, clang-tools-extra, lld,
compiler-rt, and polly all checked out).
The extra 'rm' in the apply.sh script is due to a few files in clang's test
suite using interesting unicode stuff that my python script was throwing
exceptions on. None of those files needed to be migrated, so it seemed
sufficient to ignore those cases.
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7636
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230786 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 19:29:02 +00:00
|
|
|
%X2 = getelementptr i8, i8* %p, i32 %i
|
2015-02-27 21:17:42 +00:00
|
|
|
%Y2 = load i8, i8* %X2 ; Ensure this load is not being incorrectly replaced.
|
2009-12-09 02:43:05 +00:00
|
|
|
%cond = call i1 @cond2()
|
|
|
|
|
br i1 %cond, label %cont, label %out
|
|
|
|
|
|
|
|
|
|
cont:
|
[opaque pointer type] Add textual IR support for explicit type parameter to getelementptr instruction
One of several parallel first steps to remove the target type of pointers,
replacing them with a single opaque pointer type.
This adds an explicit type parameter to the gep instruction so that when the
first parameter becomes an opaque pointer type, the type to gep through is
still available to the instructions.
* This doesn't modify gep operators, only instructions (operators will be
handled separately)
* Textual IR changes only. Bitcode (including upgrade) and changing the
in-memory representation will be in separate changes.
* geps of vectors are transformed as:
getelementptr <4 x float*> %x, ...
->getelementptr float, <4 x float*> %x, ...
Then, once the opaque pointer type is introduced, this will ultimately look
like:
getelementptr float, <4 x ptr> %x
with the unambiguous interpretation that it is a vector of pointers to float.
* address spaces remain on the pointer, not the type:
getelementptr float addrspace(1)* %x
->getelementptr float, float addrspace(1)* %x
Then, eventually:
getelementptr float, ptr addrspace(1) %x
Importantly, the massive amount of test case churn has been automated by
same crappy python code. I had to manually update a few test cases that
wouldn't fit the script's model (r228970,r229196,r229197,r229198). The
python script just massages stdin and writes the result to stdout, I
then wrapped that in a shell script to handle replacing files, then
using the usual find+xargs to migrate all the files.
update.py:
import fileinput
import sys
import re
ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
def conv(match, line):
if not match:
return line
line = match.groups()[0]
if len(match.groups()[5]) == 0:
line += match.groups()[2]
line += match.groups()[3]
line += ", "
line += match.groups()[1]
line += "\n"
return line
for line in sys.stdin:
if line.find("getelementptr ") == line.find("getelementptr inbounds"):
if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("):
line = conv(re.match(ibrep, line), line)
elif line.find("getelementptr ") != line.find("getelementptr ("):
line = conv(re.match(normrep, line), line)
sys.stdout.write(line)
apply.sh:
for name in "$@"
do
python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name"
rm -f "$name.tmp"
done
The actual commands:
From llvm/src:
find test/ -name *.ll | xargs ./apply.sh
From llvm/src/tools/clang:
find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}"
From llvm/src/tools/polly:
find test/ -name *.ll | xargs ./apply.sh
After that, check-all (with llvm, clang, clang-tools-extra, lld,
compiler-rt, and polly all checked out).
The extra 'rm' in the apply.sh script is due to a few files in clang's test
suite using interesting unicode stuff that my python script was throwing
exceptions on. None of those files needed to be migrated, so it seemed
sufficient to ignore those cases.
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7636
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230786 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 19:29:02 +00:00
|
|
|
%Z = getelementptr i8, i8* %X2, i32 -1
|
2009-12-09 02:43:05 +00:00
|
|
|
%Z2 = bitcast i8 *%Z to i32*
|
|
|
|
|
store i32 50462976, i32* %Z2 ;; (1 << 8) | (2 << 16) | (3 << 24)
|
2009-12-09 18:21:46 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
; CHECK: store i32
|
[opaque pointer type] Add textual IR support for explicit type parameter to getelementptr instruction
One of several parallel first steps to remove the target type of pointers,
replacing them with a single opaque pointer type.
This adds an explicit type parameter to the gep instruction so that when the
first parameter becomes an opaque pointer type, the type to gep through is
still available to the instructions.
* This doesn't modify gep operators, only instructions (operators will be
handled separately)
* Textual IR changes only. Bitcode (including upgrade) and changing the
in-memory representation will be in separate changes.
* geps of vectors are transformed as:
getelementptr <4 x float*> %x, ...
->getelementptr float, <4 x float*> %x, ...
Then, once the opaque pointer type is introduced, this will ultimately look
like:
getelementptr float, <4 x ptr> %x
with the unambiguous interpretation that it is a vector of pointers to float.
* address spaces remain on the pointer, not the type:
getelementptr float addrspace(1)* %x
->getelementptr float, float addrspace(1)* %x
Then, eventually:
getelementptr float, ptr addrspace(1) %x
Importantly, the massive amount of test case churn has been automated by
same crappy python code. I had to manually update a few test cases that
wouldn't fit the script's model (r228970,r229196,r229197,r229198). The
python script just massages stdin and writes the result to stdout, I
then wrapped that in a shell script to handle replacing files, then
using the usual find+xargs to migrate all the files.
update.py:
import fileinput
import sys
import re
ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
def conv(match, line):
if not match:
return line
line = match.groups()[0]
if len(match.groups()[5]) == 0:
line += match.groups()[2]
line += match.groups()[3]
line += ", "
line += match.groups()[1]
line += "\n"
return line
for line in sys.stdin:
if line.find("getelementptr ") == line.find("getelementptr inbounds"):
if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("):
line = conv(re.match(ibrep, line), line)
elif line.find("getelementptr ") != line.find("getelementptr ("):
line = conv(re.match(normrep, line), line)
sys.stdout.write(line)
apply.sh:
for name in "$@"
do
python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name"
rm -f "$name.tmp"
done
The actual commands:
From llvm/src:
find test/ -name *.ll | xargs ./apply.sh
From llvm/src/tools/clang:
find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}"
From llvm/src/tools/polly:
find test/ -name *.ll | xargs ./apply.sh
After that, check-all (with llvm, clang, clang-tools-extra, lld,
compiler-rt, and polly all checked out).
The extra 'rm' in the apply.sh script is due to a few files in clang's test
suite using interesting unicode stuff that my python script was throwing
exceptions on. None of those files needed to be migrated, so it seemed
sufficient to ignore those cases.
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7636
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230786 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 19:29:02 +00:00
|
|
|
; CHECK-NEXT: getelementptr i8, i8* %p, i32 3
|
2015-02-27 21:17:42 +00:00
|
|
|
; CHECK-NEXT: load i8, i8*
|
2009-12-09 02:43:05 +00:00
|
|
|
br label %loop
|
|
|
|
|
|
|
|
|
|
out:
|
|
|
|
|
%R = add i8 %Y, %Y2
|
|
|
|
|
ret i8 %R
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
2010-03-25 05:58:19 +00:00
|
|
|
; PR6642
|
|
|
|
|
define i32 @memset_to_load() nounwind readnone {
|
|
|
|
|
entry:
|
|
|
|
|
%x = alloca [256 x i32], align 4 ; <[256 x i32]*> [#uses=2]
|
|
|
|
|
%tmp = bitcast [256 x i32]* %x to i8* ; <i8*> [#uses=1]
|
2011-06-18 06:05:24 +00:00
|
|
|
call void @llvm.memset.p0i8.i64(i8* %tmp, i8 0, i64 1024, i32 4, i1 false)
|
[opaque pointer type] Add textual IR support for explicit type parameter to getelementptr instruction
One of several parallel first steps to remove the target type of pointers,
replacing them with a single opaque pointer type.
This adds an explicit type parameter to the gep instruction so that when the
first parameter becomes an opaque pointer type, the type to gep through is
still available to the instructions.
* This doesn't modify gep operators, only instructions (operators will be
handled separately)
* Textual IR changes only. Bitcode (including upgrade) and changing the
in-memory representation will be in separate changes.
* geps of vectors are transformed as:
getelementptr <4 x float*> %x, ...
->getelementptr float, <4 x float*> %x, ...
Then, once the opaque pointer type is introduced, this will ultimately look
like:
getelementptr float, <4 x ptr> %x
with the unambiguous interpretation that it is a vector of pointers to float.
* address spaces remain on the pointer, not the type:
getelementptr float addrspace(1)* %x
->getelementptr float, float addrspace(1)* %x
Then, eventually:
getelementptr float, ptr addrspace(1) %x
Importantly, the massive amount of test case churn has been automated by
same crappy python code. I had to manually update a few test cases that
wouldn't fit the script's model (r228970,r229196,r229197,r229198). The
python script just massages stdin and writes the result to stdout, I
then wrapped that in a shell script to handle replacing files, then
using the usual find+xargs to migrate all the files.
update.py:
import fileinput
import sys
import re
ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
def conv(match, line):
if not match:
return line
line = match.groups()[0]
if len(match.groups()[5]) == 0:
line += match.groups()[2]
line += match.groups()[3]
line += ", "
line += match.groups()[1]
line += "\n"
return line
for line in sys.stdin:
if line.find("getelementptr ") == line.find("getelementptr inbounds"):
if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("):
line = conv(re.match(ibrep, line), line)
elif line.find("getelementptr ") != line.find("getelementptr ("):
line = conv(re.match(normrep, line), line)
sys.stdout.write(line)
apply.sh:
for name in "$@"
do
python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name"
rm -f "$name.tmp"
done
The actual commands:
From llvm/src:
find test/ -name *.ll | xargs ./apply.sh
From llvm/src/tools/clang:
find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}"
From llvm/src/tools/polly:
find test/ -name *.ll | xargs ./apply.sh
After that, check-all (with llvm, clang, clang-tools-extra, lld,
compiler-rt, and polly all checked out).
The extra 'rm' in the apply.sh script is due to a few files in clang's test
suite using interesting unicode stuff that my python script was throwing
exceptions on. None of those files needed to be migrated, so it seemed
sufficient to ignore those cases.
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7636
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230786 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 19:29:02 +00:00
|
|
|
%arraydecay = getelementptr inbounds [256 x i32], [256 x i32]* %x, i32 0, i32 0 ; <i32*>
|
2015-02-27 21:17:42 +00:00
|
|
|
%tmp1 = load i32, i32* %arraydecay ; <i32> [#uses=1]
|
2010-03-25 05:58:19 +00:00
|
|
|
ret i32 %tmp1
|
2013-07-14 01:42:54 +00:00
|
|
|
; CHECK-LABEL: @memset_to_load(
|
2010-03-25 05:58:19 +00:00
|
|
|
; CHECK: ret i32 0
|
|
|
|
|
}
|
2009-12-09 02:43:05 +00:00
|
|
|
|
2011-04-26 01:21:15 +00:00
|
|
|
|
|
|
|
|
;;===----------------------------------------------------------------------===;;
|
|
|
|
|
;; Load -> Load forwarding in partial alias case.
|
|
|
|
|
;;===----------------------------------------------------------------------===;;
|
|
|
|
|
|
|
|
|
|
define i32 @load_load_partial_alias(i8* %P) nounwind ssp {
|
|
|
|
|
entry:
|
|
|
|
|
%0 = bitcast i8* %P to i32*
|
2015-02-27 21:17:42 +00:00
|
|
|
%tmp2 = load i32, i32* %0
|
[opaque pointer type] Add textual IR support for explicit type parameter to getelementptr instruction
One of several parallel first steps to remove the target type of pointers,
replacing them with a single opaque pointer type.
This adds an explicit type parameter to the gep instruction so that when the
first parameter becomes an opaque pointer type, the type to gep through is
still available to the instructions.
* This doesn't modify gep operators, only instructions (operators will be
handled separately)
* Textual IR changes only. Bitcode (including upgrade) and changing the
in-memory representation will be in separate changes.
* geps of vectors are transformed as:
getelementptr <4 x float*> %x, ...
->getelementptr float, <4 x float*> %x, ...
Then, once the opaque pointer type is introduced, this will ultimately look
like:
getelementptr float, <4 x ptr> %x
with the unambiguous interpretation that it is a vector of pointers to float.
* address spaces remain on the pointer, not the type:
getelementptr float addrspace(1)* %x
->getelementptr float, float addrspace(1)* %x
Then, eventually:
getelementptr float, ptr addrspace(1) %x
Importantly, the massive amount of test case churn has been automated by
same crappy python code. I had to manually update a few test cases that
wouldn't fit the script's model (r228970,r229196,r229197,r229198). The
python script just massages stdin and writes the result to stdout, I
then wrapped that in a shell script to handle replacing files, then
using the usual find+xargs to migrate all the files.
update.py:
import fileinput
import sys
import re
ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
def conv(match, line):
if not match:
return line
line = match.groups()[0]
if len(match.groups()[5]) == 0:
line += match.groups()[2]
line += match.groups()[3]
line += ", "
line += match.groups()[1]
line += "\n"
return line
for line in sys.stdin:
if line.find("getelementptr ") == line.find("getelementptr inbounds"):
if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("):
line = conv(re.match(ibrep, line), line)
elif line.find("getelementptr ") != line.find("getelementptr ("):
line = conv(re.match(normrep, line), line)
sys.stdout.write(line)
apply.sh:
for name in "$@"
do
python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name"
rm -f "$name.tmp"
done
The actual commands:
From llvm/src:
find test/ -name *.ll | xargs ./apply.sh
From llvm/src/tools/clang:
find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}"
From llvm/src/tools/polly:
find test/ -name *.ll | xargs ./apply.sh
After that, check-all (with llvm, clang, clang-tools-extra, lld,
compiler-rt, and polly all checked out).
The extra 'rm' in the apply.sh script is due to a few files in clang's test
suite using interesting unicode stuff that my python script was throwing
exceptions on. None of those files needed to be migrated, so it seemed
sufficient to ignore those cases.
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7636
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230786 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 19:29:02 +00:00
|
|
|
%add.ptr = getelementptr inbounds i8, i8* %P, i64 1
|
2015-02-27 21:17:42 +00:00
|
|
|
%tmp5 = load i8, i8* %add.ptr
|
2011-04-26 01:21:15 +00:00
|
|
|
%conv = zext i8 %tmp5 to i32
|
|
|
|
|
%add = add nsw i32 %tmp2, %conv
|
|
|
|
|
ret i32 %add
|
|
|
|
|
|
2013-07-14 01:42:54 +00:00
|
|
|
; TEMPORARILYDISABLED-LABEL: @load_load_partial_alias(
|
2015-02-27 21:17:42 +00:00
|
|
|
; TEMPORARILYDISABLED: load i32, i32*
|
2011-06-04 06:48:50 +00:00
|
|
|
; TEMPORARILYDISABLED-NOT: load
|
|
|
|
|
; TEMPORARILYDISABLED: lshr i32 {{.*}}, 8
|
|
|
|
|
; TEMPORARILYDISABLED-NOT: load
|
|
|
|
|
; TEMPORARILYDISABLED: trunc i32 {{.*}} to i8
|
|
|
|
|
; TEMPORARILYDISABLED-NOT: load
|
|
|
|
|
; TEMPORARILYDISABLED: ret i32
|
2011-04-26 01:21:15 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
2011-04-26 17:41:02 +00:00
|
|
|
; Cross block partial alias case.
|
|
|
|
|
define i32 @load_load_partial_alias_cross_block(i8* %P) nounwind ssp {
|
|
|
|
|
entry:
|
|
|
|
|
%xx = bitcast i8* %P to i32*
|
2015-02-27 21:17:42 +00:00
|
|
|
%x1 = load i32, i32* %xx, align 4
|
2011-04-26 17:41:02 +00:00
|
|
|
%cmp = icmp eq i32 %x1, 127
|
|
|
|
|
br i1 %cmp, label %land.lhs.true, label %if.end
|
|
|
|
|
|
|
|
|
|
land.lhs.true: ; preds = %entry
|
[opaque pointer type] Add textual IR support for explicit type parameter to getelementptr instruction
One of several parallel first steps to remove the target type of pointers,
replacing them with a single opaque pointer type.
This adds an explicit type parameter to the gep instruction so that when the
first parameter becomes an opaque pointer type, the type to gep through is
still available to the instructions.
* This doesn't modify gep operators, only instructions (operators will be
handled separately)
* Textual IR changes only. Bitcode (including upgrade) and changing the
in-memory representation will be in separate changes.
* geps of vectors are transformed as:
getelementptr <4 x float*> %x, ...
->getelementptr float, <4 x float*> %x, ...
Then, once the opaque pointer type is introduced, this will ultimately look
like:
getelementptr float, <4 x ptr> %x
with the unambiguous interpretation that it is a vector of pointers to float.
* address spaces remain on the pointer, not the type:
getelementptr float addrspace(1)* %x
->getelementptr float, float addrspace(1)* %x
Then, eventually:
getelementptr float, ptr addrspace(1) %x
Importantly, the massive amount of test case churn has been automated by
same crappy python code. I had to manually update a few test cases that
wouldn't fit the script's model (r228970,r229196,r229197,r229198). The
python script just massages stdin and writes the result to stdout, I
then wrapped that in a shell script to handle replacing files, then
using the usual find+xargs to migrate all the files.
update.py:
import fileinput
import sys
import re
ibrep = re.compile(r"(^.*?[^%\w]getelementptr inbounds )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
normrep = re.compile( r"(^.*?[^%\w]getelementptr )(((?:<\d* x )?)(.*?)(| addrspace\(\d\)) *\*(|>)(?:$| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$))")
def conv(match, line):
if not match:
return line
line = match.groups()[0]
if len(match.groups()[5]) == 0:
line += match.groups()[2]
line += match.groups()[3]
line += ", "
line += match.groups()[1]
line += "\n"
return line
for line in sys.stdin:
if line.find("getelementptr ") == line.find("getelementptr inbounds"):
if line.find("getelementptr inbounds") != line.find("getelementptr inbounds ("):
line = conv(re.match(ibrep, line), line)
elif line.find("getelementptr ") != line.find("getelementptr ("):
line = conv(re.match(normrep, line), line)
sys.stdout.write(line)
apply.sh:
for name in "$@"
do
python3 `dirname "$0"`/update.py < "$name" > "$name.tmp" && mv "$name.tmp" "$name"
rm -f "$name.tmp"
done
The actual commands:
From llvm/src:
find test/ -name *.ll | xargs ./apply.sh
From llvm/src/tools/clang:
find test/ -name *.mm -o -name *.m -o -name *.cpp -o -name *.c | xargs -I '{}' ../../apply.sh "{}"
From llvm/src/tools/polly:
find test/ -name *.ll | xargs ./apply.sh
After that, check-all (with llvm, clang, clang-tools-extra, lld,
compiler-rt, and polly all checked out).
The extra 'rm' in the apply.sh script is due to a few files in clang's test
suite using interesting unicode stuff that my python script was throwing
exceptions on. None of those files needed to be migrated, so it seemed
sufficient to ignore those cases.
Reviewers: rafael, dexonsmith, grosser
Differential Revision: http://reviews.llvm.org/D7636
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@230786 91177308-0d34-0410-b5e6-96231b3b80d8
2015-02-27 19:29:02 +00:00
|
|
|
%arrayidx4 = getelementptr inbounds i8, i8* %P, i64 1
|
2015-02-27 21:17:42 +00:00
|
|
|
%tmp5 = load i8, i8* %arrayidx4, align 1
|
2011-04-26 17:41:02 +00:00
|
|
|
%conv6 = zext i8 %tmp5 to i32
|
|
|
|
|
ret i32 %conv6
|
|
|
|
|
|
|
|
|
|
if.end:
|
|
|
|
|
ret i32 52
|
2013-07-14 01:42:54 +00:00
|
|
|
; TEMPORARILY_DISABLED-LABEL: @load_load_partial_alias_cross_block(
|
2011-06-04 06:48:50 +00:00
|
|
|
; TEMPORARILY_DISABLED: land.lhs.true:
|
|
|
|
|
; TEMPORARILY_DISABLED-NOT: load i8
|
|
|
|
|
; TEMPORARILY_DISABLED: ret i32 %conv6
|
2011-04-26 17:41:02 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
2011-04-28 07:29:08 +00:00
|
|
|
;;===----------------------------------------------------------------------===;;
|
|
|
|
|
;; Load Widening
|
|
|
|
|
;;===----------------------------------------------------------------------===;;
|
|
|
|
|
|
2011-04-28 18:15:47 +00:00
|
|
|
%widening1 = type { i32, i8, i8, i8, i8 }
|
2011-04-28 07:29:08 +00:00
|
|
|
|
|
|
|
|
@f = global %widening1 zeroinitializer, align 4
|
|
|
|
|
|
2011-04-28 18:15:47 +00:00
|
|
|
define i32 @test_widening1(i8* %P) nounwind ssp noredzone {
|
2011-04-28 07:29:08 +00:00
|
|
|
entry:
|
2015-03-13 18:20:45 +00:00
|
|
|
%tmp = load i8, i8* getelementptr inbounds (%widening1, %widening1* @f, i64 0, i32 1), align 4
|
2011-04-28 07:29:08 +00:00
|
|
|
%conv = zext i8 %tmp to i32
|
2015-03-13 18:20:45 +00:00
|
|
|
%tmp1 = load i8, i8* getelementptr inbounds (%widening1, %widening1* @f, i64 0, i32 2), align 1
|
2011-04-28 07:29:08 +00:00
|
|
|
%conv2 = zext i8 %tmp1 to i32
|
|
|
|
|
%add = add nsw i32 %conv, %conv2
|
|
|
|
|
ret i32 %add
|
2013-07-14 01:42:54 +00:00
|
|
|
; CHECK-LABEL: @test_widening1(
|
2011-04-28 07:29:08 +00:00
|
|
|
; CHECK-NOT: load
|
2015-02-27 21:17:42 +00:00
|
|
|
; CHECK: load i16, i16*
|
2011-04-28 07:29:08 +00:00
|
|
|
; CHECK-NOT: load
|
2012-08-17 12:28:26 +00:00
|
|
|
; CHECK: ret i32
|
2011-04-28 07:29:08 +00:00
|
|
|
}
|
|
|
|
|
|
2011-04-28 18:15:47 +00:00
|
|
|
define i32 @test_widening2() nounwind ssp noredzone {
|
|
|
|
|
entry:
|
2015-03-13 18:20:45 +00:00
|
|
|
%tmp = load i8, i8* getelementptr inbounds (%widening1, %widening1* @f, i64 0, i32 1), align 4
|
2011-04-28 18:15:47 +00:00
|
|
|
%conv = zext i8 %tmp to i32
|
2015-03-13 18:20:45 +00:00
|
|
|
%tmp1 = load i8, i8* getelementptr inbounds (%widening1, %widening1* @f, i64 0, i32 2), align 1
|
2011-04-28 18:15:47 +00:00
|
|
|
%conv2 = zext i8 %tmp1 to i32
|
|
|
|
|
%add = add nsw i32 %conv, %conv2
|
|
|
|
|
|
2015-03-13 18:20:45 +00:00
|
|
|
%tmp2 = load i8, i8* getelementptr inbounds (%widening1, %widening1* @f, i64 0, i32 3), align 2
|
2011-04-28 18:15:47 +00:00
|
|
|
%conv3 = zext i8 %tmp2 to i32
|
|
|
|
|
%add2 = add nsw i32 %add, %conv3
|
|
|
|
|
|
2015-03-13 18:20:45 +00:00
|
|
|
%tmp3 = load i8, i8* getelementptr inbounds (%widening1, %widening1* @f, i64 0, i32 4), align 1
|
2011-04-28 18:15:47 +00:00
|
|
|
%conv4 = zext i8 %tmp3 to i32
|
|
|
|
|
%add3 = add nsw i32 %add2, %conv3
|
|
|
|
|
|
|
|
|
|
ret i32 %add3
|
2013-07-14 01:42:54 +00:00
|
|
|
; CHECK-LABEL: @test_widening2(
|
2011-04-28 18:15:47 +00:00
|
|
|
; CHECK-NOT: load
|
2015-02-27 21:17:42 +00:00
|
|
|
; CHECK: load i32, i32*
|
2011-04-28 18:15:47 +00:00
|
|
|
; CHECK-NOT: load
|
2012-08-17 12:28:26 +00:00
|
|
|
; CHECK: ret i32
|
2011-04-28 18:15:47 +00:00
|
|
|
}
|
2011-04-26 01:21:15 +00:00
|
|
|
|
2011-06-18 06:05:24 +00:00
|
|
|
declare void @llvm.memset.p0i8.i64(i8* nocapture, i8, i64, i32, i1) nounwind
|
|
|
|
|
|
|
|
|
|
declare void @llvm.memcpy.p0i8.p0i8.i64(i8* nocapture, i8* nocapture, i64, i32, i1) nounwind
|
2013-10-30 19:05:41 +00:00
|
|
|
declare void @llvm.memcpy.p0i8.p1i8.i64(i8* nocapture, i8 addrspace(1)* nocapture, i64, i32, i1) nounwind
|
|
|
|
|
|
2011-06-18 06:05:24 +00:00
|
|
|
|
2011-11-14 22:49:42 +00:00
|
|
|
;;===----------------------------------------------------------------------===;;
|
|
|
|
|
;; Load -> Store dependency which isn't interfered with by a call that happens
|
|
|
|
|
;; before the pointer was captured.
|
|
|
|
|
;;===----------------------------------------------------------------------===;;
|
|
|
|
|
|
|
|
|
|
%class.X = type { [8 x i8] }
|
|
|
|
|
|
|
|
|
|
@_ZTV1X = weak_odr constant [5 x i8*] zeroinitializer
|
|
|
|
|
@_ZTV1Y = weak_odr constant [5 x i8*] zeroinitializer
|
|
|
|
|
|
|
|
|
|
declare void @use()
|
|
|
|
|
declare void @use3(i8***, i8**)
|
|
|
|
|
|
|
|
|
|
; PR8908
|
|
|
|
|
define void @test_escape1() nounwind {
|
|
|
|
|
%x = alloca i8**, align 8
|
2015-03-13 18:20:45 +00:00
|
|
|
store i8** getelementptr inbounds ([5 x i8*], [5 x i8*]* @_ZTV1X, i64 0, i64 2), i8*** %x, align 8
|
2011-11-14 22:49:42 +00:00
|
|
|
call void @use() nounwind
|
2015-02-27 21:17:42 +00:00
|
|
|
%DEAD = load i8**, i8*** %x, align 8
|
2011-11-14 22:49:42 +00:00
|
|
|
call void @use3(i8*** %x, i8** %DEAD) nounwind
|
|
|
|
|
ret void
|
|
|
|
|
; CHECK: test_escape1
|
|
|
|
|
; CHECK-NOT: DEAD
|
|
|
|
|
; CHECK: ret
|
|
|
|
|
}
|
