[DAGCombiner] do not fold (fmul (fadd X, 1), Y) -> (fmad X, Y, Y) by default
Summary:
When X = 0 and Y = inf, the original code produces inf, but the transformed
code produces nan. So this transform (and its relatives) should only be
used when the no-infs-fp-math flag is explicitly enabled.
Also disable the transform using fmad (intermediate rounding) when unsafe-math
is not enabled, since it can reduce the precision of the result; consider this
example with binary floating point numbers with two bits of mantissa:
x = 1.01
y = 111
x * (y + 1) = 1.01 * 1000 = 1010 (this is the exact result; no rounding occurs at any step)
x * y + x = 1000.11 + 1.01 =r 1000 + 1.01 = 1001.01 =r 1000 (with rounding towards zero)
The example relies on rounding towards zero at least in the second step.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98578
Reviewers: RKSimon, tstellarAMD, spatel, arsenm
Subscribers: wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D26602
llvm-svn: 288506
2016-12-03 00:06:18 +08:00
|
|
|
; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs -fp-contract=fast < %s | FileCheck -check-prefix=SI-NOFMA -check-prefix=SI -check-prefix=FUNC %s
|
|
|
|
|
; RUN: llc -march=amdgcn -mcpu=verde -verify-machineinstrs -fp-contract=fast < %s | FileCheck -check-prefix=SI-NOFMA -check-prefix=SI -check-prefix=FUNC %s
|
|
|
|
|
; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs -fp-contract=fast -enable-no-infs-fp-math -mattr=+fp32-denormals < %s | FileCheck -check-prefix=SI-FMA -check-prefix=SI -check-prefix=FUNC %s
|
|
|
|
|
|
|
|
|
|
; Note: The SI-FMA conversions of type x * (y + 1) --> x * y + x would be
|
|
|
|
|
; beneficial even without fp32 denormals, but they do require no-infs-fp-math
|
|
|
|
|
; for correctness.
|
2015-01-30 03:34:32 +08:00
|
|
|
|
2016-02-11 14:02:01 +08:00
|
|
|
declare i32 @llvm.amdgcn.workitem.id.x() #0
|
2015-01-30 03:34:32 +08:00
|
|
|
declare double @llvm.fabs.f64(double) #0
|
|
|
|
|
declare double @llvm.fma.f64(double, double, double) #0
|
|
|
|
|
declare float @llvm.fma.f32(float, float, float) #0
|
|
|
|
|
|
|
|
|
|
; (fadd (fmul x, y), z) -> (fma x, y, z)
|
|
|
|
|
; FUNC-LABEL: {{^}}combine_to_fma_f64_0:
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[A:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[B:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[C:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:16{{$}}
|
|
|
|
|
; SI: v_fma_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[A]], [[B]], [[C]]
|
|
|
|
|
; SI: buffer_store_dwordx2 [[RESULT]]
|
|
|
|
|
define void @combine_to_fma_f64_0(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) #1 {
|
2016-02-11 14:02:01 +08:00
|
|
|
%tid = tail call i32 @llvm.amdgcn.workitem.id.x() #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
llvm-svn: 230786
2015-02-28 03:29:02 +08:00
|
|
|
%gep.0 = getelementptr double, double addrspace(1)* %in, i32 %tid
|
|
|
|
|
%gep.1 = getelementptr double, double addrspace(1)* %gep.0, i32 1
|
|
|
|
|
%gep.2 = getelementptr double, double addrspace(1)* %gep.0, i32 2
|
|
|
|
|
%gep.out = getelementptr double, double addrspace(1)* %out, i32 %tid
|
2015-01-30 03:34:32 +08:00
|
|
|
|
2016-04-12 21:38:18 +08:00
|
|
|
%a = load volatile double, double addrspace(1)* %gep.0
|
|
|
|
|
%b = load volatile double, double addrspace(1)* %gep.1
|
|
|
|
|
%c = load volatile double, double addrspace(1)* %gep.2
|
2015-01-30 03:34:32 +08:00
|
|
|
|
|
|
|
|
%mul = fmul double %a, %b
|
|
|
|
|
%fma = fadd double %mul, %c
|
|
|
|
|
store double %fma, double addrspace(1)* %gep.out
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; (fadd (fmul x, y), z) -> (fma x, y, z)
|
|
|
|
|
; FUNC-LABEL: {{^}}combine_to_fma_f64_0_2use:
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[A:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[B:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[C:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:16{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[D:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:24{{$}}
|
|
|
|
|
; SI-DAG: v_fma_f64 [[RESULT0:v\[[0-9]+:[0-9]+\]]], [[A]], [[B]], [[C]]
|
|
|
|
|
; SI-DAG: v_fma_f64 [[RESULT1:v\[[0-9]+:[0-9]+\]]], [[A]], [[B]], [[D]]
|
|
|
|
|
; SI-DAG: buffer_store_dwordx2 [[RESULT0]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
|
|
|
|
|
; SI-DAG: buffer_store_dwordx2 [[RESULT1]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
|
|
|
|
|
; SI: s_endpgm
|
|
|
|
|
define void @combine_to_fma_f64_0_2use(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) #1 {
|
2016-02-11 14:02:01 +08:00
|
|
|
%tid = tail call i32 @llvm.amdgcn.workitem.id.x() #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
llvm-svn: 230786
2015-02-28 03:29:02 +08:00
|
|
|
%gep.0 = getelementptr double, double addrspace(1)* %in, i32 %tid
|
|
|
|
|
%gep.1 = getelementptr double, double addrspace(1)* %gep.0, i32 1
|
|
|
|
|
%gep.2 = getelementptr double, double addrspace(1)* %gep.0, i32 2
|
|
|
|
|
%gep.3 = getelementptr double, double addrspace(1)* %gep.0, i32 3
|
|
|
|
|
%gep.out.0 = getelementptr double, double addrspace(1)* %out, i32 %tid
|
|
|
|
|
%gep.out.1 = getelementptr double, double addrspace(1)* %gep.out.0, i32 1
|
2015-01-30 03:34:32 +08:00
|
|
|
|
2016-04-12 21:38:18 +08:00
|
|
|
%a = load volatile double, double addrspace(1)* %gep.0
|
|
|
|
|
%b = load volatile double, double addrspace(1)* %gep.1
|
|
|
|
|
%c = load volatile double, double addrspace(1)* %gep.2
|
|
|
|
|
%d = load volatile double, double addrspace(1)* %gep.3
|
2015-01-30 03:34:32 +08:00
|
|
|
|
|
|
|
|
%mul = fmul double %a, %b
|
|
|
|
|
%fma0 = fadd double %mul, %c
|
|
|
|
|
%fma1 = fadd double %mul, %d
|
2016-04-12 21:38:18 +08:00
|
|
|
store volatile double %fma0, double addrspace(1)* %gep.out.0
|
|
|
|
|
store volatile double %fma1, double addrspace(1)* %gep.out.1
|
2015-01-30 03:34:32 +08:00
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; (fadd x, (fmul y, z)) -> (fma y, z, x)
|
|
|
|
|
; FUNC-LABEL: {{^}}combine_to_fma_f64_1:
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[A:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[B:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[C:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:16{{$}}
|
|
|
|
|
; SI: v_fma_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[A]], [[B]], [[C]]
|
|
|
|
|
; SI: buffer_store_dwordx2 [[RESULT]]
|
|
|
|
|
define void @combine_to_fma_f64_1(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) #1 {
|
2016-02-11 14:02:01 +08:00
|
|
|
%tid = tail call i32 @llvm.amdgcn.workitem.id.x() #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
llvm-svn: 230786
2015-02-28 03:29:02 +08:00
|
|
|
%gep.0 = getelementptr double, double addrspace(1)* %in, i32 %tid
|
|
|
|
|
%gep.1 = getelementptr double, double addrspace(1)* %gep.0, i32 1
|
|
|
|
|
%gep.2 = getelementptr double, double addrspace(1)* %gep.0, i32 2
|
|
|
|
|
%gep.out = getelementptr double, double addrspace(1)* %out, i32 %tid
|
2015-01-30 03:34:32 +08:00
|
|
|
|
2016-04-12 21:38:18 +08:00
|
|
|
%a = load volatile double, double addrspace(1)* %gep.0
|
|
|
|
|
%b = load volatile double, double addrspace(1)* %gep.1
|
|
|
|
|
%c = load volatile double, double addrspace(1)* %gep.2
|
2015-01-30 03:34:32 +08:00
|
|
|
|
|
|
|
|
%mul = fmul double %a, %b
|
|
|
|
|
%fma = fadd double %c, %mul
|
|
|
|
|
store double %fma, double addrspace(1)* %gep.out
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
|
|
|
|
|
; FUNC-LABEL: {{^}}combine_to_fma_fsub_0_f64:
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[A:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[B:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[C:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:16{{$}}
|
|
|
|
|
; SI: v_fma_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[A]], [[B]], -[[C]]
|
|
|
|
|
; SI: buffer_store_dwordx2 [[RESULT]]
|
|
|
|
|
define void @combine_to_fma_fsub_0_f64(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) #1 {
|
2016-02-11 14:02:01 +08:00
|
|
|
%tid = tail call i32 @llvm.amdgcn.workitem.id.x() #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
llvm-svn: 230786
2015-02-28 03:29:02 +08:00
|
|
|
%gep.0 = getelementptr double, double addrspace(1)* %in, i32 %tid
|
|
|
|
|
%gep.1 = getelementptr double, double addrspace(1)* %gep.0, i32 1
|
|
|
|
|
%gep.2 = getelementptr double, double addrspace(1)* %gep.0, i32 2
|
|
|
|
|
%gep.out = getelementptr double, double addrspace(1)* %out, i32 %tid
|
2015-01-30 03:34:32 +08:00
|
|
|
|
2016-04-12 21:38:18 +08:00
|
|
|
%a = load volatile double, double addrspace(1)* %gep.0
|
|
|
|
|
%b = load volatile double, double addrspace(1)* %gep.1
|
|
|
|
|
%c = load volatile double, double addrspace(1)* %gep.2
|
2015-01-30 03:34:32 +08:00
|
|
|
|
|
|
|
|
%mul = fmul double %a, %b
|
|
|
|
|
%fma = fsub double %mul, %c
|
|
|
|
|
store double %fma, double addrspace(1)* %gep.out
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
|
|
|
|
|
; FUNC-LABEL: {{^}}combine_to_fma_fsub_f64_0_2use:
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[A:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[B:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[C:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:16{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[D:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:24{{$}}
|
|
|
|
|
; SI-DAG: v_fma_f64 [[RESULT0:v\[[0-9]+:[0-9]+\]]], [[A]], [[B]], -[[C]]
|
|
|
|
|
; SI-DAG: v_fma_f64 [[RESULT1:v\[[0-9]+:[0-9]+\]]], [[A]], [[B]], -[[D]]
|
|
|
|
|
; SI-DAG: buffer_store_dwordx2 [[RESULT0]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
|
|
|
|
|
; SI-DAG: buffer_store_dwordx2 [[RESULT1]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
|
|
|
|
|
; SI: s_endpgm
|
|
|
|
|
define void @combine_to_fma_fsub_f64_0_2use(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) #1 {
|
2016-02-11 14:02:01 +08:00
|
|
|
%tid = tail call i32 @llvm.amdgcn.workitem.id.x() #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
llvm-svn: 230786
2015-02-28 03:29:02 +08:00
|
|
|
%gep.0 = getelementptr double, double addrspace(1)* %in, i32 %tid
|
|
|
|
|
%gep.1 = getelementptr double, double addrspace(1)* %gep.0, i32 1
|
|
|
|
|
%gep.2 = getelementptr double, double addrspace(1)* %gep.0, i32 2
|
|
|
|
|
%gep.3 = getelementptr double, double addrspace(1)* %gep.0, i32 3
|
|
|
|
|
%gep.out.0 = getelementptr double, double addrspace(1)* %out, i32 %tid
|
|
|
|
|
%gep.out.1 = getelementptr double, double addrspace(1)* %gep.out.0, i32 1
|
2015-01-30 03:34:32 +08:00
|
|
|
|
2016-04-12 21:38:18 +08:00
|
|
|
%a = load volatile double, double addrspace(1)* %gep.0
|
|
|
|
|
%b = load volatile double, double addrspace(1)* %gep.1
|
|
|
|
|
%c = load volatile double, double addrspace(1)* %gep.2
|
|
|
|
|
%d = load volatile double, double addrspace(1)* %gep.3
|
2015-01-30 03:34:32 +08:00
|
|
|
|
|
|
|
|
%mul = fmul double %a, %b
|
|
|
|
|
%fma0 = fsub double %mul, %c
|
|
|
|
|
%fma1 = fsub double %mul, %d
|
2016-04-12 21:38:18 +08:00
|
|
|
store volatile double %fma0, double addrspace(1)* %gep.out.0
|
|
|
|
|
store volatile double %fma1, double addrspace(1)* %gep.out.1
|
2015-01-30 03:34:32 +08:00
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
|
|
|
|
|
; FUNC-LABEL: {{^}}combine_to_fma_fsub_1_f64:
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[A:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[B:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[C:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:16{{$}}
|
|
|
|
|
; SI: v_fma_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], -[[A]], [[B]], [[C]]
|
|
|
|
|
; SI: buffer_store_dwordx2 [[RESULT]]
|
|
|
|
|
define void @combine_to_fma_fsub_1_f64(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) #1 {
|
2016-02-11 14:02:01 +08:00
|
|
|
%tid = tail call i32 @llvm.amdgcn.workitem.id.x() #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
llvm-svn: 230786
2015-02-28 03:29:02 +08:00
|
|
|
%gep.0 = getelementptr double, double addrspace(1)* %in, i32 %tid
|
|
|
|
|
%gep.1 = getelementptr double, double addrspace(1)* %gep.0, i32 1
|
|
|
|
|
%gep.2 = getelementptr double, double addrspace(1)* %gep.0, i32 2
|
|
|
|
|
%gep.out = getelementptr double, double addrspace(1)* %out, i32 %tid
|
2015-01-30 03:34:32 +08:00
|
|
|
|
2016-04-12 21:38:18 +08:00
|
|
|
%a = load volatile double, double addrspace(1)* %gep.0
|
|
|
|
|
%b = load volatile double, double addrspace(1)* %gep.1
|
|
|
|
|
%c = load volatile double, double addrspace(1)* %gep.2
|
2015-01-30 03:34:32 +08:00
|
|
|
|
|
|
|
|
%mul = fmul double %a, %b
|
|
|
|
|
%fma = fsub double %c, %mul
|
|
|
|
|
store double %fma, double addrspace(1)* %gep.out
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
|
|
|
|
|
; FUNC-LABEL: {{^}}combine_to_fma_fsub_1_f64_2use:
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[A:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[B:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[C:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:16{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[D:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:24{{$}}
|
|
|
|
|
; SI-DAG: v_fma_f64 [[RESULT0:v\[[0-9]+:[0-9]+\]]], -[[A]], [[B]], [[C]]
|
|
|
|
|
; SI-DAG: v_fma_f64 [[RESULT1:v\[[0-9]+:[0-9]+\]]], -[[A]], [[B]], [[D]]
|
|
|
|
|
; SI-DAG: buffer_store_dwordx2 [[RESULT0]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
|
|
|
|
|
; SI-DAG: buffer_store_dwordx2 [[RESULT1]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
|
|
|
|
|
; SI: s_endpgm
|
|
|
|
|
define void @combine_to_fma_fsub_1_f64_2use(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) #1 {
|
2016-02-11 14:02:01 +08:00
|
|
|
%tid = tail call i32 @llvm.amdgcn.workitem.id.x() #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
llvm-svn: 230786
2015-02-28 03:29:02 +08:00
|
|
|
%gep.0 = getelementptr double, double addrspace(1)* %in, i32 %tid
|
|
|
|
|
%gep.1 = getelementptr double, double addrspace(1)* %gep.0, i32 1
|
|
|
|
|
%gep.2 = getelementptr double, double addrspace(1)* %gep.0, i32 2
|
|
|
|
|
%gep.3 = getelementptr double, double addrspace(1)* %gep.0, i32 3
|
|
|
|
|
%gep.out.0 = getelementptr double, double addrspace(1)* %out, i32 %tid
|
|
|
|
|
%gep.out.1 = getelementptr double, double addrspace(1)* %gep.out.0, i32 1
|
2015-01-30 03:34:32 +08:00
|
|
|
|
2016-04-12 21:38:18 +08:00
|
|
|
%a = load volatile double, double addrspace(1)* %gep.0
|
|
|
|
|
%b = load volatile double, double addrspace(1)* %gep.1
|
|
|
|
|
%c = load volatile double, double addrspace(1)* %gep.2
|
|
|
|
|
%d = load volatile double, double addrspace(1)* %gep.3
|
2015-01-30 03:34:32 +08:00
|
|
|
|
|
|
|
|
%mul = fmul double %a, %b
|
|
|
|
|
%fma0 = fsub double %c, %mul
|
|
|
|
|
%fma1 = fsub double %d, %mul
|
2016-04-12 21:38:18 +08:00
|
|
|
store volatile double %fma0, double addrspace(1)* %gep.out.0
|
|
|
|
|
store volatile double %fma1, double addrspace(1)* %gep.out.1
|
2015-01-30 03:34:32 +08:00
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; (fsub (fneg (fmul x, y)), z) -> (fma (fneg x), y, (fneg z))
|
|
|
|
|
; FUNC-LABEL: {{^}}combine_to_fma_fsub_2_f64:
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[A:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[B:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[C:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:16{{$}}
|
|
|
|
|
; SI: v_fma_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], -[[A]], [[B]], -[[C]]
|
|
|
|
|
; SI: buffer_store_dwordx2 [[RESULT]]
|
|
|
|
|
define void @combine_to_fma_fsub_2_f64(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) #1 {
|
2016-02-11 14:02:01 +08:00
|
|
|
%tid = tail call i32 @llvm.amdgcn.workitem.id.x() #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
llvm-svn: 230786
2015-02-28 03:29:02 +08:00
|
|
|
%gep.0 = getelementptr double, double addrspace(1)* %in, i32 %tid
|
|
|
|
|
%gep.1 = getelementptr double, double addrspace(1)* %gep.0, i32 1
|
|
|
|
|
%gep.2 = getelementptr double, double addrspace(1)* %gep.0, i32 2
|
|
|
|
|
%gep.out = getelementptr double, double addrspace(1)* %out, i32 %tid
|
2015-01-30 03:34:32 +08:00
|
|
|
|
2016-04-12 21:38:18 +08:00
|
|
|
%a = load volatile double, double addrspace(1)* %gep.0
|
|
|
|
|
%b = load volatile double, double addrspace(1)* %gep.1
|
|
|
|
|
%c = load volatile double, double addrspace(1)* %gep.2
|
2015-01-30 03:34:32 +08:00
|
|
|
|
|
|
|
|
%mul = fmul double %a, %b
|
|
|
|
|
%mul.neg = fsub double -0.0, %mul
|
|
|
|
|
%fma = fsub double %mul.neg, %c
|
|
|
|
|
|
|
|
|
|
store double %fma, double addrspace(1)* %gep.out
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; (fsub (fneg (fmul x, y)), z) -> (fma (fneg x), y, (fneg z))
|
|
|
|
|
; FUNC-LABEL: {{^}}combine_to_fma_fsub_2_f64_2uses_neg:
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[A:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[B:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[C:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:16{{$}}
|
|
|
|
|
; SI-DAG: v_fma_f64 [[RESULT0:v\[[0-9]+:[0-9]+\]]], -[[A]], [[B]], -[[C]]
|
|
|
|
|
; SI-DAG: v_fma_f64 [[RESULT1:v\[[0-9]+:[0-9]+\]]], -[[A]], [[B]], -[[D]]
|
|
|
|
|
; SI-DAG: buffer_store_dwordx2 [[RESULT0]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
|
|
|
|
|
; SI-DAG: buffer_store_dwordx2 [[RESULT1]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
|
|
|
|
|
; SI: s_endpgm
|
|
|
|
|
define void @combine_to_fma_fsub_2_f64_2uses_neg(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) #1 {
|
2016-02-11 14:02:01 +08:00
|
|
|
%tid = tail call i32 @llvm.amdgcn.workitem.id.x() #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
llvm-svn: 230786
2015-02-28 03:29:02 +08:00
|
|
|
%gep.0 = getelementptr double, double addrspace(1)* %in, i32 %tid
|
|
|
|
|
%gep.1 = getelementptr double, double addrspace(1)* %gep.0, i32 1
|
|
|
|
|
%gep.2 = getelementptr double, double addrspace(1)* %gep.0, i32 2
|
|
|
|
|
%gep.3 = getelementptr double, double addrspace(1)* %gep.0, i32 3
|
|
|
|
|
%gep.out.0 = getelementptr double, double addrspace(1)* %out, i32 %tid
|
|
|
|
|
%gep.out.1 = getelementptr double, double addrspace(1)* %gep.out.0, i32 1
|
2015-01-30 03:34:32 +08:00
|
|
|
|
2016-04-12 21:38:18 +08:00
|
|
|
%a = load volatile double, double addrspace(1)* %gep.0
|
|
|
|
|
%b = load volatile double, double addrspace(1)* %gep.1
|
|
|
|
|
%c = load volatile double, double addrspace(1)* %gep.2
|
|
|
|
|
%d = load volatile double, double addrspace(1)* %gep.3
|
2015-01-30 03:34:32 +08:00
|
|
|
|
|
|
|
|
%mul = fmul double %a, %b
|
|
|
|
|
%mul.neg = fsub double -0.0, %mul
|
|
|
|
|
%fma0 = fsub double %mul.neg, %c
|
|
|
|
|
%fma1 = fsub double %mul.neg, %d
|
|
|
|
|
|
2016-04-12 21:38:18 +08:00
|
|
|
store volatile double %fma0, double addrspace(1)* %gep.out.0
|
|
|
|
|
store volatile double %fma1, double addrspace(1)* %gep.out.1
|
2015-01-30 03:34:32 +08:00
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; (fsub (fneg (fmul x, y)), z) -> (fma (fneg x), y, (fneg z))
|
|
|
|
|
; FUNC-LABEL: {{^}}combine_to_fma_fsub_2_f64_2uses_mul:
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[A:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[B:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[C:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:16{{$}}
|
|
|
|
|
; SI-DAG: v_fma_f64 [[RESULT0:v\[[0-9]+:[0-9]+\]]], -[[A]], [[B]], -[[C]]
|
|
|
|
|
; SI-DAG: v_fma_f64 [[RESULT1:v\[[0-9]+:[0-9]+\]]], [[A]], [[B]], -[[D]]
|
|
|
|
|
; SI-DAG: buffer_store_dwordx2 [[RESULT0]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
|
|
|
|
|
; SI-DAG: buffer_store_dwordx2 [[RESULT1]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
|
|
|
|
|
; SI: s_endpgm
|
|
|
|
|
define void @combine_to_fma_fsub_2_f64_2uses_mul(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) #1 {
|
2016-02-11 14:02:01 +08:00
|
|
|
%tid = tail call i32 @llvm.amdgcn.workitem.id.x() #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
llvm-svn: 230786
2015-02-28 03:29:02 +08:00
|
|
|
%gep.0 = getelementptr double, double addrspace(1)* %in, i32 %tid
|
|
|
|
|
%gep.1 = getelementptr double, double addrspace(1)* %gep.0, i32 1
|
|
|
|
|
%gep.2 = getelementptr double, double addrspace(1)* %gep.0, i32 2
|
|
|
|
|
%gep.3 = getelementptr double, double addrspace(1)* %gep.0, i32 3
|
|
|
|
|
%gep.out.0 = getelementptr double, double addrspace(1)* %out, i32 %tid
|
|
|
|
|
%gep.out.1 = getelementptr double, double addrspace(1)* %gep.out.0, i32 1
|
2015-01-30 03:34:32 +08:00
|
|
|
|
2016-04-12 21:38:18 +08:00
|
|
|
%a = load volatile double, double addrspace(1)* %gep.0
|
|
|
|
|
%b = load volatile double, double addrspace(1)* %gep.1
|
|
|
|
|
%c = load volatile double, double addrspace(1)* %gep.2
|
|
|
|
|
%d = load volatile double, double addrspace(1)* %gep.3
|
2015-01-30 03:34:32 +08:00
|
|
|
|
|
|
|
|
%mul = fmul double %a, %b
|
|
|
|
|
%mul.neg = fsub double -0.0, %mul
|
|
|
|
|
%fma0 = fsub double %mul.neg, %c
|
|
|
|
|
%fma1 = fsub double %mul, %d
|
|
|
|
|
|
2016-04-12 21:38:18 +08:00
|
|
|
store volatile double %fma0, double addrspace(1)* %gep.out.0
|
|
|
|
|
store volatile double %fma1, double addrspace(1)* %gep.out.1
|
2015-01-30 03:34:32 +08:00
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; fold (fsub (fma x, y, (fmul u, v)), z) -> (fma x, y (fma u, v, (fneg z)))
|
|
|
|
|
|
|
|
|
|
; FUNC-LABEL: {{^}}aggressive_combine_to_fma_fsub_0_f64:
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[X:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[Y:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[Z:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:16{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[U:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:24{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[V:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:32{{$}}
|
|
|
|
|
; SI: v_fma_f64 [[FMA0:v\[[0-9]+:[0-9]+\]]], [[U]], [[V]], -[[Z]]
|
|
|
|
|
; SI: v_fma_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[X]], [[Y]], [[FMA0]]
|
|
|
|
|
; SI: buffer_store_dwordx2 [[RESULT]]
|
|
|
|
|
define void @aggressive_combine_to_fma_fsub_0_f64(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) #1 {
|
2016-02-11 14:02:01 +08:00
|
|
|
%tid = tail call i32 @llvm.amdgcn.workitem.id.x() #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
llvm-svn: 230786
2015-02-28 03:29:02 +08:00
|
|
|
%gep.0 = getelementptr double, double addrspace(1)* %in, i32 %tid
|
|
|
|
|
%gep.1 = getelementptr double, double addrspace(1)* %gep.0, i32 1
|
|
|
|
|
%gep.2 = getelementptr double, double addrspace(1)* %gep.0, i32 2
|
|
|
|
|
%gep.3 = getelementptr double, double addrspace(1)* %gep.0, i32 3
|
|
|
|
|
%gep.4 = getelementptr double, double addrspace(1)* %gep.0, i32 4
|
|
|
|
|
%gep.out = getelementptr double, double addrspace(1)* %out, i32 %tid
|
2015-01-30 03:34:32 +08:00
|
|
|
|
2016-04-12 21:38:18 +08:00
|
|
|
%x = load volatile double, double addrspace(1)* %gep.0
|
|
|
|
|
%y = load volatile double, double addrspace(1)* %gep.1
|
|
|
|
|
%z = load volatile double, double addrspace(1)* %gep.2
|
|
|
|
|
%u = load volatile double, double addrspace(1)* %gep.3
|
|
|
|
|
%v = load volatile double, double addrspace(1)* %gep.4
|
2015-01-30 03:34:32 +08:00
|
|
|
|
|
|
|
|
%tmp0 = fmul double %u, %v
|
|
|
|
|
%tmp1 = call double @llvm.fma.f64(double %x, double %y, double %tmp0) #0
|
|
|
|
|
%tmp2 = fsub double %tmp1, %z
|
|
|
|
|
|
|
|
|
|
store double %tmp2, double addrspace(1)* %gep.out
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; fold (fsub x, (fma y, z, (fmul u, v)))
|
|
|
|
|
; -> (fma (fneg y), z, (fma (fneg u), v, x))
|
|
|
|
|
|
|
|
|
|
; FUNC-LABEL: {{^}}aggressive_combine_to_fma_fsub_1_f64:
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[X:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[Y:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[Z:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:16{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[U:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:24{{$}}
|
|
|
|
|
; SI-DAG: buffer_load_dwordx2 [[V:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:32{{$}}
|
|
|
|
|
; SI: v_fma_f64 [[FMA0:v\[[0-9]+:[0-9]+\]]], -[[U]], [[V]], [[X]]
|
|
|
|
|
; SI: v_fma_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], -[[Y]], [[Z]], [[FMA0]]
|
|
|
|
|
; SI: buffer_store_dwordx2 [[RESULT]]
|
|
|
|
|
define void @aggressive_combine_to_fma_fsub_1_f64(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) #1 {
|
2016-02-11 14:02:01 +08:00
|
|
|
%tid = tail call i32 @llvm.amdgcn.workitem.id.x() #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
llvm-svn: 230786
2015-02-28 03:29:02 +08:00
|
|
|
%gep.0 = getelementptr double, double addrspace(1)* %in, i32 %tid
|
|
|
|
|
%gep.1 = getelementptr double, double addrspace(1)* %gep.0, i32 1
|
|
|
|
|
%gep.2 = getelementptr double, double addrspace(1)* %gep.0, i32 2
|
|
|
|
|
%gep.3 = getelementptr double, double addrspace(1)* %gep.0, i32 3
|
|
|
|
|
%gep.4 = getelementptr double, double addrspace(1)* %gep.0, i32 4
|
|
|
|
|
%gep.out = getelementptr double, double addrspace(1)* %out, i32 %tid
|
2015-01-30 03:34:32 +08:00
|
|
|
|
2016-04-12 21:38:18 +08:00
|
|
|
%x = load volatile double, double addrspace(1)* %gep.0
|
|
|
|
|
%y = load volatile double, double addrspace(1)* %gep.1
|
|
|
|
|
%z = load volatile double, double addrspace(1)* %gep.2
|
|
|
|
|
%u = load volatile double, double addrspace(1)* %gep.3
|
|
|
|
|
%v = load volatile double, double addrspace(1)* %gep.4
|
2015-01-30 03:34:32 +08:00
|
|
|
|
|
|
|
|
%tmp0 = fmul double %u, %v
|
|
|
|
|
%tmp1 = call double @llvm.fma.f64(double %y, double %z, double %tmp0) #0
|
|
|
|
|
%tmp2 = fsub double %x, %tmp1
|
|
|
|
|
|
|
|
|
|
store double %tmp2, double addrspace(1)* %gep.out
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
2015-09-22 04:32:48 +08:00
|
|
|
;
|
|
|
|
|
; Patterns (+ fneg variants): mul(add(1.0,x),y), mul(sub(1.0,x),y), mul(sub(x,1.0),y)
|
|
|
|
|
;
|
|
|
|
|
|
|
|
|
|
; FUNC-LABEL: {{^}}test_f32_mul_add_x_one_y:
|
[DAGCombiner] do not fold (fmul (fadd X, 1), Y) -> (fmad X, Y, Y) by default
Summary:
When X = 0 and Y = inf, the original code produces inf, but the transformed
code produces nan. So this transform (and its relatives) should only be
used when the no-infs-fp-math flag is explicitly enabled.
Also disable the transform using fmad (intermediate rounding) when unsafe-math
is not enabled, since it can reduce the precision of the result; consider this
example with binary floating point numbers with two bits of mantissa:
x = 1.01
y = 111
x * (y + 1) = 1.01 * 1000 = 1010 (this is the exact result; no rounding occurs at any step)
x * y + x = 1000.11 + 1.01 =r 1000 + 1.01 = 1001.01 =r 1000 (with rounding towards zero)
The example relies on rounding towards zero at least in the second step.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98578
Reviewers: RKSimon, tstellarAMD, spatel, arsenm
Subscribers: wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D26602
llvm-svn: 288506
2016-12-03 00:06:18 +08:00
|
|
|
; SI-NOFMA: v_add_f32_e32 [[VS:v[0-9]]], 1.0, [[VX:v[0-9]]]
|
|
|
|
|
; SI-NOFMA: v_mul_f32_e32 {{v[0-9]}}, [[VY:v[0-9]]], [[VS]]
|
|
|
|
|
;
|
|
|
|
|
; SI-FMA: v_fma_f32 {{v[0-9]}}, [[VX:v[0-9]]], [[VY:v[0-9]]], [[VY:v[0-9]]]
|
2015-09-22 04:32:48 +08:00
|
|
|
define void @test_f32_mul_add_x_one_y(float addrspace(1)* %out,
|
|
|
|
|
float addrspace(1)* %in1,
|
|
|
|
|
float addrspace(1)* %in2) {
|
2016-04-12 21:38:18 +08:00
|
|
|
%x = load volatile float, float addrspace(1)* %in1
|
|
|
|
|
%y = load volatile float, float addrspace(1)* %in2
|
2015-09-22 04:32:48 +08:00
|
|
|
%a = fadd float %x, 1.0
|
|
|
|
|
%m = fmul float %a, %y
|
|
|
|
|
store float %m, float addrspace(1)* %out
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; FUNC-LABEL: {{^}}test_f32_mul_y_add_x_one:
|
[DAGCombiner] do not fold (fmul (fadd X, 1), Y) -> (fmad X, Y, Y) by default
Summary:
When X = 0 and Y = inf, the original code produces inf, but the transformed
code produces nan. So this transform (and its relatives) should only be
used when the no-infs-fp-math flag is explicitly enabled.
Also disable the transform using fmad (intermediate rounding) when unsafe-math
is not enabled, since it can reduce the precision of the result; consider this
example with binary floating point numbers with two bits of mantissa:
x = 1.01
y = 111
x * (y + 1) = 1.01 * 1000 = 1010 (this is the exact result; no rounding occurs at any step)
x * y + x = 1000.11 + 1.01 =r 1000 + 1.01 = 1001.01 =r 1000 (with rounding towards zero)
The example relies on rounding towards zero at least in the second step.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98578
Reviewers: RKSimon, tstellarAMD, spatel, arsenm
Subscribers: wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D26602
llvm-svn: 288506
2016-12-03 00:06:18 +08:00
|
|
|
; SI-NOFMA: v_add_f32_e32 [[VS:v[0-9]]], 1.0, [[VX:v[0-9]]]
|
|
|
|
|
; SI-NOFMA: v_mul_f32_e32 {{v[0-9]}}, [[VS]], [[VY:v[0-9]]]
|
|
|
|
|
;
|
|
|
|
|
; SI-FMA: v_fma_f32 {{v[0-9]}}, [[VX:v[0-9]]], [[VY:v[0-9]]], [[VY:v[0-9]]]
|
2015-09-22 04:32:48 +08:00
|
|
|
define void @test_f32_mul_y_add_x_one(float addrspace(1)* %out,
|
|
|
|
|
float addrspace(1)* %in1,
|
|
|
|
|
float addrspace(1)* %in2) {
|
2016-04-12 21:38:18 +08:00
|
|
|
%x = load volatile float, float addrspace(1)* %in1
|
|
|
|
|
%y = load volatile float, float addrspace(1)* %in2
|
2015-09-22 04:32:48 +08:00
|
|
|
%a = fadd float %x, 1.0
|
|
|
|
|
%m = fmul float %y, %a
|
|
|
|
|
store float %m, float addrspace(1)* %out
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; FUNC-LABEL: {{^}}test_f32_mul_add_x_negone_y:
|
[DAGCombiner] do not fold (fmul (fadd X, 1), Y) -> (fmad X, Y, Y) by default
Summary:
When X = 0 and Y = inf, the original code produces inf, but the transformed
code produces nan. So this transform (and its relatives) should only be
used when the no-infs-fp-math flag is explicitly enabled.
Also disable the transform using fmad (intermediate rounding) when unsafe-math
is not enabled, since it can reduce the precision of the result; consider this
example with binary floating point numbers with two bits of mantissa:
x = 1.01
y = 111
x * (y + 1) = 1.01 * 1000 = 1010 (this is the exact result; no rounding occurs at any step)
x * y + x = 1000.11 + 1.01 =r 1000 + 1.01 = 1001.01 =r 1000 (with rounding towards zero)
The example relies on rounding towards zero at least in the second step.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98578
Reviewers: RKSimon, tstellarAMD, spatel, arsenm
Subscribers: wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D26602
llvm-svn: 288506
2016-12-03 00:06:18 +08:00
|
|
|
; SI-NOFMA: v_add_f32_e32 [[VS:v[0-9]]], -1.0, [[VX:v[0-9]]]
|
|
|
|
|
; SI-NOFMA: v_mul_f32_e32 {{v[0-9]}}, [[VY:v[0-9]]], [[VS]]
|
|
|
|
|
;
|
|
|
|
|
; SI-FMA: v_fma_f32 {{v[0-9]}}, [[VX:v[0-9]]], [[VY:v[0-9]]], -[[VY:v[0-9]]]
|
2015-09-22 04:32:48 +08:00
|
|
|
define void @test_f32_mul_add_x_negone_y(float addrspace(1)* %out,
|
|
|
|
|
float addrspace(1)* %in1,
|
|
|
|
|
float addrspace(1)* %in2) {
|
|
|
|
|
%x = load float, float addrspace(1)* %in1
|
|
|
|
|
%y = load float, float addrspace(1)* %in2
|
|
|
|
|
%a = fadd float %x, -1.0
|
|
|
|
|
%m = fmul float %a, %y
|
|
|
|
|
store float %m, float addrspace(1)* %out
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; FUNC-LABEL: {{^}}test_f32_mul_y_add_x_negone:
|
[DAGCombiner] do not fold (fmul (fadd X, 1), Y) -> (fmad X, Y, Y) by default
Summary:
When X = 0 and Y = inf, the original code produces inf, but the transformed
code produces nan. So this transform (and its relatives) should only be
used when the no-infs-fp-math flag is explicitly enabled.
Also disable the transform using fmad (intermediate rounding) when unsafe-math
is not enabled, since it can reduce the precision of the result; consider this
example with binary floating point numbers with two bits of mantissa:
x = 1.01
y = 111
x * (y + 1) = 1.01 * 1000 = 1010 (this is the exact result; no rounding occurs at any step)
x * y + x = 1000.11 + 1.01 =r 1000 + 1.01 = 1001.01 =r 1000 (with rounding towards zero)
The example relies on rounding towards zero at least in the second step.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98578
Reviewers: RKSimon, tstellarAMD, spatel, arsenm
Subscribers: wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D26602
llvm-svn: 288506
2016-12-03 00:06:18 +08:00
|
|
|
; SI-NOFMA: v_add_f32_e32 [[VS:v[0-9]]], -1.0, [[VX:v[0-9]]]
|
|
|
|
|
; SI-NOFMA: v_mul_f32_e32 {{v[0-9]}}, [[VS]], [[VY:v[0-9]]]
|
|
|
|
|
;
|
|
|
|
|
; SI-FMA: v_fma_f32 {{v[0-9]}}, [[VX:v[0-9]]], [[VY:v[0-9]]], -[[VY:v[0-9]]]
|
2015-09-22 04:32:48 +08:00
|
|
|
define void @test_f32_mul_y_add_x_negone(float addrspace(1)* %out,
|
|
|
|
|
float addrspace(1)* %in1,
|
|
|
|
|
float addrspace(1)* %in2) {
|
|
|
|
|
%x = load float, float addrspace(1)* %in1
|
|
|
|
|
%y = load float, float addrspace(1)* %in2
|
|
|
|
|
%a = fadd float %x, -1.0
|
|
|
|
|
%m = fmul float %y, %a
|
|
|
|
|
store float %m, float addrspace(1)* %out
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; FUNC-LABEL: {{^}}test_f32_mul_sub_one_x_y:
|
[DAGCombiner] do not fold (fmul (fadd X, 1), Y) -> (fmad X, Y, Y) by default
Summary:
When X = 0 and Y = inf, the original code produces inf, but the transformed
code produces nan. So this transform (and its relatives) should only be
used when the no-infs-fp-math flag is explicitly enabled.
Also disable the transform using fmad (intermediate rounding) when unsafe-math
is not enabled, since it can reduce the precision of the result; consider this
example with binary floating point numbers with two bits of mantissa:
x = 1.01
y = 111
x * (y + 1) = 1.01 * 1000 = 1010 (this is the exact result; no rounding occurs at any step)
x * y + x = 1000.11 + 1.01 =r 1000 + 1.01 = 1001.01 =r 1000 (with rounding towards zero)
The example relies on rounding towards zero at least in the second step.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98578
Reviewers: RKSimon, tstellarAMD, spatel, arsenm
Subscribers: wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D26602
llvm-svn: 288506
2016-12-03 00:06:18 +08:00
|
|
|
; SI-NOFMA: v_sub_f32_e32 [[VS:v[0-9]]], 1.0, [[VX:v[0-9]]]
|
|
|
|
|
; SI-NOFMA: v_mul_f32_e32 {{v[0-9]}}, [[VY:v[0-9]]], [[VS]]
|
|
|
|
|
;
|
|
|
|
|
; SI-FMA: v_fma_f32 {{v[0-9]}}, -[[VX:v[0-9]]], [[VY:v[0-9]]], [[VY:v[0-9]]]
|
2015-09-22 04:32:48 +08:00
|
|
|
define void @test_f32_mul_sub_one_x_y(float addrspace(1)* %out,
|
|
|
|
|
float addrspace(1)* %in1,
|
|
|
|
|
float addrspace(1)* %in2) {
|
|
|
|
|
%x = load float, float addrspace(1)* %in1
|
|
|
|
|
%y = load float, float addrspace(1)* %in2
|
|
|
|
|
%s = fsub float 1.0, %x
|
|
|
|
|
%m = fmul float %s, %y
|
|
|
|
|
store float %m, float addrspace(1)* %out
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; FUNC-LABEL: {{^}}test_f32_mul_y_sub_one_x:
|
[DAGCombiner] do not fold (fmul (fadd X, 1), Y) -> (fmad X, Y, Y) by default
Summary:
When X = 0 and Y = inf, the original code produces inf, but the transformed
code produces nan. So this transform (and its relatives) should only be
used when the no-infs-fp-math flag is explicitly enabled.
Also disable the transform using fmad (intermediate rounding) when unsafe-math
is not enabled, since it can reduce the precision of the result; consider this
example with binary floating point numbers with two bits of mantissa:
x = 1.01
y = 111
x * (y + 1) = 1.01 * 1000 = 1010 (this is the exact result; no rounding occurs at any step)
x * y + x = 1000.11 + 1.01 =r 1000 + 1.01 = 1001.01 =r 1000 (with rounding towards zero)
The example relies on rounding towards zero at least in the second step.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98578
Reviewers: RKSimon, tstellarAMD, spatel, arsenm
Subscribers: wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D26602
llvm-svn: 288506
2016-12-03 00:06:18 +08:00
|
|
|
; SI-NOFMA: v_sub_f32_e32 [[VS:v[0-9]]], 1.0, [[VX:v[0-9]]]
|
|
|
|
|
; SI-NOFMA: v_mul_f32_e32 {{v[0-9]}}, [[VS]], [[VY:v[0-9]]]
|
|
|
|
|
;
|
|
|
|
|
; SI-FMA: v_fma_f32 {{v[0-9]}}, -[[VX:v[0-9]]], [[VY:v[0-9]]], [[VY:v[0-9]]]
|
2015-09-22 04:32:48 +08:00
|
|
|
define void @test_f32_mul_y_sub_one_x(float addrspace(1)* %out,
|
|
|
|
|
float addrspace(1)* %in1,
|
|
|
|
|
float addrspace(1)* %in2) {
|
|
|
|
|
%x = load float, float addrspace(1)* %in1
|
|
|
|
|
%y = load float, float addrspace(1)* %in2
|
|
|
|
|
%s = fsub float 1.0, %x
|
|
|
|
|
%m = fmul float %y, %s
|
|
|
|
|
store float %m, float addrspace(1)* %out
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; FUNC-LABEL: {{^}}test_f32_mul_sub_negone_x_y:
|
[DAGCombiner] do not fold (fmul (fadd X, 1), Y) -> (fmad X, Y, Y) by default
Summary:
When X = 0 and Y = inf, the original code produces inf, but the transformed
code produces nan. So this transform (and its relatives) should only be
used when the no-infs-fp-math flag is explicitly enabled.
Also disable the transform using fmad (intermediate rounding) when unsafe-math
is not enabled, since it can reduce the precision of the result; consider this
example with binary floating point numbers with two bits of mantissa:
x = 1.01
y = 111
x * (y + 1) = 1.01 * 1000 = 1010 (this is the exact result; no rounding occurs at any step)
x * y + x = 1000.11 + 1.01 =r 1000 + 1.01 = 1001.01 =r 1000 (with rounding towards zero)
The example relies on rounding towards zero at least in the second step.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98578
Reviewers: RKSimon, tstellarAMD, spatel, arsenm
Subscribers: wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D26602
llvm-svn: 288506
2016-12-03 00:06:18 +08:00
|
|
|
; SI-NOFMA: v_sub_f32_e32 [[VS:v[0-9]]], -1.0, [[VX:v[0-9]]]
|
|
|
|
|
; SI-NOFMA: v_mul_f32_e32 {{v[0-9]}}, [[VY:v[0-9]]], [[VS]]
|
|
|
|
|
;
|
|
|
|
|
; SI-FMA: v_fma_f32 {{v[0-9]}}, -[[VX:v[0-9]]], [[VY:v[0-9]]], -[[VY:v[0-9]]]
|
2015-09-22 04:32:48 +08:00
|
|
|
define void @test_f32_mul_sub_negone_x_y(float addrspace(1)* %out,
|
|
|
|
|
float addrspace(1)* %in1,
|
|
|
|
|
float addrspace(1)* %in2) {
|
|
|
|
|
%x = load float, float addrspace(1)* %in1
|
|
|
|
|
%y = load float, float addrspace(1)* %in2
|
|
|
|
|
%s = fsub float -1.0, %x
|
|
|
|
|
%m = fmul float %s, %y
|
|
|
|
|
store float %m, float addrspace(1)* %out
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; FUNC-LABEL: {{^}}test_f32_mul_y_sub_negone_x:
|
[DAGCombiner] do not fold (fmul (fadd X, 1), Y) -> (fmad X, Y, Y) by default
Summary:
When X = 0 and Y = inf, the original code produces inf, but the transformed
code produces nan. So this transform (and its relatives) should only be
used when the no-infs-fp-math flag is explicitly enabled.
Also disable the transform using fmad (intermediate rounding) when unsafe-math
is not enabled, since it can reduce the precision of the result; consider this
example with binary floating point numbers with two bits of mantissa:
x = 1.01
y = 111
x * (y + 1) = 1.01 * 1000 = 1010 (this is the exact result; no rounding occurs at any step)
x * y + x = 1000.11 + 1.01 =r 1000 + 1.01 = 1001.01 =r 1000 (with rounding towards zero)
The example relies on rounding towards zero at least in the second step.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98578
Reviewers: RKSimon, tstellarAMD, spatel, arsenm
Subscribers: wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D26602
llvm-svn: 288506
2016-12-03 00:06:18 +08:00
|
|
|
; SI-NOFMA: v_sub_f32_e32 [[VS:v[0-9]]], -1.0, [[VX:v[0-9]]]
|
|
|
|
|
; SI-NOFMA: v_mul_f32_e32 {{v[0-9]}}, [[VS]], [[VY:v[0-9]]]
|
|
|
|
|
;
|
|
|
|
|
; SI-FMA: v_fma_f32 {{v[0-9]}}, -[[VX:v[0-9]]], [[VY:v[0-9]]], -[[VY:v[0-9]]]
|
2015-09-22 04:32:48 +08:00
|
|
|
define void @test_f32_mul_y_sub_negone_x(float addrspace(1)* %out,
|
|
|
|
|
float addrspace(1)* %in1,
|
|
|
|
|
float addrspace(1)* %in2) {
|
|
|
|
|
%x = load float, float addrspace(1)* %in1
|
|
|
|
|
%y = load float, float addrspace(1)* %in2
|
|
|
|
|
%s = fsub float -1.0, %x
|
|
|
|
|
%m = fmul float %y, %s
|
|
|
|
|
store float %m, float addrspace(1)* %out
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; FUNC-LABEL: {{^}}test_f32_mul_sub_x_one_y:
|
[DAGCombiner] do not fold (fmul (fadd X, 1), Y) -> (fmad X, Y, Y) by default
Summary:
When X = 0 and Y = inf, the original code produces inf, but the transformed
code produces nan. So this transform (and its relatives) should only be
used when the no-infs-fp-math flag is explicitly enabled.
Also disable the transform using fmad (intermediate rounding) when unsafe-math
is not enabled, since it can reduce the precision of the result; consider this
example with binary floating point numbers with two bits of mantissa:
x = 1.01
y = 111
x * (y + 1) = 1.01 * 1000 = 1010 (this is the exact result; no rounding occurs at any step)
x * y + x = 1000.11 + 1.01 =r 1000 + 1.01 = 1001.01 =r 1000 (with rounding towards zero)
The example relies on rounding towards zero at least in the second step.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98578
Reviewers: RKSimon, tstellarAMD, spatel, arsenm
Subscribers: wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D26602
llvm-svn: 288506
2016-12-03 00:06:18 +08:00
|
|
|
; SI-NOFMA: v_add_f32_e32 [[VS:v[0-9]]], -1.0, [[VX:v[0-9]]]
|
|
|
|
|
; SI-NOFMA: v_mul_f32_e32 {{v[0-9]}}, [[VY:v[0-9]]], [[VS]]
|
|
|
|
|
;
|
|
|
|
|
; SI-FMA: v_fma_f32 {{v[0-9]}}, [[VX:v[0-9]]], [[VY:v[0-9]]], -[[VY:v[0-9]]]
|
2015-09-22 04:32:48 +08:00
|
|
|
define void @test_f32_mul_sub_x_one_y(float addrspace(1)* %out,
|
|
|
|
|
float addrspace(1)* %in1,
|
|
|
|
|
float addrspace(1)* %in2) {
|
|
|
|
|
%x = load float, float addrspace(1)* %in1
|
|
|
|
|
%y = load float, float addrspace(1)* %in2
|
|
|
|
|
%s = fsub float %x, 1.0
|
|
|
|
|
%m = fmul float %s, %y
|
|
|
|
|
store float %m, float addrspace(1)* %out
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; FUNC-LABEL: {{^}}test_f32_mul_y_sub_x_one:
|
[DAGCombiner] do not fold (fmul (fadd X, 1), Y) -> (fmad X, Y, Y) by default
Summary:
When X = 0 and Y = inf, the original code produces inf, but the transformed
code produces nan. So this transform (and its relatives) should only be
used when the no-infs-fp-math flag is explicitly enabled.
Also disable the transform using fmad (intermediate rounding) when unsafe-math
is not enabled, since it can reduce the precision of the result; consider this
example with binary floating point numbers with two bits of mantissa:
x = 1.01
y = 111
x * (y + 1) = 1.01 * 1000 = 1010 (this is the exact result; no rounding occurs at any step)
x * y + x = 1000.11 + 1.01 =r 1000 + 1.01 = 1001.01 =r 1000 (with rounding towards zero)
The example relies on rounding towards zero at least in the second step.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98578
Reviewers: RKSimon, tstellarAMD, spatel, arsenm
Subscribers: wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D26602
llvm-svn: 288506
2016-12-03 00:06:18 +08:00
|
|
|
; SI-NOFMA: v_add_f32_e32 [[VS:v[0-9]]], -1.0, [[VX:v[0-9]]]
|
|
|
|
|
; SI-NOFMA: v_mul_f32_e32 {{v[0-9]}}, [[VS]], [[VY:v[0-9]]]
|
|
|
|
|
;
|
|
|
|
|
; SI-FMA: v_fma_f32 {{v[0-9]}}, [[VX:v[0-9]]], [[VY:v[0-9]]], -[[VY:v[0-9]]]
|
2015-09-22 04:32:48 +08:00
|
|
|
define void @test_f32_mul_y_sub_x_one(float addrspace(1)* %out,
|
|
|
|
|
float addrspace(1)* %in1,
|
|
|
|
|
float addrspace(1)* %in2) {
|
|
|
|
|
%x = load float, float addrspace(1)* %in1
|
|
|
|
|
%y = load float, float addrspace(1)* %in2
|
|
|
|
|
%s = fsub float %x, 1.0
|
|
|
|
|
%m = fmul float %y, %s
|
|
|
|
|
store float %m, float addrspace(1)* %out
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; FUNC-LABEL: {{^}}test_f32_mul_sub_x_negone_y:
|
[DAGCombiner] do not fold (fmul (fadd X, 1), Y) -> (fmad X, Y, Y) by default
Summary:
When X = 0 and Y = inf, the original code produces inf, but the transformed
code produces nan. So this transform (and its relatives) should only be
used when the no-infs-fp-math flag is explicitly enabled.
Also disable the transform using fmad (intermediate rounding) when unsafe-math
is not enabled, since it can reduce the precision of the result; consider this
example with binary floating point numbers with two bits of mantissa:
x = 1.01
y = 111
x * (y + 1) = 1.01 * 1000 = 1010 (this is the exact result; no rounding occurs at any step)
x * y + x = 1000.11 + 1.01 =r 1000 + 1.01 = 1001.01 =r 1000 (with rounding towards zero)
The example relies on rounding towards zero at least in the second step.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98578
Reviewers: RKSimon, tstellarAMD, spatel, arsenm
Subscribers: wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D26602
llvm-svn: 288506
2016-12-03 00:06:18 +08:00
|
|
|
; SI-NOFMA: v_add_f32_e32 [[VS:v[0-9]]], 1.0, [[VX:v[0-9]]]
|
|
|
|
|
; SI-NOFMA: v_mul_f32_e32 {{v[0-9]}}, [[VY:v[0-9]]], [[VS]]
|
|
|
|
|
;
|
|
|
|
|
; SI-FMA: v_fma_f32 {{v[0-9]}}, [[VX:v[0-9]]], [[VY:v[0-9]]], [[VY:v[0-9]]]
|
2015-09-22 04:32:48 +08:00
|
|
|
define void @test_f32_mul_sub_x_negone_y(float addrspace(1)* %out,
|
|
|
|
|
float addrspace(1)* %in1,
|
|
|
|
|
float addrspace(1)* %in2) {
|
|
|
|
|
%x = load float, float addrspace(1)* %in1
|
|
|
|
|
%y = load float, float addrspace(1)* %in2
|
|
|
|
|
%s = fsub float %x, -1.0
|
|
|
|
|
%m = fmul float %s, %y
|
|
|
|
|
store float %m, float addrspace(1)* %out
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; FUNC-LABEL: {{^}}test_f32_mul_y_sub_x_negone:
|
[DAGCombiner] do not fold (fmul (fadd X, 1), Y) -> (fmad X, Y, Y) by default
Summary:
When X = 0 and Y = inf, the original code produces inf, but the transformed
code produces nan. So this transform (and its relatives) should only be
used when the no-infs-fp-math flag is explicitly enabled.
Also disable the transform using fmad (intermediate rounding) when unsafe-math
is not enabled, since it can reduce the precision of the result; consider this
example with binary floating point numbers with two bits of mantissa:
x = 1.01
y = 111
x * (y + 1) = 1.01 * 1000 = 1010 (this is the exact result; no rounding occurs at any step)
x * y + x = 1000.11 + 1.01 =r 1000 + 1.01 = 1001.01 =r 1000 (with rounding towards zero)
The example relies on rounding towards zero at least in the second step.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98578
Reviewers: RKSimon, tstellarAMD, spatel, arsenm
Subscribers: wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D26602
llvm-svn: 288506
2016-12-03 00:06:18 +08:00
|
|
|
; SI-NOFMA: v_add_f32_e32 [[VS:v[0-9]]], 1.0, [[VX:v[0-9]]]
|
|
|
|
|
; SI-NOFMA: v_mul_f32_e32 {{v[0-9]}}, [[VS]], [[VY:v[0-9]]]
|
|
|
|
|
;
|
|
|
|
|
; SI-FMA: v_fma_f32 {{v[0-9]}}, [[VX:v[0-9]]], [[VY:v[0-9]]], [[VY:v[0-9]]]
|
2015-09-22 04:32:48 +08:00
|
|
|
define void @test_f32_mul_y_sub_x_negone(float addrspace(1)* %out,
|
|
|
|
|
float addrspace(1)* %in1,
|
|
|
|
|
float addrspace(1)* %in2) {
|
|
|
|
|
%x = load float, float addrspace(1)* %in1
|
|
|
|
|
%y = load float, float addrspace(1)* %in2
|
|
|
|
|
%s = fsub float %x, -1.0
|
|
|
|
|
%m = fmul float %y, %s
|
|
|
|
|
store float %m, float addrspace(1)* %out
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
;
|
|
|
|
|
; Interpolation Patterns: add(mul(x,t),mul(sub(1.0,t),y))
|
|
|
|
|
;
|
|
|
|
|
|
|
|
|
|
; FUNC-LABEL: {{^}}test_f32_interp:
|
[DAGCombiner] do not fold (fmul (fadd X, 1), Y) -> (fmad X, Y, Y) by default
Summary:
When X = 0 and Y = inf, the original code produces inf, but the transformed
code produces nan. So this transform (and its relatives) should only be
used when the no-infs-fp-math flag is explicitly enabled.
Also disable the transform using fmad (intermediate rounding) when unsafe-math
is not enabled, since it can reduce the precision of the result; consider this
example with binary floating point numbers with two bits of mantissa:
x = 1.01
y = 111
x * (y + 1) = 1.01 * 1000 = 1010 (this is the exact result; no rounding occurs at any step)
x * y + x = 1000.11 + 1.01 =r 1000 + 1.01 = 1001.01 =r 1000 (with rounding towards zero)
The example relies on rounding towards zero at least in the second step.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98578
Reviewers: RKSimon, tstellarAMD, spatel, arsenm
Subscribers: wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D26602
llvm-svn: 288506
2016-12-03 00:06:18 +08:00
|
|
|
; SI-NOFMA: v_sub_f32_e32 [[VT1:v[0-9]]], 1.0, [[VT:v[0-9]]]
|
|
|
|
|
; SI-NOFMA: v_mul_f32_e32 [[VTY:v[0-9]]], [[VT1]], [[VY:v[0-9]]]
|
|
|
|
|
; SI-NOFMA: v_mac_f32_e32 [[VTY]], [[VT]], [[VX:v[0-9]]]
|
|
|
|
|
;
|
|
|
|
|
; SI-FMA: v_fma_f32 [[VR:v[0-9]]], -[[VT:v[0-9]]], [[VY:v[0-9]]], [[VY]]
|
|
|
|
|
; SI-FMA: v_fma_f32 {{v[0-9]}}, [[VX:v[0-9]]], [[VT]], [[VR]]
|
2015-09-22 04:32:48 +08:00
|
|
|
define void @test_f32_interp(float addrspace(1)* %out,
|
|
|
|
|
float addrspace(1)* %in1,
|
|
|
|
|
float addrspace(1)* %in2,
|
|
|
|
|
float addrspace(1)* %in3) {
|
|
|
|
|
%x = load float, float addrspace(1)* %in1
|
|
|
|
|
%y = load float, float addrspace(1)* %in2
|
|
|
|
|
%t = load float, float addrspace(1)* %in3
|
|
|
|
|
%t1 = fsub float 1.0, %t
|
|
|
|
|
%tx = fmul float %x, %t
|
|
|
|
|
%ty = fmul float %y, %t1
|
|
|
|
|
%r = fadd float %tx, %ty
|
|
|
|
|
store float %r, float addrspace(1)* %out
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; FUNC-LABEL: {{^}}test_f64_interp:
|
[DAGCombiner] do not fold (fmul (fadd X, 1), Y) -> (fmad X, Y, Y) by default
Summary:
When X = 0 and Y = inf, the original code produces inf, but the transformed
code produces nan. So this transform (and its relatives) should only be
used when the no-infs-fp-math flag is explicitly enabled.
Also disable the transform using fmad (intermediate rounding) when unsafe-math
is not enabled, since it can reduce the precision of the result; consider this
example with binary floating point numbers with two bits of mantissa:
x = 1.01
y = 111
x * (y + 1) = 1.01 * 1000 = 1010 (this is the exact result; no rounding occurs at any step)
x * y + x = 1000.11 + 1.01 =r 1000 + 1.01 = 1001.01 =r 1000 (with rounding towards zero)
The example relies on rounding towards zero at least in the second step.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98578
Reviewers: RKSimon, tstellarAMD, spatel, arsenm
Subscribers: wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D26602
llvm-svn: 288506
2016-12-03 00:06:18 +08:00
|
|
|
; SI-NOFMA: v_add_f64 [[VT1:v\[[0-9]+:[0-9]+\]]], -[[VT:v\[[0-9]+:[0-9]+\]]], 1.0
|
|
|
|
|
; SI-NOFMA: v_mul_f64 [[VTY:v\[[0-9]+:[0-9]+\]]], [[VY:v\[[0-9]+:[0-9]+\]]], [[VT1]]
|
|
|
|
|
; SI-NOFMA: v_fma_f64 v{{\[[0-9]+:[0-9]+\]}}, [[VX:v\[[0-9]+:[0-9]+\]]], [[VT]], [[VTY]]
|
|
|
|
|
;
|
|
|
|
|
; SI-FMA: v_fma_f64 [[VR:v\[[0-9]+:[0-9]+\]]], -[[VT:v\[[0-9]+:[0-9]+\]]], [[VY:v\[[0-9]+:[0-9]+\]]], [[VY]]
|
|
|
|
|
; SI-FMA: v_fma_f64 v{{\[[0-9]+:[0-9]+\]}}, [[VX:v\[[0-9]+:[0-9]+\]]], [[VT]], [[VR]]
|
2015-09-22 04:32:48 +08:00
|
|
|
define void @test_f64_interp(double addrspace(1)* %out,
|
|
|
|
|
double addrspace(1)* %in1,
|
|
|
|
|
double addrspace(1)* %in2,
|
|
|
|
|
double addrspace(1)* %in3) {
|
|
|
|
|
%x = load double, double addrspace(1)* %in1
|
|
|
|
|
%y = load double, double addrspace(1)* %in2
|
|
|
|
|
%t = load double, double addrspace(1)* %in3
|
|
|
|
|
%t1 = fsub double 1.0, %t
|
|
|
|
|
%tx = fmul double %x, %t
|
|
|
|
|
%ty = fmul double %y, %t1
|
|
|
|
|
%r = fadd double %tx, %ty
|
|
|
|
|
store double %r, double addrspace(1)* %out
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
2015-01-30 03:34:32 +08:00
|
|
|
attributes #0 = { nounwind readnone }
|
|
|
|
|
attributes #1 = { nounwind }
|