maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/llvm/test/CodeGen/AMDGPU/shl_add_ptr.ll

389 lines
17 KiB
LLVM
Raw Normal View History

AMDGPU: Fix multi-use shl/add combine This was using a custom function that didn't handle the addressing modes properly for private. Use isLegalAddressingMode to avoid duplicating this. Additionally, skip the combine if there is only one use since the standard combine will handle it. llvm-svn: 318013
2017-11-13 13:11:54 +08:00
; RUN: llc -march=amdgcn -mcpu=bonaire -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN %s
; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN %s
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
; Test that doing a shift of a pointer with a constant add will be
; folded into the constant offset addressing mode even if the add has
; multiple uses. This is relevant to accessing 2 separate, adjacent
; LDS globals.
AMDGPU: Remove some old intrinsic uses from tests llvm-svn: 260493
2016-02-11 14:02:01 +08:00
declare i32 @llvm.amdgcn.workitem.id.x() #1
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
R600: Error on initializer for LDS. Also give a proper error for other address spaces. llvm-svn: 221917
2014-11-14 03:56:13 +08:00
@lds0 = addrspace(3) global [512 x float] undef, align 4
@lds1 = addrspace(3) global [512 x float] undef, align 4
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
; Make sure the (add tid, 2) << 2 gets folded into the ds's offset as (tid << 2) + 8
AMDGPU: Fix multi-use shl/add combine This was using a custom function that didn't handle the addressing modes properly for private. Use isLegalAddressingMode to avoid duplicating this. Additionally, skip the combine if there is only one use since the standard combine will handle it. llvm-svn: 318013
2017-11-13 13:11:54 +08:00
; GCN-LABEL: {{^}}load_shl_base_lds_0:
; GCN: v_lshlrev_b32_e32 [[PTR:v[0-9]+]], 2, {{v[0-9]+}}
; GCN: ds_read_b32 {{v[0-9]+}}, [[PTR]] offset:8
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @load_shl_base_lds_0(float addrspace(1)* %out, i32 addrspace(1)* %add_use) #0 {
AMDGPU: Remove some old intrinsic uses from tests llvm-svn: 260493
2016-02-11 14:02:01 +08:00
%tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%idx.0 = add nsw i32 %tid.x, 2
[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
%arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds0, i32 0, i32 %idx.0
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%val0 = load float, float addrspace(3)* %arrayidx0, align 4
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
store i32 %idx.0, i32 addrspace(1)* %add_use, align 4
store float %val0, float addrspace(1)* %out
ret void
}
; Make sure once the first use is folded into the addressing mode, the
; remaining add use goes through the normal shl + add constant fold.
AMDGPU: Fix multi-use shl/add combine This was using a custom function that didn't handle the addressing modes properly for private. Use isLegalAddressingMode to avoid duplicating this. Additionally, skip the combine if there is only one use since the standard combine will handle it. llvm-svn: 318013
2017-11-13 13:11:54 +08:00
; GCN-LABEL: {{^}}load_shl_base_lds_1:
; GCN: v_lshlrev_b32_e32 [[PTR:v[0-9]+]], 2, {{v[0-9]+}}
; GCN: ds_read_b32 [[RESULT:v[0-9]+]], [[PTR]] offset:8
; GCN: v_add_i32_e32 [[ADDUSE:v[0-9]+]], vcc, 8, v{{[0-9]+}}
; GCN-DAG: buffer_store_dword [[RESULT]]
; GCN-DAG: buffer_store_dword [[ADDUSE]]
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @load_shl_base_lds_1(float addrspace(1)* %out, i32 addrspace(1)* %add_use) #0 {
AMDGPU: Remove some old intrinsic uses from tests llvm-svn: 260493
2016-02-11 14:02:01 +08:00
%tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%idx.0 = add nsw i32 %tid.x, 2
[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
%arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds0, i32 0, i32 %idx.0
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%val0 = load float, float addrspace(3)* %arrayidx0, align 4
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%shl_add_use = shl i32 %idx.0, 2
store i32 %shl_add_use, i32 addrspace(1)* %add_use, align 4
store float %val0, float addrspace(1)* %out
ret void
}
R600: Error on initializer for LDS. Also give a proper error for other address spaces. llvm-svn: 221917
2014-11-14 03:56:13 +08:00
@maxlds = addrspace(3) global [65536 x i8] undef, align 4
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
AMDGPU: Fix multi-use shl/add combine This was using a custom function that didn't handle the addressing modes properly for private. Use isLegalAddressingMode to avoid duplicating this. Additionally, skip the combine if there is only one use since the standard combine will handle it. llvm-svn: 318013
2017-11-13 13:11:54 +08:00
; GCN-LABEL: {{^}}load_shl_base_lds_max_offset
; GCN: ds_read_u8 v{{[0-9]+}}, v{{[0-9]+}} offset:65535
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @load_shl_base_lds_max_offset(i8 addrspace(1)* %out, i8 addrspace(3)* %lds, i32 addrspace(1)* %add_use) #0 {
AMDGPU: Remove some old intrinsic uses from tests llvm-svn: 260493
2016-02-11 14:02:01 +08:00
%tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%idx.0 = add nsw i32 %tid.x, 65535
[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
%arrayidx0 = getelementptr inbounds [65536 x i8], [65536 x i8] addrspace(3)* @maxlds, i32 0, i32 %idx.0
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%val0 = load i8, i8 addrspace(3)* %arrayidx0
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
store i32 %idx.0, i32 addrspace(1)* %add_use
store i8 %val0, i8 addrspace(1)* %out
ret void
}
; The two globals are placed adjacent in memory, so the same base
; pointer can be used with an offset into the second one.
AMDGPU: Fix multi-use shl/add combine This was using a custom function that didn't handle the addressing modes properly for private. Use isLegalAddressingMode to avoid duplicating this. Additionally, skip the combine if there is only one use since the standard combine will handle it. llvm-svn: 318013
2017-11-13 13:11:54 +08:00
; GCN-LABEL: {{^}}load_shl_base_lds_2:
; GCN: v_lshlrev_b32_e32 [[PTR:v[0-9]+]], 2, {{v[0-9]+}}
; GCN: s_mov_b32 m0, -1
; GCN-NEXT: ds_read2st64_b32 {{v\[[0-9]+:[0-9]+\]}}, [[PTR]] offset0:1 offset1:9
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @load_shl_base_lds_2(float addrspace(1)* %out) #0 {
AMDGPU: Remove some old intrinsic uses from tests llvm-svn: 260493
2016-02-11 14:02:01 +08:00
%tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%idx.0 = add nsw i32 %tid.x, 64
[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
%arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds0, i32 0, i32 %idx.0
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%val0 = load float, float addrspace(3)* %arrayidx0, align 4
[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
%arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds1, i32 0, i32 %idx.0
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%val1 = load float, float addrspace(3)* %arrayidx1, align 4
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%sum = fadd float %val0, %val1
store float %sum, float addrspace(1)* %out, align 4
ret void
}
AMDGPU: Fix multi-use shl/add combine This was using a custom function that didn't handle the addressing modes properly for private. Use isLegalAddressingMode to avoid duplicating this. Additionally, skip the combine if there is only one use since the standard combine will handle it. llvm-svn: 318013
2017-11-13 13:11:54 +08:00
; GCN-LABEL: {{^}}store_shl_base_lds_0:
; GCN: v_lshlrev_b32_e32 [[PTR:v[0-9]+]], 2, {{v[0-9]+}}
; GCN: ds_write_b32 [[PTR]], {{v[0-9]+}} offset:8
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @store_shl_base_lds_0(float addrspace(1)* %out, i32 addrspace(1)* %add_use) #0 {
AMDGPU: Remove some old intrinsic uses from tests llvm-svn: 260493
2016-02-11 14:02:01 +08:00
%tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%idx.0 = add nsw i32 %tid.x, 2
[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
%arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds0, i32 0, i32 %idx.0
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
store float 1.0, float addrspace(3)* %arrayidx0, align 4
store i32 %idx.0, i32 addrspace(1)* %add_use, align 4
ret void
}
; --------------------------------------------------------------------------------
; Atomics.
R600: Error on initializer for LDS. Also give a proper error for other address spaces. llvm-svn: 221917
2014-11-14 03:56:13 +08:00
@lds2 = addrspace(3) global [512 x i32] undef, align 4
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
; define amdgpu_kernel void @atomic_load_shl_base_lds_0(i32 addrspace(1)* %out, i32 addrspace(1)* %add_use) #0 {
AMDGPU: Remove some old intrinsic uses from tests llvm-svn: 260493
2016-02-11 14:02:01 +08:00
; %tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
; %idx.0 = add nsw i32 %tid.x, 2
[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
; %arrayidx0 = getelementptr inbounds [512 x i32], [512 x i32] addrspace(3)* @lds2, i32 0, i32 %idx.0
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
; %val = load atomic i32, i32 addrspace(3)* %arrayidx0 seq_cst, align 4
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
; store i32 %val, i32 addrspace(1)* %out, align 4
; store i32 %idx.0, i32 addrspace(1)* %add_use, align 4
; ret void
; }
AMDGPU: Fix multi-use shl/add combine This was using a custom function that didn't handle the addressing modes properly for private. Use isLegalAddressingMode to avoid duplicating this. Additionally, skip the combine if there is only one use since the standard combine will handle it. llvm-svn: 318013
2017-11-13 13:11:54 +08:00
; GCN-LABEL: {{^}}atomic_cmpxchg_shl_base_lds_0:
; GCN: v_lshlrev_b32_e32 [[PTR:v[0-9]+]], 2, {{v[0-9]+}}
; GCN: ds_cmpst_rtn_b32 {{v[0-9]+}}, [[PTR]], {{v[0-9]+}}, {{v[0-9]+}} offset:8
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @atomic_cmpxchg_shl_base_lds_0(i32 addrspace(1)* %out, i32 addrspace(1)* %add_use, i32 %swap) #0 {
AMDGPU: Remove some old intrinsic uses from tests llvm-svn: 260493
2016-02-11 14:02:01 +08:00
%tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%idx.0 = add nsw i32 %tid.x, 2
[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
%arrayidx0 = getelementptr inbounds [512 x i32], [512 x i32] addrspace(3)* @lds2, i32 0, i32 %idx.0
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%pair = cmpxchg i32 addrspace(3)* %arrayidx0, i32 7, i32 %swap seq_cst monotonic
%result = extractvalue { i32, i1 } %pair, 0
store i32 %result, i32 addrspace(1)* %out, align 4
store i32 %idx.0, i32 addrspace(1)* %add_use, align 4
ret void
}
AMDGPU: Fix multi-use shl/add combine This was using a custom function that didn't handle the addressing modes properly for private. Use isLegalAddressingMode to avoid duplicating this. Additionally, skip the combine if there is only one use since the standard combine will handle it. llvm-svn: 318013
2017-11-13 13:11:54 +08:00
; GCN-LABEL: {{^}}atomic_swap_shl_base_lds_0:
; GCN: v_lshlrev_b32_e32 [[PTR:v[0-9]+]], 2, {{v[0-9]+}}
; GCN: ds_wrxchg_rtn_b32 {{v[0-9]+}}, [[PTR]], {{v[0-9]+}} offset:8
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @atomic_swap_shl_base_lds_0(i32 addrspace(1)* %out, i32 addrspace(1)* %add_use) #0 {
AMDGPU: Remove some old intrinsic uses from tests llvm-svn: 260493
2016-02-11 14:02:01 +08:00
%tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%idx.0 = add nsw i32 %tid.x, 2
[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
%arrayidx0 = getelementptr inbounds [512 x i32], [512 x i32] addrspace(3)* @lds2, i32 0, i32 %idx.0
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%val = atomicrmw xchg i32 addrspace(3)* %arrayidx0, i32 3 seq_cst
store i32 %val, i32 addrspace(1)* %out, align 4
store i32 %idx.0, i32 addrspace(1)* %add_use, align 4
ret void
}
AMDGPU: Fix multi-use shl/add combine This was using a custom function that didn't handle the addressing modes properly for private. Use isLegalAddressingMode to avoid duplicating this. Additionally, skip the combine if there is only one use since the standard combine will handle it. llvm-svn: 318013
2017-11-13 13:11:54 +08:00
; GCN-LABEL: {{^}}atomic_add_shl_base_lds_0:
; GCN: v_lshlrev_b32_e32 [[PTR:v[0-9]+]], 2, {{v[0-9]+}}
; GCN: ds_add_rtn_u32 {{v[0-9]+}}, [[PTR]], {{v[0-9]+}} offset:8
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @atomic_add_shl_base_lds_0(i32 addrspace(1)* %out, i32 addrspace(1)* %add_use) #0 {
AMDGPU: Remove some old intrinsic uses from tests llvm-svn: 260493
2016-02-11 14:02:01 +08:00
%tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%idx.0 = add nsw i32 %tid.x, 2
[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
%arrayidx0 = getelementptr inbounds [512 x i32], [512 x i32] addrspace(3)* @lds2, i32 0, i32 %idx.0
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%val = atomicrmw add i32 addrspace(3)* %arrayidx0, i32 3 seq_cst
store i32 %val, i32 addrspace(1)* %out, align 4
store i32 %idx.0, i32 addrspace(1)* %add_use, align 4
ret void
}
AMDGPU: Fix multi-use shl/add combine This was using a custom function that didn't handle the addressing modes properly for private. Use isLegalAddressingMode to avoid duplicating this. Additionally, skip the combine if there is only one use since the standard combine will handle it. llvm-svn: 318013
2017-11-13 13:11:54 +08:00
; GCN-LABEL: {{^}}atomic_sub_shl_base_lds_0:
; GCN: v_lshlrev_b32_e32 [[PTR:v[0-9]+]], 2, {{v[0-9]+}}
; GCN: ds_sub_rtn_u32 {{v[0-9]+}}, [[PTR]], {{v[0-9]+}} offset:8
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @atomic_sub_shl_base_lds_0(i32 addrspace(1)* %out, i32 addrspace(1)* %add_use) #0 {
AMDGPU: Remove some old intrinsic uses from tests llvm-svn: 260493
2016-02-11 14:02:01 +08:00
%tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%idx.0 = add nsw i32 %tid.x, 2
[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
%arrayidx0 = getelementptr inbounds [512 x i32], [512 x i32] addrspace(3)* @lds2, i32 0, i32 %idx.0
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%val = atomicrmw sub i32 addrspace(3)* %arrayidx0, i32 3 seq_cst
store i32 %val, i32 addrspace(1)* %out, align 4
store i32 %idx.0, i32 addrspace(1)* %add_use, align 4
ret void
}
AMDGPU: Fix multi-use shl/add combine This was using a custom function that didn't handle the addressing modes properly for private. Use isLegalAddressingMode to avoid duplicating this. Additionally, skip the combine if there is only one use since the standard combine will handle it. llvm-svn: 318013
2017-11-13 13:11:54 +08:00
; GCN-LABEL: {{^}}atomic_and_shl_base_lds_0:
; GCN: v_lshlrev_b32_e32 [[PTR:v[0-9]+]], 2, {{v[0-9]+}}
; GCN: ds_and_rtn_b32 {{v[0-9]+}}, [[PTR]], {{v[0-9]+}} offset:8
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @atomic_and_shl_base_lds_0(i32 addrspace(1)* %out, i32 addrspace(1)* %add_use) #0 {
AMDGPU: Remove some old intrinsic uses from tests llvm-svn: 260493
2016-02-11 14:02:01 +08:00
%tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%idx.0 = add nsw i32 %tid.x, 2
[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
%arrayidx0 = getelementptr inbounds [512 x i32], [512 x i32] addrspace(3)* @lds2, i32 0, i32 %idx.0
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%val = atomicrmw and i32 addrspace(3)* %arrayidx0, i32 3 seq_cst
store i32 %val, i32 addrspace(1)* %out, align 4
store i32 %idx.0, i32 addrspace(1)* %add_use, align 4
ret void
}
AMDGPU: Fix multi-use shl/add combine This was using a custom function that didn't handle the addressing modes properly for private. Use isLegalAddressingMode to avoid duplicating this. Additionally, skip the combine if there is only one use since the standard combine will handle it. llvm-svn: 318013
2017-11-13 13:11:54 +08:00
; GCN-LABEL: {{^}}atomic_or_shl_base_lds_0:
; GCN: v_lshlrev_b32_e32 [[PTR:v[0-9]+]], 2, {{v[0-9]+}}
; GCN: ds_or_rtn_b32 {{v[0-9]+}}, [[PTR]], {{v[0-9]+}} offset:8
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @atomic_or_shl_base_lds_0(i32 addrspace(1)* %out, i32 addrspace(1)* %add_use) #0 {
AMDGPU: Remove some old intrinsic uses from tests llvm-svn: 260493
2016-02-11 14:02:01 +08:00
%tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%idx.0 = add nsw i32 %tid.x, 2
[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
%arrayidx0 = getelementptr inbounds [512 x i32], [512 x i32] addrspace(3)* @lds2, i32 0, i32 %idx.0
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%val = atomicrmw or i32 addrspace(3)* %arrayidx0, i32 3 seq_cst
store i32 %val, i32 addrspace(1)* %out, align 4
store i32 %idx.0, i32 addrspace(1)* %add_use, align 4
ret void
}
AMDGPU: Fix multi-use shl/add combine This was using a custom function that didn't handle the addressing modes properly for private. Use isLegalAddressingMode to avoid duplicating this. Additionally, skip the combine if there is only one use since the standard combine will handle it. llvm-svn: 318013
2017-11-13 13:11:54 +08:00
; GCN-LABEL: {{^}}atomic_xor_shl_base_lds_0:
; GCN: v_lshlrev_b32_e32 [[PTR:v[0-9]+]], 2, {{v[0-9]+}}
; GCN: ds_xor_rtn_b32 {{v[0-9]+}}, [[PTR]], {{v[0-9]+}} offset:8
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @atomic_xor_shl_base_lds_0(i32 addrspace(1)* %out, i32 addrspace(1)* %add_use) #0 {
AMDGPU: Remove some old intrinsic uses from tests llvm-svn: 260493
2016-02-11 14:02:01 +08:00
%tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%idx.0 = add nsw i32 %tid.x, 2
[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
%arrayidx0 = getelementptr inbounds [512 x i32], [512 x i32] addrspace(3)* @lds2, i32 0, i32 %idx.0
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%val = atomicrmw xor i32 addrspace(3)* %arrayidx0, i32 3 seq_cst
store i32 %val, i32 addrspace(1)* %out, align 4
store i32 %idx.0, i32 addrspace(1)* %add_use, align 4
ret void
}
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
; define amdgpu_kernel void @atomic_nand_shl_base_lds_0(i32 addrspace(1)* %out, i32 addrspace(1)* %add_use) #0 {
AMDGPU: Remove some old intrinsic uses from tests llvm-svn: 260493
2016-02-11 14:02:01 +08:00
; %tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
; %idx.0 = add nsw i32 %tid.x, 2
[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
; %arrayidx0 = getelementptr inbounds [512 x i32], [512 x i32] addrspace(3)* @lds2, i32 0, i32 %idx.0
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
; %val = atomicrmw nand i32 addrspace(3)* %arrayidx0, i32 3 seq_cst
; store i32 %val, i32 addrspace(1)* %out, align 4
; store i32 %idx.0, i32 addrspace(1)* %add_use, align 4
; ret void
; }
AMDGPU: Fix multi-use shl/add combine This was using a custom function that didn't handle the addressing modes properly for private. Use isLegalAddressingMode to avoid duplicating this. Additionally, skip the combine if there is only one use since the standard combine will handle it. llvm-svn: 318013
2017-11-13 13:11:54 +08:00
; GCN-LABEL: {{^}}atomic_min_shl_base_lds_0:
; GCN: v_lshlrev_b32_e32 [[PTR:v[0-9]+]], 2, {{v[0-9]+}}
; GCN: ds_min_rtn_i32 {{v[0-9]+}}, [[PTR]], {{v[0-9]+}} offset:8
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @atomic_min_shl_base_lds_0(i32 addrspace(1)* %out, i32 addrspace(1)* %add_use) #0 {
AMDGPU: Remove some old intrinsic uses from tests llvm-svn: 260493
2016-02-11 14:02:01 +08:00
%tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%idx.0 = add nsw i32 %tid.x, 2
[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
%arrayidx0 = getelementptr inbounds [512 x i32], [512 x i32] addrspace(3)* @lds2, i32 0, i32 %idx.0
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%val = atomicrmw min i32 addrspace(3)* %arrayidx0, i32 3 seq_cst
store i32 %val, i32 addrspace(1)* %out, align 4
store i32 %idx.0, i32 addrspace(1)* %add_use, align 4
ret void
}
AMDGPU: Fix multi-use shl/add combine This was using a custom function that didn't handle the addressing modes properly for private. Use isLegalAddressingMode to avoid duplicating this. Additionally, skip the combine if there is only one use since the standard combine will handle it. llvm-svn: 318013
2017-11-13 13:11:54 +08:00
; GCN-LABEL: {{^}}atomic_max_shl_base_lds_0:
; GCN: v_lshlrev_b32_e32 [[PTR:v[0-9]+]], 2, {{v[0-9]+}}
; GCN: ds_max_rtn_i32 {{v[0-9]+}}, [[PTR]], {{v[0-9]+}} offset:8
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @atomic_max_shl_base_lds_0(i32 addrspace(1)* %out, i32 addrspace(1)* %add_use) #0 {
AMDGPU: Remove some old intrinsic uses from tests llvm-svn: 260493
2016-02-11 14:02:01 +08:00
%tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%idx.0 = add nsw i32 %tid.x, 2
[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
%arrayidx0 = getelementptr inbounds [512 x i32], [512 x i32] addrspace(3)* @lds2, i32 0, i32 %idx.0
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%val = atomicrmw max i32 addrspace(3)* %arrayidx0, i32 3 seq_cst
store i32 %val, i32 addrspace(1)* %out, align 4
store i32 %idx.0, i32 addrspace(1)* %add_use, align 4
ret void
}
AMDGPU: Fix multi-use shl/add combine This was using a custom function that didn't handle the addressing modes properly for private. Use isLegalAddressingMode to avoid duplicating this. Additionally, skip the combine if there is only one use since the standard combine will handle it. llvm-svn: 318013
2017-11-13 13:11:54 +08:00
; GCN-LABEL: {{^}}atomic_umin_shl_base_lds_0:
; GCN: v_lshlrev_b32_e32 [[PTR:v[0-9]+]], 2, {{v[0-9]+}}
; GCN: ds_min_rtn_u32 {{v[0-9]+}}, [[PTR]], {{v[0-9]+}} offset:8
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @atomic_umin_shl_base_lds_0(i32 addrspace(1)* %out, i32 addrspace(1)* %add_use) #0 {
AMDGPU: Remove some old intrinsic uses from tests llvm-svn: 260493
2016-02-11 14:02:01 +08:00
%tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%idx.0 = add nsw i32 %tid.x, 2
[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
%arrayidx0 = getelementptr inbounds [512 x i32], [512 x i32] addrspace(3)* @lds2, i32 0, i32 %idx.0
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%val = atomicrmw umin i32 addrspace(3)* %arrayidx0, i32 3 seq_cst
store i32 %val, i32 addrspace(1)* %out, align 4
store i32 %idx.0, i32 addrspace(1)* %add_use, align 4
ret void
}
AMDGPU: Fix multi-use shl/add combine This was using a custom function that didn't handle the addressing modes properly for private. Use isLegalAddressingMode to avoid duplicating this. Additionally, skip the combine if there is only one use since the standard combine will handle it. llvm-svn: 318013
2017-11-13 13:11:54 +08:00
; GCN-LABEL: {{^}}atomic_umax_shl_base_lds_0:
; GCN: v_lshlrev_b32_e32 [[PTR:v[0-9]+]], 2, {{v[0-9]+}}
; GCN: ds_max_rtn_u32 {{v[0-9]+}}, [[PTR]], {{v[0-9]+}} offset:8
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @atomic_umax_shl_base_lds_0(i32 addrspace(1)* %out, i32 addrspace(1)* %add_use) #0 {
AMDGPU: Remove some old intrinsic uses from tests llvm-svn: 260493
2016-02-11 14:02:01 +08:00
%tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%idx.0 = add nsw i32 %tid.x, 2
[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
%arrayidx0 = getelementptr inbounds [512 x i32], [512 x i32] addrspace(3)* @lds2, i32 0, i32 %idx.0
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
%val = atomicrmw umax i32 addrspace(3)* %arrayidx0, i32 3 seq_cst
store i32 %val, i32 addrspace(1)* %out, align 4
store i32 %idx.0, i32 addrspace(1)* %add_use, align 4
ret void
}
AMDGPU: Fix multi-use shl/add combine This was using a custom function that didn't handle the addressing modes properly for private. Use isLegalAddressingMode to avoid duplicating this. Additionally, skip the combine if there is only one use since the standard combine will handle it. llvm-svn: 318013
2017-11-13 13:11:54 +08:00
; GCN-LABEL: {{^}}shl_add_ptr_combine_2use_lds:
; GCN: v_lshlrev_b32_e32 [[SCALE0:v[0-9]+]], 3, v0
; GCN: ds_write_b32 [[SCALE0]], v{{[0-9]+}} offset:32
; GCN: v_lshlrev_b32_e32 [[SCALE1:v[0-9]+]], 4, v0
; GCN: ds_write_b32 [[SCALE1]], v{{[0-9]+}} offset:64
define void @shl_add_ptr_combine_2use_lds(i32 %idx) #0 {
%idx.add = add nuw i32 %idx, 4
%shl0 = shl i32 %idx.add, 3
%shl1 = shl i32 %idx.add, 4
%ptr0 = inttoptr i32 %shl0 to i32 addrspace(3)*
%ptr1 = inttoptr i32 %shl1 to i32 addrspace(3)*
store volatile i32 9, i32 addrspace(3)* %ptr0
store volatile i32 10, i32 addrspace(3)* %ptr1
ret void
}
; GCN-LABEL: {{^}}shl_add_ptr_combine_2use_max_lds_offset:
; GCN-DAG: v_lshlrev_b32_e32 [[SCALE0:v[0-9]+]], 3, v0
; GCN-DAG: v_lshlrev_b32_e32 [[SCALE1:v[0-9]+]], 4, v0
; GCN-DAG: ds_write_b32 [[SCALE0]], v{{[0-9]+}} offset:65528
; GCN-DAG: v_add_i32_e32 [[ADD1:v[0-9]+]], vcc, 0x1fff0, [[SCALE1]]
; GCN: ds_write_b32 [[ADD1]], v{{[0-9]+$}}
define void @shl_add_ptr_combine_2use_max_lds_offset(i32 %idx) #0 {
%idx.add = add nuw i32 %idx, 8191
%shl0 = shl i32 %idx.add, 3
%shl1 = shl i32 %idx.add, 4
%ptr0 = inttoptr i32 %shl0 to i32 addrspace(3)*
%ptr1 = inttoptr i32 %shl1 to i32 addrspace(3)*
store volatile i32 9, i32 addrspace(3)* %ptr0
store volatile i32 10, i32 addrspace(3)* %ptr1
ret void
}
; GCN-LABEL: {{^}}shl_add_ptr_combine_2use_both_max_lds_offset:
; GCN: v_add_i32_e32 [[ADD:v[0-9]+]], vcc, 0x1000, v0
; GCN-DAG: v_lshlrev_b32_e32 [[SCALE0:v[0-9]+]], 4, [[ADD]]
; GCN-DAG: v_lshlrev_b32_e32 [[SCALE1:v[0-9]+]], 5, [[ADD]]
; GCN-DAG: ds_write_b32 [[SCALE0]], v{{[0-9]+$}}
; GCN: ds_write_b32 [[SCALE1]], v{{[0-9]+$}}
define void @shl_add_ptr_combine_2use_both_max_lds_offset(i32 %idx) #0 {
%idx.add = add nuw i32 %idx, 4096
%shl0 = shl i32 %idx.add, 4
%shl1 = shl i32 %idx.add, 5
%ptr0 = inttoptr i32 %shl0 to i32 addrspace(3)*
%ptr1 = inttoptr i32 %shl1 to i32 addrspace(3)*
store volatile i32 9, i32 addrspace(3)* %ptr0
store volatile i32 10, i32 addrspace(3)* %ptr1
ret void
}
; GCN-LABEL: {{^}}shl_add_ptr_combine_2use_private:
; GCN: v_lshlrev_b32_e32 [[SCALE0:v[0-9]+]], 2, v0
; GCN: buffer_store_dword v{{[0-9]+}}, [[SCALE0]], s[0:3], s4 offen offset:16
; GCN: v_lshlrev_b32_e32 [[SCALE1:v[0-9]+]], 3, v0
; GCN: buffer_store_dword v{{[0-9]+}}, [[SCALE1]], s[0:3], s4 offen offset:32
define void @shl_add_ptr_combine_2use_private(i16 zeroext %idx.arg) #0 {
%idx = zext i16 %idx.arg to i32
%idx.add = add nuw i32 %idx, 4
%shl0 = shl i32 %idx.add, 2
%shl1 = shl i32 %idx.add, 3
%ptr0 = inttoptr i32 %shl0 to i32*
%ptr1 = inttoptr i32 %shl1 to i32*
store volatile i32 9, i32* %ptr0
store volatile i32 10, i32* %ptr1
ret void
}
; GCN-LABEL: {{^}}shl_add_ptr_combine_2use_max_private_offset:
; GCN-DAG: v_lshlrev_b32_e32 [[SCALE0:v[0-9]+]], 3, v0
; GCN-DAG: v_lshlrev_b32_e32 [[SCALE1:v[0-9]+]], 4, v0
; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[SCALE0]], s[0:3], s4 offen offset:4088
; GCN-DAG: v_add_i32_e32 [[ADD:v[0-9]+]], vcc, 0x1ff0, [[SCALE1]]
; GCN: buffer_store_dword v{{[0-9]+}}, [[ADD]], s[0:3], s4 offen{{$}}
define void @shl_add_ptr_combine_2use_max_private_offset(i16 zeroext %idx.arg) #0 {
%idx = zext i16 %idx.arg to i32
%idx.add = add nuw i32 %idx, 511
%shl0 = shl i32 %idx.add, 3
%shl1 = shl i32 %idx.add, 4
%ptr0 = inttoptr i32 %shl0 to i32*
%ptr1 = inttoptr i32 %shl1 to i32*
store volatile i32 9, i32* %ptr0
store volatile i32 10, i32* %ptr1
ret void
}
; GCN-LABEL: {{^}}shl_add_ptr_combine_2use_both_max_private_offset:
; GCN: v_add_i32_e32 [[ADD:v[0-9]+]], vcc, 0x100, v0
; GCN-DAG: v_lshlrev_b32_e32 [[SCALE0:v[0-9]+]], 4, [[ADD]]
; GCN-DAG: v_lshlrev_b32_e32 [[SCALE1:v[0-9]+]], 5, [[ADD]]
; GCN-DAG: buffer_store_dword v{{[0-9]+}}, [[SCALE0]], s[0:3], s4 offen{{$}}
; GCN: buffer_store_dword v{{[0-9]+}}, [[SCALE1]], s[0:3], s4 offen{{$}}
define void @shl_add_ptr_combine_2use_both_max_private_offset(i16 zeroext %idx.arg) #0 {
%idx = zext i16 %idx.arg to i32
%idx.add = add nuw i32 %idx, 256
%shl0 = shl i32 %idx.add, 4
%shl1 = shl i32 %idx.add, 5
%ptr0 = inttoptr i32 %shl0 to i32*
%ptr1 = inttoptr i32 %shl1 to i32*
store volatile i32 9, i32* %ptr0
store volatile i32 10, i32* %ptr1
ret void
}
R600/SI: Fix offset folding in some cases with shifted pointers. Ordinarily (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) is only done if the add has one use. If the resulting constant add can be folded into an addressing mode, force this to happen for the pointer operand. This ends up happening a lot because of how LDS objects are allocated. Since the globals are allocated next to each other, acessing the first element of the second object is directly indexed by a shifted pointer. llvm-svn: 215739
2014-08-16 01:49:05 +08:00
attributes #0 = { nounwind }
attributes #1 = { nounwind readnone }