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/X86/unaligned-32-byte-memops.ll

279 lines
11 KiB
LLVM
Raw Normal View History

fixed to test attributes and use better checks 1. Used update_llc_test_checks.py to tighten checks 2. Fixed triple (nothing Darwin-specific here) 3. Replaced CPU specifiers with attributes 4. Fixed comments 5. Removed IvyBridge run because it did not add any coverage llvm-svn: 240058
2015-06-19 05:12:24 +08:00
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx,+slow-unaligned-mem-32 | FileCheck %s --check-prefix=AVXSLOW
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx,-slow-unaligned-mem-32 | FileCheck %s --check-prefix=AVXFAST
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx2 | FileCheck %s --check-prefix=AVX2
Add a feature flag for slow 32-byte unaligned memory accesses [x86]. This patch adds a feature flag to avoid unaligned 32-byte load/store AVX codegen for Sandy Bridge and Ivy Bridge. There is no functionality change intended for those chips. Previously, the absence of AVX2 was being used as a proxy to detect this feature. But that hindered codegen for AVX-enabled AMD chips such as btver2 that do not have the 32-byte unaligned access slowdown. Performance measurements are included in PR21541 ( http://llvm.org/bugs/show_bug.cgi?id=21541 ). Differential Revision: http://reviews.llvm.org/D6355 llvm-svn: 222544
2014-11-22 01:40:04 +08:00
fixed to test attributes and use better checks 1. Used update_llc_test_checks.py to tighten checks 2. Fixed triple (nothing Darwin-specific here) 3. Replaced CPU specifiers with attributes 4. Fixed comments 5. Removed IvyBridge run because it did not add any coverage llvm-svn: 240058
2015-06-19 05:12:24 +08:00
; Don't generate an unaligned 32-byte load on this test if that is slower than two 16-byte loads.
Add a feature flag for slow 32-byte unaligned memory accesses [x86]. This patch adds a feature flag to avoid unaligned 32-byte load/store AVX codegen for Sandy Bridge and Ivy Bridge. There is no functionality change intended for those chips. Previously, the absence of AVX2 was being used as a proxy to detect this feature. But that hindered codegen for AVX-enabled AMD chips such as btver2 that do not have the 32-byte unaligned access slowdown. Performance measurements are included in PR21541 ( http://llvm.org/bugs/show_bug.cgi?id=21541 ). Differential Revision: http://reviews.llvm.org/D6355 llvm-svn: 222544
2014-11-22 01:40:04 +08:00
define <8 x float> @load32bytes(<8 x float>* %Ap) {
fixed to test attributes and use better checks 1. Used update_llc_test_checks.py to tighten checks 2. Fixed triple (nothing Darwin-specific here) 3. Replaced CPU specifiers with attributes 4. Fixed comments 5. Removed IvyBridge run because it did not add any coverage llvm-svn: 240058
2015-06-19 05:12:24 +08:00
; AVXSLOW-LABEL: load32bytes:
; AVXSLOW: # BB#0:
; AVXSLOW-NEXT: vmovaps (%rdi), %xmm0
; AVXSLOW-NEXT: vinsertf128 $1, 16(%rdi), %ymm0, %ymm0
; AVXSLOW-NEXT: retq
;
; AVXFAST-LABEL: load32bytes:
; AVXFAST: # BB#0:
; AVXFAST-NEXT: vmovups (%rdi), %ymm0
; AVXFAST-NEXT: retq
;
; AVX2-LABEL: load32bytes:
; AVX2: # BB#0:
; AVX2-NEXT: vmovups (%rdi), %ymm0
; AVX2-NEXT: retq
[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
%A = load <8 x float>, <8 x float>* %Ap, align 16
Add a feature flag for slow 32-byte unaligned memory accesses [x86]. This patch adds a feature flag to avoid unaligned 32-byte load/store AVX codegen for Sandy Bridge and Ivy Bridge. There is no functionality change intended for those chips. Previously, the absence of AVX2 was being used as a proxy to detect this feature. But that hindered codegen for AVX-enabled AMD chips such as btver2 that do not have the 32-byte unaligned access slowdown. Performance measurements are included in PR21541 ( http://llvm.org/bugs/show_bug.cgi?id=21541 ). Differential Revision: http://reviews.llvm.org/D6355 llvm-svn: 222544
2014-11-22 01:40:04 +08:00
ret <8 x float> %A
}
fixed to test attributes and use better checks 1. Used update_llc_test_checks.py to tighten checks 2. Fixed triple (nothing Darwin-specific here) 3. Replaced CPU specifiers with attributes 4. Fixed comments 5. Removed IvyBridge run because it did not add any coverage llvm-svn: 240058
2015-06-19 05:12:24 +08:00
; Don't generate an unaligned 32-byte store on this test if that is slower than two 16-byte loads.
Add a feature flag for slow 32-byte unaligned memory accesses [x86]. This patch adds a feature flag to avoid unaligned 32-byte load/store AVX codegen for Sandy Bridge and Ivy Bridge. There is no functionality change intended for those chips. Previously, the absence of AVX2 was being used as a proxy to detect this feature. But that hindered codegen for AVX-enabled AMD chips such as btver2 that do not have the 32-byte unaligned access slowdown. Performance measurements are included in PR21541 ( http://llvm.org/bugs/show_bug.cgi?id=21541 ). Differential Revision: http://reviews.llvm.org/D6355 llvm-svn: 222544
2014-11-22 01:40:04 +08:00
define void @store32bytes(<8 x float> %A, <8 x float>* %P) {
fixed to test attributes and use better checks 1. Used update_llc_test_checks.py to tighten checks 2. Fixed triple (nothing Darwin-specific here) 3. Replaced CPU specifiers with attributes 4. Fixed comments 5. Removed IvyBridge run because it did not add any coverage llvm-svn: 240058
2015-06-19 05:12:24 +08:00
; AVXSLOW-LABEL: store32bytes:
; AVXSLOW: # BB#0:
; AVXSLOW-NEXT: vextractf128 $1, %ymm0, 16(%rdi)
; AVXSLOW-NEXT: vmovaps %xmm0, (%rdi)
; AVXSLOW-NEXT: vzeroupper
; AVXSLOW-NEXT: retq
;
; AVXFAST-LABEL: store32bytes:
; AVXFAST: # BB#0:
; AVXFAST-NEXT: vmovups %ymm0, (%rdi)
; AVXFAST-NEXT: vzeroupper
; AVXFAST-NEXT: retq
;
; AVX2-LABEL: store32bytes:
; AVX2: # BB#0:
; AVX2-NEXT: vmovups %ymm0, (%rdi)
; AVX2-NEXT: vzeroupper
; AVX2-NEXT: retq
Add a feature flag for slow 32-byte unaligned memory accesses [x86]. This patch adds a feature flag to avoid unaligned 32-byte load/store AVX codegen for Sandy Bridge and Ivy Bridge. There is no functionality change intended for those chips. Previously, the absence of AVX2 was being used as a proxy to detect this feature. But that hindered codegen for AVX-enabled AMD chips such as btver2 that do not have the 32-byte unaligned access slowdown. Performance measurements are included in PR21541 ( http://llvm.org/bugs/show_bug.cgi?id=21541 ). Differential Revision: http://reviews.llvm.org/D6355 llvm-svn: 222544
2014-11-22 01:40:04 +08:00
store <8 x float> %A, <8 x float>* %P, align 16
ret void
}
combine consecutive subvector 16-byte loads into one 32-byte load This is a fix for PR21709 ( http://llvm.org/bugs/show_bug.cgi?id=21709 ). When we have 2 consecutive 16-byte loads that are merged into one 32-byte vector, we can use a single 32-byte load instead. But we don't do this for SandyBridge / IvyBridge because they have slower 32-byte memops. We also don't bother using 32-byte *integer* loads on a machine that only has AVX1 (btver2) because those operands would have to be split in half anyway since there is no support for 32-byte integer math ops. Differential Revision: http://reviews.llvm.org/D6492 llvm-svn: 224344
2014-12-17 00:30:01 +08:00
fixed to test attributes and use better checks 1. Used update_llc_test_checks.py to tighten checks 2. Fixed triple (nothing Darwin-specific here) 3. Replaced CPU specifiers with attributes 4. Fixed comments 5. Removed IvyBridge run because it did not add any coverage llvm-svn: 240058
2015-06-19 05:12:24 +08:00
; Merge two consecutive 16-byte subvector loads into a single 32-byte load if it's faster.
combine consecutive subvector 16-byte loads into one 32-byte load This is a fix for PR21709 ( http://llvm.org/bugs/show_bug.cgi?id=21709 ). When we have 2 consecutive 16-byte loads that are merged into one 32-byte vector, we can use a single 32-byte load instead. But we don't do this for SandyBridge / IvyBridge because they have slower 32-byte memops. We also don't bother using 32-byte *integer* loads on a machine that only has AVX1 (btver2) because those operands would have to be split in half anyway since there is no support for 32-byte integer math ops. Differential Revision: http://reviews.llvm.org/D6492 llvm-svn: 224344
2014-12-17 00:30:01 +08:00
define <8 x float> @combine_16_byte_loads_no_intrinsic(<4 x float>* %ptr) {
fixed to test attributes and use better checks 1. Used update_llc_test_checks.py to tighten checks 2. Fixed triple (nothing Darwin-specific here) 3. Replaced CPU specifiers with attributes 4. Fixed comments 5. Removed IvyBridge run because it did not add any coverage llvm-svn: 240058
2015-06-19 05:12:24 +08:00
; AVXSLOW-LABEL: combine_16_byte_loads_no_intrinsic:
; AVXSLOW: # BB#0:
; AVXSLOW-NEXT: vmovups 48(%rdi), %xmm0
; AVXSLOW-NEXT: vinsertf128 $1, 64(%rdi), %ymm0, %ymm0
; AVXSLOW-NEXT: retq
;
; AVXFAST-LABEL: combine_16_byte_loads_no_intrinsic:
; AVXFAST: # BB#0:
; AVXFAST-NEXT: vmovups 48(%rdi), %ymm0
; AVXFAST-NEXT: retq
;
; AVX2-LABEL: combine_16_byte_loads_no_intrinsic:
; AVX2: # BB#0:
; AVX2-NEXT: vmovups 48(%rdi), %ymm0
; AVX2-NEXT: retq
[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
%ptr1 = getelementptr inbounds <4 x float>, <4 x float>* %ptr, i64 3
%ptr2 = getelementptr inbounds <4 x float>, <4 x float>* %ptr, i64 4
[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
%v1 = load <4 x float>, <4 x float>* %ptr1, align 1
%v2 = load <4 x float>, <4 x float>* %ptr2, align 1
combine consecutive subvector 16-byte loads into one 32-byte load This is a fix for PR21709 ( http://llvm.org/bugs/show_bug.cgi?id=21709 ). When we have 2 consecutive 16-byte loads that are merged into one 32-byte vector, we can use a single 32-byte load instead. But we don't do this for SandyBridge / IvyBridge because they have slower 32-byte memops. We also don't bother using 32-byte *integer* loads on a machine that only has AVX1 (btver2) because those operands would have to be split in half anyway since there is no support for 32-byte integer math ops. Differential Revision: http://reviews.llvm.org/D6492 llvm-svn: 224344
2014-12-17 00:30:01 +08:00
%v3 = shufflevector <4 x float> %v1, <4 x float> %v2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
ret <8 x float> %v3
}
[x86] fix allowsMisalignedMemoryAccess() implementation This patch fixes the x86 implementation of allowsMisalignedMemoryAccess() to correctly return the 'Fast' output parameter for 32-byte accesses. To test that, an existing load merging optimization is changed to use the TLI hook. This exposes a shortcoming in the current logic and results in the regression test update. Changing other direct users of the isUnalignedMem32Slow() x86 CPU attribute would be a follow-on patch. Without the fix in allowsMisalignedMemoryAccesses(), we will infinite loop when targeting SandyBridge because LowerINSERT_SUBVECTOR() creates 32-byte loads from two 16-byte loads while PerformLOADCombine() splits them back into 16-byte loads. Differential Revision: http://reviews.llvm.org/D10662 llvm-svn: 245075
2015-08-15 01:53:40 +08:00
; If the first load is 32-byte aligned, then the loads should be merged in all cases.
add test to show suboptimal load merging behavior llvm-svn: 240063
2015-06-19 05:34:26 +08:00
define <8 x float> @combine_16_byte_loads_aligned(<4 x float>* %ptr) {
; AVXSLOW-LABEL: combine_16_byte_loads_aligned:
; AVXSLOW: # BB#0:
[x86] fix allowsMisalignedMemoryAccess() implementation This patch fixes the x86 implementation of allowsMisalignedMemoryAccess() to correctly return the 'Fast' output parameter for 32-byte accesses. To test that, an existing load merging optimization is changed to use the TLI hook. This exposes a shortcoming in the current logic and results in the regression test update. Changing other direct users of the isUnalignedMem32Slow() x86 CPU attribute would be a follow-on patch. Without the fix in allowsMisalignedMemoryAccesses(), we will infinite loop when targeting SandyBridge because LowerINSERT_SUBVECTOR() creates 32-byte loads from two 16-byte loads while PerformLOADCombine() splits them back into 16-byte loads. Differential Revision: http://reviews.llvm.org/D10662 llvm-svn: 245075
2015-08-15 01:53:40 +08:00
; AVXSLOW-NEXT: vmovaps 48(%rdi), %ymm0
add test to show suboptimal load merging behavior llvm-svn: 240063
2015-06-19 05:34:26 +08:00
; AVXSLOW-NEXT: retq
;
; AVXFAST-LABEL: combine_16_byte_loads_aligned:
; AVXFAST: # BB#0:
; AVXFAST-NEXT: vmovaps 48(%rdi), %ymm0
; AVXFAST-NEXT: retq
;
; AVX2-LABEL: combine_16_byte_loads_aligned:
; AVX2: # BB#0:
; AVX2-NEXT: vmovaps 48(%rdi), %ymm0
; AVX2-NEXT: retq
%ptr1 = getelementptr inbounds <4 x float>, <4 x float>* %ptr, i64 3
%ptr2 = getelementptr inbounds <4 x float>, <4 x float>* %ptr, i64 4
%v1 = load <4 x float>, <4 x float>* %ptr1, align 32
%v2 = load <4 x float>, <4 x float>* %ptr2, align 1
%v3 = shufflevector <4 x float> %v1, <4 x float> %v2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
ret <8 x float> %v3
}
fixed to test attributes and use better checks 1. Used update_llc_test_checks.py to tighten checks 2. Fixed triple (nothing Darwin-specific here) 3. Replaced CPU specifiers with attributes 4. Fixed comments 5. Removed IvyBridge run because it did not add any coverage llvm-svn: 240058
2015-06-19 05:12:24 +08:00
; Swap the order of the shufflevector operands to ensure that the pattern still matches.
combine consecutive subvector 16-byte loads into one 32-byte load This is a fix for PR21709 ( http://llvm.org/bugs/show_bug.cgi?id=21709 ). When we have 2 consecutive 16-byte loads that are merged into one 32-byte vector, we can use a single 32-byte load instead. But we don't do this for SandyBridge / IvyBridge because they have slower 32-byte memops. We also don't bother using 32-byte *integer* loads on a machine that only has AVX1 (btver2) because those operands would have to be split in half anyway since there is no support for 32-byte integer math ops. Differential Revision: http://reviews.llvm.org/D6492 llvm-svn: 224344
2014-12-17 00:30:01 +08:00
fixed to test attributes and use better checks 1. Used update_llc_test_checks.py to tighten checks 2. Fixed triple (nothing Darwin-specific here) 3. Replaced CPU specifiers with attributes 4. Fixed comments 5. Removed IvyBridge run because it did not add any coverage llvm-svn: 240058
2015-06-19 05:12:24 +08:00
define <8 x float> @combine_16_byte_loads_no_intrinsic_swap(<4 x float>* %ptr) {
; AVXSLOW-LABEL: combine_16_byte_loads_no_intrinsic_swap:
; AVXSLOW: # BB#0:
; AVXSLOW-NEXT: vmovups 64(%rdi), %xmm0
; AVXSLOW-NEXT: vinsertf128 $1, 80(%rdi), %ymm0, %ymm0
; AVXSLOW-NEXT: retq
;
; AVXFAST-LABEL: combine_16_byte_loads_no_intrinsic_swap:
; AVXFAST: # BB#0:
; AVXFAST-NEXT: vmovups 64(%rdi), %ymm0
; AVXFAST-NEXT: retq
;
; AVX2-LABEL: combine_16_byte_loads_no_intrinsic_swap:
; AVX2: # BB#0:
; AVX2-NEXT: vmovups 64(%rdi), %ymm0
; AVX2-NEXT: retq
[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
%ptr1 = getelementptr inbounds <4 x float>, <4 x float>* %ptr, i64 4
%ptr2 = getelementptr inbounds <4 x float>, <4 x float>* %ptr, i64 5
[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
%v1 = load <4 x float>, <4 x float>* %ptr1, align 1
%v2 = load <4 x float>, <4 x float>* %ptr2, align 1
combine consecutive subvector 16-byte loads into one 32-byte load This is a fix for PR21709 ( http://llvm.org/bugs/show_bug.cgi?id=21709 ). When we have 2 consecutive 16-byte loads that are merged into one 32-byte vector, we can use a single 32-byte load instead. But we don't do this for SandyBridge / IvyBridge because they have slower 32-byte memops. We also don't bother using 32-byte *integer* loads on a machine that only has AVX1 (btver2) because those operands would have to be split in half anyway since there is no support for 32-byte integer math ops. Differential Revision: http://reviews.llvm.org/D6492 llvm-svn: 224344
2014-12-17 00:30:01 +08:00
%v3 = shufflevector <4 x float> %v2, <4 x float> %v1, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 0, i32 1, i32 2, i32 3>
ret <8 x float> %v3
}
; Check each element type other than float to make sure it is handled correctly.
; Use the loaded values with an 'add' to make sure we're using the correct load type.
fixed to test attributes and use better checks 1. Used update_llc_test_checks.py to tighten checks 2. Fixed triple (nothing Darwin-specific here) 3. Replaced CPU specifiers with attributes 4. Fixed comments 5. Removed IvyBridge run because it did not add any coverage llvm-svn: 240058
2015-06-19 05:12:24 +08:00
; Don't generate 32-byte loads for integer ops unless we have AVX2.
combine consecutive subvector 16-byte loads into one 32-byte load This is a fix for PR21709 ( http://llvm.org/bugs/show_bug.cgi?id=21709 ). When we have 2 consecutive 16-byte loads that are merged into one 32-byte vector, we can use a single 32-byte load instead. But we don't do this for SandyBridge / IvyBridge because they have slower 32-byte memops. We also don't bother using 32-byte *integer* loads on a machine that only has AVX1 (btver2) because those operands would have to be split in half anyway since there is no support for 32-byte integer math ops. Differential Revision: http://reviews.llvm.org/D6492 llvm-svn: 224344
2014-12-17 00:30:01 +08:00
define <4 x i64> @combine_16_byte_loads_i64(<2 x i64>* %ptr, <4 x i64> %x) {
fixed to test attributes and use better checks 1. Used update_llc_test_checks.py to tighten checks 2. Fixed triple (nothing Darwin-specific here) 3. Replaced CPU specifiers with attributes 4. Fixed comments 5. Removed IvyBridge run because it did not add any coverage llvm-svn: 240058
2015-06-19 05:12:24 +08:00
; AVXSLOW-LABEL: combine_16_byte_loads_i64:
; AVXSLOW: # BB#0:
; AVXSLOW-NEXT: vextractf128 $1, %ymm0, %xmm1
; AVXSLOW-NEXT: vpaddq 96(%rdi), %xmm1, %xmm1
; AVXSLOW-NEXT: vpaddq 80(%rdi), %xmm0, %xmm0
; AVXSLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVXSLOW-NEXT: retq
;
; AVXFAST-LABEL: combine_16_byte_loads_i64:
; AVXFAST: # BB#0:
; AVXFAST-NEXT: vextractf128 $1, %ymm0, %xmm1
; AVXFAST-NEXT: vpaddq 96(%rdi), %xmm1, %xmm1
; AVXFAST-NEXT: vpaddq 80(%rdi), %xmm0, %xmm0
; AVXFAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVXFAST-NEXT: retq
;
; AVX2-LABEL: combine_16_byte_loads_i64:
; AVX2: # BB#0:
; AVX2-NEXT: vpaddq 80(%rdi), %ymm0, %ymm0
; AVX2-NEXT: retq
[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
%ptr1 = getelementptr inbounds <2 x i64>, <2 x i64>* %ptr, i64 5
%ptr2 = getelementptr inbounds <2 x i64>, <2 x i64>* %ptr, i64 6
[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
%v1 = load <2 x i64>, <2 x i64>* %ptr1, align 1
%v2 = load <2 x i64>, <2 x i64>* %ptr2, align 1
combine consecutive subvector 16-byte loads into one 32-byte load This is a fix for PR21709 ( http://llvm.org/bugs/show_bug.cgi?id=21709 ). When we have 2 consecutive 16-byte loads that are merged into one 32-byte vector, we can use a single 32-byte load instead. But we don't do this for SandyBridge / IvyBridge because they have slower 32-byte memops. We also don't bother using 32-byte *integer* loads on a machine that only has AVX1 (btver2) because those operands would have to be split in half anyway since there is no support for 32-byte integer math ops. Differential Revision: http://reviews.llvm.org/D6492 llvm-svn: 224344
2014-12-17 00:30:01 +08:00
%v3 = shufflevector <2 x i64> %v1, <2 x i64> %v2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
%v4 = add <4 x i64> %v3, %x
ret <4 x i64> %v4
}
define <8 x i32> @combine_16_byte_loads_i32(<4 x i32>* %ptr, <8 x i32> %x) {
fixed to test attributes and use better checks 1. Used update_llc_test_checks.py to tighten checks 2. Fixed triple (nothing Darwin-specific here) 3. Replaced CPU specifiers with attributes 4. Fixed comments 5. Removed IvyBridge run because it did not add any coverage llvm-svn: 240058
2015-06-19 05:12:24 +08:00
; AVXSLOW-LABEL: combine_16_byte_loads_i32:
; AVXSLOW: # BB#0:
; AVXSLOW-NEXT: vextractf128 $1, %ymm0, %xmm1
; AVXSLOW-NEXT: vpaddd 112(%rdi), %xmm1, %xmm1
; AVXSLOW-NEXT: vpaddd 96(%rdi), %xmm0, %xmm0
; AVXSLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVXSLOW-NEXT: retq
;
; AVXFAST-LABEL: combine_16_byte_loads_i32:
; AVXFAST: # BB#0:
; AVXFAST-NEXT: vextractf128 $1, %ymm0, %xmm1
; AVXFAST-NEXT: vpaddd 112(%rdi), %xmm1, %xmm1
; AVXFAST-NEXT: vpaddd 96(%rdi), %xmm0, %xmm0
; AVXFAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVXFAST-NEXT: retq
;
; AVX2-LABEL: combine_16_byte_loads_i32:
; AVX2: # BB#0:
; AVX2-NEXT: vpaddd 96(%rdi), %ymm0, %ymm0
; AVX2-NEXT: retq
[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
%ptr1 = getelementptr inbounds <4 x i32>, <4 x i32>* %ptr, i64 6
%ptr2 = getelementptr inbounds <4 x i32>, <4 x i32>* %ptr, i64 7
[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
%v1 = load <4 x i32>, <4 x i32>* %ptr1, align 1
%v2 = load <4 x i32>, <4 x i32>* %ptr2, align 1
combine consecutive subvector 16-byte loads into one 32-byte load This is a fix for PR21709 ( http://llvm.org/bugs/show_bug.cgi?id=21709 ). When we have 2 consecutive 16-byte loads that are merged into one 32-byte vector, we can use a single 32-byte load instead. But we don't do this for SandyBridge / IvyBridge because they have slower 32-byte memops. We also don't bother using 32-byte *integer* loads on a machine that only has AVX1 (btver2) because those operands would have to be split in half anyway since there is no support for 32-byte integer math ops. Differential Revision: http://reviews.llvm.org/D6492 llvm-svn: 224344
2014-12-17 00:30:01 +08:00
%v3 = shufflevector <4 x i32> %v1, <4 x i32> %v2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
%v4 = add <8 x i32> %v3, %x
ret <8 x i32> %v4
}
define <16 x i16> @combine_16_byte_loads_i16(<8 x i16>* %ptr, <16 x i16> %x) {
fixed to test attributes and use better checks 1. Used update_llc_test_checks.py to tighten checks 2. Fixed triple (nothing Darwin-specific here) 3. Replaced CPU specifiers with attributes 4. Fixed comments 5. Removed IvyBridge run because it did not add any coverage llvm-svn: 240058
2015-06-19 05:12:24 +08:00
; AVXSLOW-LABEL: combine_16_byte_loads_i16:
; AVXSLOW: # BB#0:
; AVXSLOW-NEXT: vextractf128 $1, %ymm0, %xmm1
; AVXSLOW-NEXT: vpaddw 128(%rdi), %xmm1, %xmm1
; AVXSLOW-NEXT: vpaddw 112(%rdi), %xmm0, %xmm0
; AVXSLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVXSLOW-NEXT: retq
;
; AVXFAST-LABEL: combine_16_byte_loads_i16:
; AVXFAST: # BB#0:
; AVXFAST-NEXT: vextractf128 $1, %ymm0, %xmm1
; AVXFAST-NEXT: vpaddw 128(%rdi), %xmm1, %xmm1
; AVXFAST-NEXT: vpaddw 112(%rdi), %xmm0, %xmm0
; AVXFAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVXFAST-NEXT: retq
;
; AVX2-LABEL: combine_16_byte_loads_i16:
; AVX2: # BB#0:
; AVX2-NEXT: vpaddw 112(%rdi), %ymm0, %ymm0
; AVX2-NEXT: retq
[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
%ptr1 = getelementptr inbounds <8 x i16>, <8 x i16>* %ptr, i64 7
%ptr2 = getelementptr inbounds <8 x i16>, <8 x i16>* %ptr, i64 8
[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
%v1 = load <8 x i16>, <8 x i16>* %ptr1, align 1
%v2 = load <8 x i16>, <8 x i16>* %ptr2, align 1
combine consecutive subvector 16-byte loads into one 32-byte load This is a fix for PR21709 ( http://llvm.org/bugs/show_bug.cgi?id=21709 ). When we have 2 consecutive 16-byte loads that are merged into one 32-byte vector, we can use a single 32-byte load instead. But we don't do this for SandyBridge / IvyBridge because they have slower 32-byte memops. We also don't bother using 32-byte *integer* loads on a machine that only has AVX1 (btver2) because those operands would have to be split in half anyway since there is no support for 32-byte integer math ops. Differential Revision: http://reviews.llvm.org/D6492 llvm-svn: 224344
2014-12-17 00:30:01 +08:00
%v3 = shufflevector <8 x i16> %v1, <8 x i16> %v2, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
%v4 = add <16 x i16> %v3, %x
ret <16 x i16> %v4
}
define <32 x i8> @combine_16_byte_loads_i8(<16 x i8>* %ptr, <32 x i8> %x) {
fixed to test attributes and use better checks 1. Used update_llc_test_checks.py to tighten checks 2. Fixed triple (nothing Darwin-specific here) 3. Replaced CPU specifiers with attributes 4. Fixed comments 5. Removed IvyBridge run because it did not add any coverage llvm-svn: 240058
2015-06-19 05:12:24 +08:00
; AVXSLOW-LABEL: combine_16_byte_loads_i8:
; AVXSLOW: # BB#0:
; AVXSLOW-NEXT: vextractf128 $1, %ymm0, %xmm1
; AVXSLOW-NEXT: vpaddb 144(%rdi), %xmm1, %xmm1
; AVXSLOW-NEXT: vpaddb 128(%rdi), %xmm0, %xmm0
; AVXSLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVXSLOW-NEXT: retq
;
; AVXFAST-LABEL: combine_16_byte_loads_i8:
; AVXFAST: # BB#0:
; AVXFAST-NEXT: vextractf128 $1, %ymm0, %xmm1
; AVXFAST-NEXT: vpaddb 144(%rdi), %xmm1, %xmm1
; AVXFAST-NEXT: vpaddb 128(%rdi), %xmm0, %xmm0
; AVXFAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
; AVXFAST-NEXT: retq
;
; AVX2-LABEL: combine_16_byte_loads_i8:
; AVX2: # BB#0:
; AVX2-NEXT: vpaddb 128(%rdi), %ymm0, %ymm0
; AVX2-NEXT: retq
[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
%ptr1 = getelementptr inbounds <16 x i8>, <16 x i8>* %ptr, i64 8
%ptr2 = getelementptr inbounds <16 x i8>, <16 x i8>* %ptr, i64 9
[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
%v1 = load <16 x i8>, <16 x i8>* %ptr1, align 1
%v2 = load <16 x i8>, <16 x i8>* %ptr2, align 1
combine consecutive subvector 16-byte loads into one 32-byte load This is a fix for PR21709 ( http://llvm.org/bugs/show_bug.cgi?id=21709 ). When we have 2 consecutive 16-byte loads that are merged into one 32-byte vector, we can use a single 32-byte load instead. But we don't do this for SandyBridge / IvyBridge because they have slower 32-byte memops. We also don't bother using 32-byte *integer* loads on a machine that only has AVX1 (btver2) because those operands would have to be split in half anyway since there is no support for 32-byte integer math ops. Differential Revision: http://reviews.llvm.org/D6492 llvm-svn: 224344
2014-12-17 00:30:01 +08:00
%v3 = shufflevector <16 x i8> %v1, <16 x i8> %v2, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
%v4 = add <32 x i8> %v3, %x
ret <32 x i8> %v4
}
define <4 x double> @combine_16_byte_loads_double(<2 x double>* %ptr, <4 x double> %x) {
fixed to test attributes and use better checks 1. Used update_llc_test_checks.py to tighten checks 2. Fixed triple (nothing Darwin-specific here) 3. Replaced CPU specifiers with attributes 4. Fixed comments 5. Removed IvyBridge run because it did not add any coverage llvm-svn: 240058
2015-06-19 05:12:24 +08:00
; AVXSLOW-LABEL: combine_16_byte_loads_double:
; AVXSLOW: # BB#0:
; AVXSLOW-NEXT: vmovupd 144(%rdi), %xmm1
; AVXSLOW-NEXT: vinsertf128 $1, 160(%rdi), %ymm1, %ymm1
; AVXSLOW-NEXT: vaddpd %ymm0, %ymm1, %ymm0
; AVXSLOW-NEXT: retq
;
; AVXFAST-LABEL: combine_16_byte_loads_double:
; AVXFAST: # BB#0:
; AVXFAST-NEXT: vaddpd 144(%rdi), %ymm0, %ymm0
; AVXFAST-NEXT: retq
;
; AVX2-LABEL: combine_16_byte_loads_double:
; AVX2: # BB#0:
; AVX2-NEXT: vaddpd 144(%rdi), %ymm0, %ymm0
; AVX2-NEXT: retq
[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
%ptr1 = getelementptr inbounds <2 x double>, <2 x double>* %ptr, i64 9
%ptr2 = getelementptr inbounds <2 x double>, <2 x double>* %ptr, i64 10
[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
%v1 = load <2 x double>, <2 x double>* %ptr1, align 1
%v2 = load <2 x double>, <2 x double>* %ptr2, align 1
combine consecutive subvector 16-byte loads into one 32-byte load This is a fix for PR21709 ( http://llvm.org/bugs/show_bug.cgi?id=21709 ). When we have 2 consecutive 16-byte loads that are merged into one 32-byte vector, we can use a single 32-byte load instead. But we don't do this for SandyBridge / IvyBridge because they have slower 32-byte memops. We also don't bother using 32-byte *integer* loads on a machine that only has AVX1 (btver2) because those operands would have to be split in half anyway since there is no support for 32-byte integer math ops. Differential Revision: http://reviews.llvm.org/D6492 llvm-svn: 224344
2014-12-17 00:30:01 +08:00
%v3 = shufflevector <2 x double> %v1, <2 x double> %v2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
%v4 = fadd <4 x double> %v3, %x
ret <4 x double> %v4
}