ArgumentPromotion creates an illegal type argument for AArch64 SVE VLS #69147
Labels
Milestone
Comments
|
@llvm/issue-subscribers-backend-aarch64 Author: KAWASHIMA Takahiro (kawashima-fj)
# ArgumentPromotion creates an illegal type argument for AArch64 SVE VLS
This issue is derived from #63025. ProblemCompiling the following ACLE program with #include <arm_sve.h>
typedef svfloat32_t svfloat32_256_t __attribute__((arm_sve_vector_bits(256)));
__attribute__((noinline))
static void callee(svfloat32_256_t *src, svfloat32_256_t *dst) {
*dst = *src;
}
void caller(svfloat32_256_t *src, svfloat32_256_t *dst) {
callee(src, dst);
}$ clang -S --target=aarch64-unknwon-linux-gnu --sysroot=/opt/aarch64-none-linux-gnu/libc -O3 -march=armv8.2-a+sve -msve-vector-bits=256 -fno-crash-diagnostics test.c
clang: /src/llvm/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp:7014: void analyzeCallOperands(const llvm::AArch64TargetLowering&, const llvm::AArch64Subtarget*, const llvm::TargetLowering::CallLoweringInfo&, llvm::CCState&): Assertion `!Res && "Call operand has unhandled type"' failed.
PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/ and include the crash backtrace, preprocessed source, and associated run script.
Stack dump:
0. Program arguments: clang -S --target=aarch64-unknwon-linux-gnu --sysroot=/opt/aarch64-none-linux-gnu/libc -O3 -march=armv8.2-a+sve -msve-vector-bits=256 -fno-crash-diagnostics test.c
1. <eof> parser at end of file
2. Code generation
3. Running pass 'Function Pass Manager' on module 'test.c'.
4. Running pass 'AArch64 Instruction Selection' on function '@<!-- -->caller'
#<!-- -->0 0x00007f41724fa010 llvm::sys::PrintStackTrace(llvm::raw_ostream&, int) (/opt/llvm/bin/../lib/libLLVMSupport.so.18git+0x248010)
#<!-- -->1 0x00007f41724f741f llvm::sys::RunSignalHandlers() (/opt/llvm/bin/../lib/libLLVMSupport.so.18git+0x24541f)
#<!-- -->2 0x00007f41723fba68 CrashRecoverySignalHandler(int) CrashRecoveryContext.cpp:0:0
#<!-- -->3 0x00007f4171e92520 (/lib/x86_64-linux-gnu/libc.so.6+0x42520)
#<!-- -->4 0x00007f4171ee69fc __pthread_kill_implementation ./nptl/pthread_kill.c:44:76
#<!-- -->5 0x00007f4171ee69fc __pthread_kill_internal ./nptl/pthread_kill.c:78:10
#<!-- -->6 0x00007f4171ee69fc pthread_kill ./nptl/pthread_kill.c:89:10
#<!-- -->7 0x00007f4171e92476 gsignal ./signal/../sysdeps/posix/raise.c:27:6
#<!-- -->8 0x00007f4171e787f3 abort ./stdlib/abort.c:81:7
#<!-- -->9 0x00007f4171e7871b _nl_load_domain ./intl/loadmsgcat.c:1177:9
#<!-- -->10 0x00007f4171e89e96 (/lib/x86_64-linux-gnu/libc.so.6+0x39e96)
#<!-- -->11 0x00007f417730dbbf analyzeCallOperands(llvm::AArch64TargetLowering const&, llvm::AArch64Subtarget const*, llvm::TargetLowering::CallLoweringInfo const&, llvm::CCState&) AArch64ISelLowering.cpp:0:0
#<!-- -->12 0x00007f417739f9aa llvm::AArch64TargetLowering::LowerCall(llvm::TargetLowering::CallLoweringInfo&, llvm::SmallVectorImpl<llvm::SDValue>&) const (/opt/llvm/bin/../lib/libLLVMAArch64CodeGen.so.18git+0x6a49aa)
#<!-- -->13 0x00007f41717eb86b llvm::TargetLowering::LowerCallTo(llvm::TargetLowering::CallLoweringInfo&) const (/opt/llvm/bin/../lib/../lib/libLLVMSelectionDAG.so.18git+0x35686b)
#<!-- -->14 0x00007f41717f8686 llvm::SelectionDAGBuilder::lowerInvokable(llvm::TargetLowering::CallLoweringInfo&, llvm::BasicBlock const*) (/opt/llvm/bin/../lib/../lib/libLLVMSelectionDAG.so.18git+0x363686)
#<!-- -->15 0x00007f4171818e5e llvm::SelectionDAGBuilder::LowerCallTo(llvm::CallBase const&, llvm::SDValue, bool, bool, llvm::BasicBlock const*) (/opt/llvm/bin/../lib/../lib/libLLVMSelectionDAG.so.18git+0x383e5e)
#<!-- -->16 0x00007f4171805ddd llvm::SelectionDAGBuilder::visitCall(llvm::CallInst const&) (/opt/llvm/bin/../lib/../lib/libLLVMSelectionDAG.so.18git+0x370ddd)
#<!-- -->17 0x00007f4171839a9e llvm::SelectionDAGBuilder::visit(llvm::Instruction const&) (/opt/llvm/bin/../lib/../lib/libLLVMSelectionDAG.so.18git+0x3a4a9e)
#<!-- -->18 0x00007f41718ca5b8 llvm::SelectionDAGISel::SelectBasicBlock(llvm::ilist_iterator<llvm::ilist_detail::node_options<llvm::Instruction, true, false, void>, false, true>, llvm::ilist_iterator<llvm::ilist_detail::node_options<llvm::Instruction, true, false, void>, false, true>, bool&) (/opt/llvm/bin/../lib/../lib/libLLVMSelectionDAG.so.18git+0x4355b8)
#<!-- -->19 0x00007f41718cb2a8 llvm::SelectionDAGISel::SelectAllBasicBlocks(llvm::Function const&) (/opt/llvm/bin/../lib/../lib/libLLVMSelectionDAG.so.18git+0x4362a8)
#<!-- -->20 0x00007f41718cd114 llvm::SelectionDAGISel::runOnMachineFunction(llvm::MachineFunction&) (.part.0) SelectionDAGISel.cpp:0:0
#<!-- -->21 0x00007f4175a9bf07 llvm::MachineFunctionPass::runOnFunction(llvm::Function&) (.part.0) MachineFunctionPass.cpp:0:0
#<!-- -->22 0x00007f417312488e llvm::FPPassManager::runOnFunction(llvm::Function&) (/opt/llvm/bin/../lib/libLLVMCore.so.18git+0x49888e)
#<!-- -->23 0x00007f4173124ad9 llvm::FPPassManager::runOnModule(llvm::Module&) (/opt/llvm/bin/../lib/libLLVMCore.so.18git+0x498ad9)
#<!-- -->24 0x00007f4173125415 llvm::legacy::PassManagerImpl::run(llvm::Module&) (/opt/llvm/bin/../lib/libLLVMCore.so.18git+0x499415)
#<!-- -->25 0x00007f41761bb078 clang::EmitBackendOutput(clang::DiagnosticsEngine&, clang::HeaderSearchOptions const&, clang::CodeGenOptions const&, clang::TargetOptions const&, clang::LangOptions const&, llvm::StringRef, llvm::Module*, clang::BackendAction, llvm::IntrusiveRefCntPtr<llvm::vfs::FileSystem>, std::unique_ptr<llvm::raw_pwrite_stream, std::default_delete<llvm::raw_pwrite_stream>>) (/opt/llvm/bin/../lib/libclangCodeGen.so.18git+0x2d6078)
#<!-- -->26 0x00007f4176668c3a clang::BackendConsumer::HandleTranslationUnit(clang::ASTContext&) CodeGenAction.cpp:0:0
#<!-- -->27 0x00007f416f7515f9 clang::ParseAST(clang::Sema&, bool, bool) (/opt/llvm/bin/../lib/../lib/libclangParse.so.18git+0x685f9)
#<!-- -->28 0x00007f4174cbdca9 clang::FrontendAction::Execute() (/opt/llvm/bin/../lib/libclangFrontend.so.18git+0x199ca9)
#<!-- -->29 0x00007f4174c2eb15 clang::CompilerInstance::ExecuteAction(clang::FrontendAction&) (/opt/llvm/bin/../lib/libclangFrontend.so.18git+0x10ab15)
#<!-- -->30 0x00007f4176bb4e75 clang::ExecuteCompilerInvocation(clang::CompilerInstance*) (/opt/llvm/bin/../lib/libclangFrontendTool.so.18git+0x5e75)
#<!-- -->31 0x00005576f0c83e59 cc1_main(llvm::ArrayRef<char const*>, char const*, void*) (/opt/llvm/bin/clang-18+0x19e59)
#<!-- -->32 0x00005576f0c7b693 ExecuteCC1Tool(llvm::SmallVectorImpl<char const*>&, llvm::ToolContext const&) driver.cpp:0:0
#<!-- -->33 0x00007f417490c34d void llvm::function_ref<void ()>::callback_fn<clang::driver::CC1Command::Execute(llvm::ArrayRef<std::optional<llvm::StringRef>>, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char>>*, bool*) const::'lambda'()>(long) Job.cpp:0:0
#<!-- -->34 0x00007f41723fbf30 llvm::CrashRecoveryContext::RunSafely(llvm::function_ref<void ()>) (/opt/llvm/bin/../lib/libLLVMSupport.so.18git+0x149f30)
#<!-- -->35 0x00007f417490cbce clang::driver::CC1Command::Execute(llvm::ArrayRef<std::optional<llvm::StringRef>>, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char>>*, bool*) const (.part.0) Job.cpp:0:0
#<!-- -->36 0x00007f41748cd29a clang::driver::Compilation::ExecuteCommand(clang::driver::Command const&, clang::driver::Command const*&, bool) const (/opt/llvm/bin/../lib/libclangDriver.so.18git+0xff29a)
#<!-- -->37 0x00007f41748cdd6d clang::driver::Compilation::ExecuteJobs(clang::driver::JobList const&, llvm::SmallVectorImpl<std::pair<int, clang::driver::Command const*>>&, bool) const (/opt/llvm/bin/../lib/libclangDriver.so.18git+0xffd6d)
#<!-- -->38 0x00007f41748dcd34 clang::driver::Driver::ExecuteCompilation(clang::driver::Compilation&, llvm::SmallVectorImpl<std::pair<int, clang::driver::Command const*>>&) (/opt/llvm/bin/../lib/libclangDriver.so.18git+0x10ed34)
#<!-- -->39 0x00005576f0c80bd0 clang_main(int, char**, llvm::ToolContext const&) (/opt/llvm/bin/clang-18+0x16bd0)
#<!-- -->40 0x00005576f0c911a3 main (/opt/llvm/bin/clang-18+0x271a3)
#<!-- -->41 0x00007f4171e79d90 __libc_start_call_main ./csu/../sysdeps/nptl/libc_start_call_main.h:58:16
#<!-- -->42 0x00007f4171e79e40 call_init ./csu/../csu/libc-start.c:128:20
#<!-- -->43 0x00007f4171e79e40 __libc_start_main ./csu/../csu/libc-start.c:379:5
#<!-- -->44 0x00005576f0c7a2d5 _start (/opt/llvm/bin/clang-18+0x102d5)
clang: error: clang frontend command failed with exit code 134 (use -v to see invocation)EnvironmentThe latest CauseAs described in #63025, the root cause is that ArgumentPromotion promotes an Before ArgumentPromotionPass: ; *** IR Dump Before ArgumentPromotionPass on (callee) ***
; ModuleID = 'test.c'
source_filename = "test.c"
target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
target triple = "aarch64-unknwon-linux-gnu"
; Function Attrs: nounwind uwtable vscale_range(2,2)
define dso_local void @<!-- -->caller(ptr noundef %src, ptr noundef %dst) local_unnamed_addr #<!-- -->0 {
entry:
call fastcc void @<!-- -->callee(ptr noundef %src, ptr noundef %dst)
ret void
}
; Function Attrs: noinline nounwind uwtable vscale_range(2,2)
define internal fastcc void @<!-- -->callee(ptr nocapture noundef readonly %src, ptr nocapture noundef writeonly %dst) unnamed_addr #<!-- -->1 {
entry:
%0 = load <8 x float>, ptr %src, align 16, !tbaa !6
store <8 x float> %0, ptr %dst, align 16, !tbaa !6
ret void
}
attributes #<!-- -->0 = { nounwind uwtable vscale_range(2,2) "frame-pointer"="non-leaf" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="generic" "target-features"="+crc,+fp-armv8,+fullfp16,+lse,+neon,+outline-atomics,+ras,+rdm,+sve,+v8.1a,+v8.2a,+v8a,-fmv" }
attributes #<!-- -->1 = { noinline nounwind uwtable vscale_range(2,2) "frame-pointer"="non-leaf" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="generic" "target-features"="+crc,+fp-armv8,+fullfp16,+lse,+neon,+outline-atomics,+ras,+rdm,+sve,+v8.1a,+v8.2a,+v8a,-fmv" }
!llvm.module.flags = !{!0, !1, !2, !3, !4}
!llvm.ident = !{!5}
!0 = !{i32 1, !"wchar_size", i32 4}
!1 = !{i32 8, !"PIC Level", i32 2}
!2 = !{i32 7, !"PIE Level", i32 2}
!3 = !{i32 7, !"uwtable", i32 2}
!4 = !{i32 7, !"frame-pointer", i32 1}
!5 = !{!"clang version 18.0.0"}
!6 = !{!7, !7, i64 0}
!7 = !{!"omnipotent char", !8, i64 0}
!8 = !{!"Simple C/C++ TBAA"}After ArgumentPromotionPass: ; *** IR Dump After ArgumentPromotionPass on (callee) ***
; ModuleID = 'test.c'
source_filename = "test.c"
target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
target triple = "aarch64-unknwon-linux-gnu"
; Function Attrs: nounwind uwtable vscale_range(2,2)
define dso_local void @<!-- -->caller(ptr noundef %src, ptr noundef %dst) local_unnamed_addr #<!-- -->0 {
entry:
%src.val = load <8 x float>, ptr %src, align 16, !tbaa !6
call fastcc void @<!-- -->callee(<8 x float> %src.val, ptr noundef %dst)
ret void
}
; Function Attrs: noinline nounwind uwtable vscale_range(2,2)
define internal fastcc void @<!-- -->callee(<8 x float> %src.0.val, ptr nocapture noundef writeonly %dst) unnamed_addr #<!-- -->1 {
entry:
store <8 x float> %src.0.val, ptr %dst, align 16, !tbaa !6
ret void
}
attributes #<!-- -->0 = { nounwind uwtable vscale_range(2,2) "frame-pointer"="non-leaf" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="generic" "target-features"="+crc,+fp-armv8,+fullfp16,+lse,+neon,+outline-atomics,+ras,+rdm,+sve,+v8.1a,+v8.2a,+v8a,-fmv" }
attributes #<!-- -->1 = { noinline nounwind uwtable vscale_range(2,2) "frame-pointer"="non-leaf" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="generic" "target-features"="+crc,+fp-armv8,+fullfp16,+lse,+neon,+outline-atomics,+ras,+rdm,+sve,+v8.1a,+v8.2a,+v8a,-fmv" }
!llvm.module.flags = !{!0, !1, !2, !3, !4}
!llvm.ident = !{!5}
!0 = !{i32 1, !"wchar_size", i32 4}
!1 = !{i32 8, !"PIC Level", i32 2}
!2 = !{i32 7, !"PIE Level", i32 2}
!3 = !{i32 7, !"uwtable", i32 2}
!4 = !{i32 7, !"frame-pointer", i32 1}
!5 = !{!"clang version 18.0.0"}
!6 = !{!7, !7, i64 0}
!7 = !{!"omnipotent char", !8, i64 0}
!8 = !{!"Simple C/C++ TBAA"}Diff: @@ -1,4 +1,4 @@
-; *** IR Dump Before ArgumentPromotionPass on (callee) ***
+; *** IR Dump After ArgumentPromotionPass on (callee) ***
; ModuleID = 'test.c'
source_filename = "test.c"
target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
@@ -7,15 +7,15 @@
; Function Attrs: nounwind uwtable vscale_range(2,2)
define dso_local void @<!-- -->caller(ptr noundef %src, ptr noundef %dst) local_unnamed_addr #<!-- -->0 {
entry:
- call fastcc void @<!-- -->callee(ptr noundef %src, ptr noundef %dst)
+ %src.val = load <8 x float>, ptr %src, align 16, !tbaa !6
+ call fastcc void @<!-- -->callee(<8 x float> %src.val, ptr noundef %dst)
ret void
}
; Function Attrs: noinline nounwind uwtable vscale_range(2,2)
-define internal fastcc void @<!-- -->callee(ptr nocapture noundef readonly %src, ptr nocapture noundef writeonly %dst) unnamed_addr #<!-- -->1 {
+define internal fastcc void @<!-- -->callee(<8 x float> %src.0.val, ptr nocapture noundef writeonly %dst) unnamed_addr #<!-- -->1 {
entry:
- %0 = load <8 x float>, ptr %src, align 16, !tbaa !6
- store <8 x float> %0, ptr %dst, align 16, !tbaa !6
+ store <8 x float> %src.0.val, ptr %dst, align 16, !tbaa !6
ret void
}
Possible FixGiven that the AArch64 backend won't support vector arguments which are VLS larger then 128 bits, we have two options.
ArgumentPromotion has a hook to check ABI using Are there any comments? If no objection, I'll post a PR. |
|
I'm an expert in backend. But I think it may be related to #68953 where SLPVectorizer crashed. Could you take a look at it? |
|
@XChy I took a look at #68953 a bit (see #68953 (comment)) but unfortunately I think it is not related to this issue. |
That's OK. Thanks for your reduction! |
|
I have no objections. Implementing @kawashima-fj: You mentioned creating a PR. Do you have a timeframe? I'd quite like to get this fixed so there's time to back port it into LLVM-17. Does that match your timeline or should I find somebody else to run with a fix? |
|
@paulwalker-arm The patch is almost ready. I'm now running tests. I'll submit a PR today (in Japan timezone) hopefully. |
kawashima-fj
added a commit
to kawashima-fj/pr-llvm-project
that referenced
this issue
Oct 26, 2023
This patch prevents argument promotion from promoting pointers to fixed-length vector types larger than 128 bits like `<8 x float>` into the values of the pointees. Such vector types are used for SVE VLS but there is no ABI for SVE VLS arguments and the backend cannot lower such value arguments. Fixes llvm#69147
kawashima-fj
added a commit
that referenced
this issue
Oct 26, 2023
…70034) This patch prevents argument promotion from promoting pointers to fixed-length vector types larger than 128 bits like `<8 x float>` into the values of the pointees. Such vector types are used for SVE VLS but there is no ABI for SVE VLS arguments and the backend cannot lower such value arguments. Fixes #69147
|
/cherry-pick 926173c |
|
/branch llvm/llvm-project-release-prs/issue69147 |
|
/pull-request llvm/llvm-project-release-prs#749 |
zahiraam
pushed a commit
to zahiraam/llvm-project
that referenced
this issue
Oct 26, 2023
…lvm#70034) This patch prevents argument promotion from promoting pointers to fixed-length vector types larger than 128 bits like `<8 x float>` into the values of the pointees. Such vector types are used for SVE VLS but there is no ABI for SVE VLS arguments and the backend cannot lower such value arguments. Fixes llvm#69147
tru
pushed a commit
that referenced
this issue
Oct 27, 2023
…70034) This patch prevents argument promotion from promoting pointers to fixed-length vector types larger than 128 bits like `<8 x float>` into the values of the pointees. Such vector types are used for SVE VLS but there is no ABI for SVE VLS arguments and the backend cannot lower such value arguments. Fixes #69147 (cherry picked from commit 926173c)
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
This issue is derived from #63025.
Problem
Compiling the following ACLE program with
-O3 -march=armv8.2-a+sve -msve-vector-bits=256for Linux/AArch64 crashes. This problem is specific to AArch64 SVE VLS (a.k.a. fixed-length).Environment
The latest
mainbranch (commit 97c9f9a) and at least 13.0.1, 14.0.6, 15.0.7, 16.0.6, 17.0.1 have this problem.Cause
As described in #63025, the root cause is that ArgumentPromotion promotes an
<8 x float>pointer argument to an<8 x float>value argument. There is no ABI for SVE VLS so we cannot legalize vector arguments which are larger then 128 bits.Before ArgumentPromotionPass:
After ArgumentPromotionPass:
Diff:
Possible Fix
Given that the AArch64 backend won't support vector arguments which are VLS larger then 128 bits, we have two options.
ArgumentPromotion has a hook to check ABI using
TargetTransformInfo. X86 and PPC use this hook to prevent illegal argument promotions. So I'd like to take a similar approach for 1. above.Are there any comments? If no objection, I'll post a PR.
The text was updated successfully, but these errors were encountered: