"Expected SME to be available" assert fails on memcpy

[arsenm]

Trying to produce the simplest possible testcase that emits __arm_sc_memcpy hits an assertion: https://godbolt.org/z/89GTdMYf5

; RUN: llc -mtriple=arm64-apple-darwin < %s | FileCheck %s

@dst = global [512 x i8] zeroinitializer, align 1
@src = global [512 x i8] zeroinitializer, align 1

define void @se_memcpy(i64 noundef %n) "aarch64_pstate_sm_enabled" nounwind {
entry:
  tail call void @llvm.memcpy.p0.p0.i64(ptr align 1 @dst, ptr nonnull align 1 @src, i64 %n, i1 false)
  ret void
}

Assertion failed: ((!IsStreaming || I->hasSME()) && "Expected SME to be available"), function getSubtargetImpl, file AArch64TargetMachine.cpp, line 482.
PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/ and include the crash backtrace.
Stack dump:
0.	Program arguments: llc -o - /tmp/assert-aarch64.ll
1.	Running pass 'Pre-ISel Intrinsic Lowering' on module '/tmp/assert-aarch64.ll'.
 #0 0x00000001072fcf04 llvm::sys::PrintStackTrace(llvm::raw_ostream&, int) (/usr/local/bin/llc+0x1024dcf04)
 #1 0x00000001072fb430 llvm::sys::RunSignalHandlers() (/usr/local/bin/llc+0x1024db430)
 #2 0x00000001072fd598 SignalHandler(int, __siginfo*, void*) (/usr/local/bin/llc+0x1024dd598)
 #3 0x0000000198f4ade4 (/usr/lib/system/libsystem_platform.dylib+0x180482de4)
 #4 0x0000000198f13f70 (/usr/lib/system/libsystem_pthread.dylib+0x18044bf70)
 #5 0x0000000198e20908 (/usr/lib/system/libsystem_c.dylib+0x180358908)
 #6 0x0000000198e1fc1c (/usr/lib/system/libsystem_c.dylib+0x180357c1c)
 #7 0x0000000107f1048c llvm::AArch64TargetMachine::getSubtargetImpl(llvm::Function const&) const (.cold.3) (/usr/local/bin/llc+0x1030f048c)
 #8 0x00000001051d1088 llvm::AArch64TargetMachine::getSubtargetImpl(llvm::Function const&) const (/usr/local/bin/llc+0x1003b1088)
 #9 0x00000001051d1690 llvm::AArch64TargetMachine::getTargetTransformInfo(llvm::Function const&) const (/usr/local/bin/llc+0x1003b1690)
#10 0x0000000106577ab0 llvm::TargetTransformInfoWrapperPass::getTTI(llvm::Function const&) (/usr/local/bin/llc+0x101757ab0)
#11 0x000000010690f2ac (anonymous namespace)::PreISelIntrinsicLowering::lowerIntrinsics(llvm::Module&) const (/usr/local/bin/llc+0x101aef2ac)
#12 0x0000000106910720 (anonymous namespace)::PreISelIntrinsicLoweringLegacyPass::runOnModule(llvm::Module&) (/usr/local/bin/llc+0x101af0720)

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Matt Arsenault (arsenm)

Trying to produce the simplest possible testcase that emits __arm_sc_memcpy hits an assertion: https://godbolt.org/z/89GTdMYf5

; RUN: llc -mtriple=arm64-apple-darwin < %s | FileCheck %s

@<!-- -->dst = global [512 x i8] zeroinitializer, align 1
@<!-- -->src = global [512 x i8] zeroinitializer, align 1

define void @<!-- -->se_memcpy(i64 noundef %n) "aarch64_pstate_sm_enabled" nounwind {
entry:
  tail call void @<!-- -->llvm.memcpy.p0.p0.i64(ptr align 1 @<!-- -->dst, ptr nonnull align 1 @<!-- -->src, i64 %n, i1 false)
  ret void
}

Assertion failed: ((!IsStreaming || I->hasSME()) && "Expected SME to be available"), function getSubtargetImpl, file AArch64TargetMachine.cpp, line 482.
PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/ and include the crash backtrace.
Stack dump:
0.	Program arguments: llc -o - /tmp/assert-aarch64.ll
1.	Running pass 'Pre-ISel Intrinsic Lowering' on module '/tmp/assert-aarch64.ll'.
 #<!-- -->0 0x00000001072fcf04 llvm::sys::PrintStackTrace(llvm::raw_ostream&amp;, int) (/usr/local/bin/llc+0x1024dcf04)
 #<!-- -->1 0x00000001072fb430 llvm::sys::RunSignalHandlers() (/usr/local/bin/llc+0x1024db430)
 #<!-- -->2 0x00000001072fd598 SignalHandler(int, __siginfo*, void*) (/usr/local/bin/llc+0x1024dd598)
 #<!-- -->3 0x0000000198f4ade4 (/usr/lib/system/libsystem_platform.dylib+0x180482de4)
 #<!-- -->4 0x0000000198f13f70 (/usr/lib/system/libsystem_pthread.dylib+0x18044bf70)
 #<!-- -->5 0x0000000198e20908 (/usr/lib/system/libsystem_c.dylib+0x180358908)
 #<!-- -->6 0x0000000198e1fc1c (/usr/lib/system/libsystem_c.dylib+0x180357c1c)
 #<!-- -->7 0x0000000107f1048c llvm::AArch64TargetMachine::getSubtargetImpl(llvm::Function const&amp;) const (.cold.3) (/usr/local/bin/llc+0x1030f048c)
 #<!-- -->8 0x00000001051d1088 llvm::AArch64TargetMachine::getSubtargetImpl(llvm::Function const&amp;) const (/usr/local/bin/llc+0x1003b1088)
 #<!-- -->9 0x00000001051d1690 llvm::AArch64TargetMachine::getTargetTransformInfo(llvm::Function const&amp;) const (/usr/local/bin/llc+0x1003b1690)
#<!-- -->10 0x0000000106577ab0 llvm::TargetTransformInfoWrapperPass::getTTI(llvm::Function const&amp;) (/usr/local/bin/llc+0x101757ab0)
#<!-- -->11 0x000000010690f2ac (anonymous namespace)::PreISelIntrinsicLowering::lowerIntrinsics(llvm::Module&amp;) const (/usr/local/bin/llc+0x101aef2ac)
#<!-- -->12 0x0000000106910720 (anonymous namespace)::PreISelIntrinsicLoweringLegacyPass::runOnModule(llvm::Module&amp;) (/usr/local/bin/llc+0x101af0720)

[sdesmalen-arm]

Streaming functions (annotated by aarch64_pstate_sm_enabled can only exist when +sme is available.

[arsenm]

Streaming functions (annotated by aarch64_pstate_sm_enabled can only exist when +sme is available.

There are various notions of what this condition means. In any case this cannot be an assert

[sdesmalen-arm]

Clang makes sure the corresponding C-level __arm_streaming can only be used when +sme is enabled. Otherwise, Clang will give a diagnostic, so at this point in LLVM the assert is asserting that the IR is sound. Are you suggesting this is missing a check in the IR Verifier?

[nikic]

Either an IR verifier check or usage of reportFatalUsageError() in the backend would be appropriate.

[arsenm]

Probably should not be an IR verifier check. Also it would be a lot better to avoid this kind of target-feature dependence. The features get lost easily