LLVM assertion failure when casting a SIMD type with -O3

[abrown]

In attempting to build a Wasm SIMD version of libaom using SIMDe, I ran into the following crash when I add the -O3 flag (it does not crash without this):

...
clang: /b/s/w/ir/cache/builder/emscripten-releases/llvm-project/llvm/include/llvm/Support/Casting.h:264: typename cast_retty<X, Y *>::ret_type llvm::cast(Y *) [X = llvm::ConstantInt, Y = llvm::Value]: Assertion `isa<X>(Val) && "cast<Ty>() argument of incompatible type!"' failed.
Stack dump:
0.      Program arguments: /home/abrown/Code/emsdk/upstream/bin/clang -target wasm32-unknown-emscripten -D__EMSCRIPTEN_major__=1 -D__EMSCRIPTEN_minor__=39 -D__EMSCRIPTEN_tiny__=6 -D_LIBCPP_ABI_VERSION=2 -Dunix -D__unix -D__unix__ -Werror=implicit-function-declaration -Xclang -nostdsysteminc -Xclang -isystem/home/abrown/Code/emsdk/upstream/emscripten/system/include/libcxx -Xclang -isystem/home/abrown/Code/emsdk/upstream/emscripten/system/lib/libcxxabi/include -Xclang -isystem/home/abrown/Code/emsdk/upstream/emscripten/system/include/compat -Xclang -isystem/home/abrown/Code/emsdk/upstream/emscripten/system/include -Xclang -isystem/home/abrown/Code/emsdk/upstream/emscripten/system/include/libc -Xclang -isystem/home/abrown/Code/emsdk/upstream/emscripten/system/lib/libc/musl/arch/emscripten -Xclang -isystem/home/abrown/Code/emsdk/upstream/emscripten/system/local/include -Xclang -isystem/home/abrown/.emscripten_cache/wasm-obj/include -DENABLE_SIMDE=1 -DSIMDE_ENABLE_NATIVE_ALIASES=1 -I/home/abrown/Code/aom -I/home/abrown/Code/aom-build-analyzer -I/home/abrown/Code/aom/apps -I/home/abrown/Code/aom/common -I/home/abrown/Code/aom/examples -I/home/abrown/Code/aom/stats -I/home/abrown/Code/aom/third_party/simde -I/home/abrown/Code/aom/third_party/libyuv/include -gline-tables-only -std=c99 -Wall -Wdisabled-optimization -Wextra -Wfloat-conversion -Wimplicit-function-declaration -Wpointer-arith -Wsign-compare -Wstring-conversion -Wtype-limits -Wuninitialized -Wunused -Wvla -Wshorten-64-to-32 -Wshadow -Wundef -D_LARGEFILE_SOURCE -D_FILE_OFFSET_BITS=64 -D_POSIX_SOURCE -O3 -D_POSIX_SOURCE -D_POSIX_SOURCE -fopenmp-simd -DSIMDE_ENABLE_OPENMP -mssse3 -c -DEMSCRIPTEN -msimd128 /home/abrown/Code/aom/aom_dsp/x86/intrapred_ssse3.c -Xclang -isystem/home/abrown/Code/emsdk/upstream/emscripten/system/include/SDL -c -o CMakeFiles/aom_dsp_common_ssse3_intrinsics.dir/aom_dsp/x86/intrapred_ssse3.c.o -mllvm -combiner-global-alias-analysis=false -mllvm -enable-emscripten-sjlj -mllvm -disable-lsr 
1.      <eof> parser at end of file
2.      Per-module optimization passes
3.      Running pass 'Function Pass Manager' on module '/home/abrown/Code/aom/aom_dsp/x86/intrapred_ssse3.c'.
4.      Running pass 'SLP Vectorizer' on function '@aom_paeth_predictor_16x4_ssse3'
...

This failure comes using the following emcc -v:

emcc (Emscripten gcc/clang-like replacement + linker emulating GNU ld) 1.39.6
clang version 10.0.0 (/b/s/w/ir/cache/git/chromium.googlesource.com-external-github.com-llvm-llvm--project 06cfcdcca7de9c88a1e885eff0d0c4c07090ad48)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: /home/abrown/Code/emsdk/upstream/bin
Found candidate GCC installation: /usr/lib/gcc/x86_64-redhat-linux/9
Selected GCC installation: /usr/lib/gcc/x86_64-redhat-linux/9
Candidate multilib: .;@m64
Candidate multilib: 32;@m32
Selected multilib: .;@m64
shared:INFO: (Emscripten: Running sanity checks)

The code being compiled is here and @nemequ reduced that file to a smaller test case:

#include <stddef.h>
#include <stdint.h>
#include <string.h>

typedef int32_t __m128i __attribute__((aligned(16), vector_size(16), __may_alias__));
typedef int16_t __m128i_i16 __attribute__((aligned(16), vector_size(16), __may_alias__));
typedef int8_t __m128i_i8 __attribute__((aligned(16), vector_size(16), __may_alias__));

static __m128i _mm_cvtsi32_si128(int32_t a) { return ((__m128i){a, 0, 0, 0}); }

static __m128i _mm_set1_epi16(int16_t v) {
  __m128i_i16 r16 = {v, v, v, v, v, v, v, v};
  __m128i r32;
  memcpy(&r32, &r16, sizeof(r32));
  return r32;
}

static __m128i _mm_shuffle_epi8(__m128i a, __m128i b) {
  __m128i r;
  __m128i_i8 ta, tb, tr;

  memcpy(&ta, &a, sizeof(ta));
  memcpy(&tb, &b, sizeof(tb));

  for (size_t i = 0; i < (sizeof(tr) / sizeof(tr[0])); i++) {
    tr[i] = ta[tb[i] & 15] & (tb[i] >> 7);
  }

  memcpy(&r, &tr, sizeof(r));
  return r;
}

static void _mm_store_si128(__m128i *mem_addr, __m128i a) {
  *mem_addr = a;
}

static __m128i _mm_add_epi16(__m128i a, __m128i b) {
  return a + b;
}

void aom_paeth_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t stride,
                                    const uint8_t *left) {
  // All fail
  //__m128i l = _mm_cvtsi32_si128(((const uint32_t *)left)[0]);
  //__m128i l = _mm_cvtsi32_si128(*((const uint32_t *)left));
  __m128i l; memcpy(&l, left, sizeof(l));

  __m128i rep = { 0, };

  for (int i = 0; i < 4; ++i) {
    const __m128i l16 = _mm_shuffle_epi8(l, rep);

    // Both fail
    // memcpy(dst, &l16, sizeof(l16));
    _mm_store_si128((__m128i *)dst, l16);

    dst += stride;

    rep = _mm_add_epi16(rep, _mm_set1_epi16(1));
  }
}

I'm not exactly sure what is going on here so feel free to update the title or re-route this to LLVM if it is a bug there; I felt it was best to triage in Emscripten first to see what you all think.

[nemequ]

It's worth noting that it also requires -s SIMD=1. And it happens at -O2 as well, though not at -O1.

[tlively]

Will investigate! Thanks for the report :)

[tlively]

So it looks like this is a bug in the SLP vectorizer pass. I have no idea how that pass works and it is not WebAssembly-specific, so it would be great if you could file an LLVM bug about this.

[tlively]

Also, whoever wants to fix this bug probably doesn't have emscripten installed, so it would be helpful if you uploaded the automatically generated test file and script when you file that bug report.

[nemequ]

Thanks for looking into this!

Reported to LLVM as https://bugs.llvm.org/show_bug.cgi?id=44954

[abrown]

@tlively, looks like a fix was merged. I'm not too familiar with how Emscripten tracks LLVM; when do you think the fix would get downstreamed into Emscripten?

[sbc100]

It should already be available via ./emsdk install tot.. and will be available as latest as soon as we tag a release.

[abrown]

I'm not seeing it fixed when I do ./emsdk install tot && ./emsdk activate tot; I posted a stack trace in the LLVM issue.

[tlively]

@abrown, as mentioned in the LLVM issue, this looks like a different bug. Can you cc me on any new issue you file?

[abrown]

@tlively, in case I didn't do things the right way in BugZilla, here's a link to the new issue: https://bugs.llvm.org/show_bug.cgi?id=45178.

[tlively]

This has been fixed.