SIMD Int64x2 code fails to build in optimized mode.

[juj]

Building Int64x2 utilizing code fails under optimization, when PNaCl ExpandI64.cpp attempts to operate on the vectors. Debug builds work ok.

As a result, the following SSE2 Int64x2 functions are currently not available in optimized builds and they are skipped in test_sse2_full:

_mm_add_epi64
_mm_mul_epu32
_mm_sub_epi64
_mm_cvtsd_si64
_mm_cvtsi128_si64
_mm_cvtsi64_sd
_mm_cvtsi64_si128
_mm_cvttsd_si64
_mm_move_epi64
_mm_sll_epi64
_mm_slli_epi64
_mm_srl_epi64
_mm_srli_epi64
_mm_storel_epi64
_mm_stream_si64
_mm_unpackhi_epi64
_mm_unpacklo_epi64

As a small test case, one can attempt to build the following with -O1 or higher:

#include <emmintrin.h>
#include <stdio.h>
#include <stdint.h>

__m128i set_64x2(int64_t a, int64_t b) { union { int64_t x[2]; __m128i m; } u; u.x[0] = a; u.x[1] = b; return u.m; }
int64_t get_64x2_lo(__m128i m) { union { int64_t x[2]; __m128i m; } u; u.m = m; return u.x[0]; }
int64_t get_64x2_hi(__m128i m) { union { int64_t x[2]; __m128i m; } u; u.m = m; return u.x[1]; }

int main()
{
    __m128i a = set_64x2(1, 2);
    __m128i b = set_64x2(3, 4);
    __m128i c = _mm_add_epi64(a, b);
    printf("%lld %lld\n", get_64x2_lo(c), get_64x2_hi(c));
}

which fails on LLVM ERROR: I->getType() == I->getOperand(0)->getType() here https://github.com/kripken/emscripten-fastcomp/blob/master/lib/Transforms/NaCl/ExpandI64.cpp#L831.

[kripken]

We end up with %5 = bitcast i64 %a to <2 x i32>. This can be handled by emscripten-core/emscripten-fastcomp@ce4b824

However, @sunfishcode , is it normal to have vector types like this? I thought our backend reports to LLVM that it can only support certain kinds.

Btw, after that patch, the testcase then fails on extractelement <2 x i64>. Same question about that type, should we be seeing that?

[juj]

Checked this again on current incoming, and it is still persists.

[kripken]

Note that the testcase here passes with -O2, but fails on -O1.

[kripken]

@sunfishcode: this problem happens due to SROA isVectorPromotionViable. There doesn't seem to be a way to tell it not to do this.

It seems that if integer promotion fails, it tries vector promotion. I wonder if vector promotion is ever worth it for us? Do you know if this enables other optimizations, or if we could safely just disable vector promotion in SROA?

[sunfishcode]

My sense is that we can safely disable this. SIMD.js doesn't have Int64x2 anyway, so we can revisit this issue when that changes.

[kripken]

Fixed on incoming.

[vilie]

The test case presented by @juj still fails on 1.36.0.

[kripken]

It fails for me with -O1, but provides the hint possible hint: optimize with -O0 or -O2+, and not -O1, and -O2 does indeed fix it.

[vrabaud]

Hi, I get something similar with a different integer type.

Unsupported:   %516 = bitcast <8 x i16> %481 to i128
LLVM ERROR: BitCast Instruction not yet supported for integer types larger than 64 bits

Does the same section need to be patched to also handle in16 and int8 or am I misunderstanding the issue ? Thx

[juj]

@vrabaud: would you have a test case by any chance?

[skal65535]

Hi,
Here's a repro script. I get the error:

Unsupported:   %32 = bitcast <16 x i8> %6 to i128
LLVM ERROR: BitCast Instruction not yet supported for integer types larger than 64 bits

I could not shrink the example further.

crash.txt

skal/

Note that if i move the function func() inside the main.c file, i get a slightly different error:

Unsupported:   %expanded4 = bitcast <16 x i8> <i8 extractelement (<16 x i8> bitcast (<4 x i32> <i32 3, i32 0, i32 undef, i32 undef> to <16 x i8>), i32 0), i8 extractelement (<16 x i8> bitcast (<4 x i32> <i32 3, i32 0, i32 undef, i32 undef> to <16 x i8>), i32 0), i8 extractelement (<16 x i8> bitcast (<4 x i32> <i32 3, i32 0, i32 undef, i32 undef> to <16 x i8>), i32 1), i8 extractelement (<16 x i8> bitcast (<4 x i32> <i32 3, i32 0, i32 undef, i32 undef> to <16 x i8>), i32 1), i8 extractelement (<16 x i8> bitcast (<4 x i32> <i32 3, i32 0, i32 undef, i32 undef> to <16 x i8>), i32 2), i8 extractelement (<16 x i8> bitcast (<4 x i32> <i32 3, i32 0, i32 undef, i32 undef> to <16 x i8>), i32 2), i8 extractelement (<16 x i8> bitcast (<4 x i32> <i32 3, i32 0, i32 undef, i32 undef> to <16 x i8>), i32 3), i8 extractelement (<16 x i8> bitcast (<4 x i32> <i32 3, i32 0, i32 undef, i32 undef> to <16 x i8>), i32 3), i8 extractelement (<16 x i8> bitcast (<4 x i32> <i32 3, i32 0, i32 undef, i32 undef> to <16 x i8>), i32 4), i8 extractelement (<16 x i8> bitcast (<4 x i32> <i32 3, i32 0, i32 undef, i32 undef> to <16 x i8>), i32 4), i8 extractelement (<16 x i8> bitcast (<4 x i32> <i32 3, i32 0, i32 undef, i32 undef> to <16 x i8>), i32 5), i8 extractelement (<16 x i8> bitcast (<4 x i32> <i32 3, i32 0, i32 undef, i32 undef> to <16 x i8>), i32 5), i8 extractelement (<16 x i8> bitcast (<4 x i32> <i32 3, i32 0, i32 undef, i32 undef> to <16 x i8>), i32 6), i8 extractelement (<16 x i8> bitcast (<4 x i32> <i32 3, i32 0, i32 undef, i32 undef> to <16 x i8>), i32 6), i8 extractelement (<16 x i8> bitcast (<4 x i32> <i32 3, i32 0, i32 undef, i32 undef> to <16 x i8>), i32 7), i8 extractelement (<16 x i8> bitcast (<4 x i32> <i32 3, i32 0, i32 undef, i32 undef> to <16 x i8>), i32 7)> to i128
LLVM ERROR: BitCast Instruction not yet supported for integer types larger than 64 bits

[skal65535]

Hi again,

actually, here's a single-file version that produces the same error as the original:

#include <emmintrin.h>
int main(int argc,const char** argv) {  
  __m128i ref0 = _mm_cvtsi32_si128(argv[0][0]);
  ref0 = _mm_unpacklo_epi8(ref0, ref0);
  ref0 = _mm_packus_epi16(ref0, ref0);
  return _mm_cvtsi128_si32(ref0);
}

I compile with: emcc -o main.js main.c -O3 -msse2
The error doesn't occur with -O1 or -Os, note.

hope it helps,
skal/