kripken
commented
1 year ago cc @tlively - was there some recent LLVM change to SIMD?
Closed mr-c closed 1 year ago
kripken
commented
1 year ago cc @tlively - was there some recent LLVM change to SIMD?
mr-c
commented
1 year ago FYI: this particular code path does not have any SIMD intrinsics, but many compilers might convert the vector operation to SIMD ops
tlively
commented
1 year ago I'm not aware of any relevant LLVM changes, but there must be a codegen regression. I'm working on reproducing and bisecting now.
tlively
commented
1 year ago @mr-c I'm having trouble reproducing the problem at all. Here's my attempt to extract what the test is doing:
#include <assert.h>
#include <math.h>
#include <stdio.h>
typedef long long __i64x2 __attribute__((__vector_size__(16), __aligned__(16)));
typedef double __f64x2 __attribute__((__vector_size__(16), __aligned__(16)));
__attribute__((__noinline__))
void test(__f64x2 a, __f64x2 b) {
__i64x2 m = a < b;
__f64x2 r = (__f64x2)((((__i64x2)a) & m) | (((__i64x2)b) & ~m));
assert(m[0] == 0);
assert(m[1] == 0);
assert(r[0] == -760.30);
assert(isnan(r[1]));
}
int main() {
__f64x2 a = {NAN, 532.24};
__f64x2 b = {-760.30, NAN};
test(a, b);
printf("ok\n");
}I would have expected this to fail with tot, but it doesn't. @mr-c, do you think you could take a stab at extracting a smaller reproducer?
mr-c
commented
1 year ago @tlively Here is the a, b, and expected return value that cause the failure
Some of the relevant typedefs:
typedef float simde_float32
typedef simde_float32 simde__m128 __attribute__((__aligned__((16)))) __attribute__((__vector_size__(16))) __attribute__((__may_alias__));I'm not able to reproduce using those values either. Can you either provide a minimal reproducer or instructions for how to reproduce this with a checkout of SIMDE?
mr-c
commented
1 year ago @tlively
Thanks for your persistence. Here's a recipe
git checkout https://github.com/simd-everywhere/simde.git
cd simde
mkdir -p build
meson --cross-file=docker/cross-files/emscripten.cross setup build
# you may need to edit emscripten.cross to use your local paths
meson test -C build x86/sse/emul/c x86/sse/emul/cpp wasm_simd128/pmax/emul/c wasm_simd128/pmax/emul/cppThanks, I was able to reproduce the error with that recipe and I have a bisection running now. I still may need help getting a minimal reproducer for an upstream bug report, but let's see what the bad LLVM commit is first.
tlively
commented
1 year ago Looks like this was the bad commit: https://reviews.llvm.org/D143581. I've left a comment, but I don't think we'll need a minimal repro because the problem is fairly obvious.
mr-c
commented
1 year ago That's great news, thanks! Do you have enough information to create a test case to prevent a future regression?
sparker-arm
commented
1 year ago I had swapped the 'true' and 'false' result values in my patch... hopefully this is the fix...
Hello, CI for https://github.com/simd-everywhere/simde against tot got a regression
The function at the heart of the now-failing test is , but without
SIMDE_WASM_SIMD128_NATIVEdefined, https://github.com/simd-everywhere/simde/blob/7e70d02dda84bdfb1e97777bdb9622f3f29682b4/simde/x86/sse.h#L2966Here is that function from he
-Epostprocess outputVersion of emscripten/emsdk: Please include the output
emcc -vhereworking tot: sdk-releases-c41ee404c571c6a46e1ccde357a03072706555a5-64bit https://github.com/simd-everywhere/simde/actions/runs/4155968543/jobs/7189425976
broken tot: sdk-releases-51d9739bc0b3221501654ff6a14e76b90e3dd935-64bit https://github.com/simd-everywhere/simde/actions/runs/4183801504/jobs/7248626128
Full link command and output with
-vappended:Contents of
/tmp/emscripten_temp/regression.tar.gz