_mm256_cmpeq_epi8 returns a different result when -mattr=+avx2

[MrcSnm]

import std.stdio;
import std.format;
import inteli.avx2intrin;

void main()
{
    string a = `00000"00000""""""0000000000"""00`;
    writeln = a.length;
    byte32 chunk = cast(byte32)_mm256_loadu_si256(cast(__m256i*)a.ptr);
    byte32 quoteMask = [0x22, 0x22,0x22,0x22,0x22,0x22,0x22,0x22,0x22,0x22,0x22,0x22,0x22,0x22,0x22,0x22,0x22, 0x22,0x22,0x22,0x22,0x22,0x22,0x22,0x22,0x22,0x22,0x22,0x22,0x22,0x22,0x22,];
    // ulong chunk = *cast(ulong*)a.ptr;

    // enum quoteMask = 0x22_22_22_22_22_22_22_22;

    // char[2] b = "1\"";

    // enum quoteMask = 0x22_22;

    // ushort chunk = *cast(ushort*)b.ptr;
    // enum minus = 0x01_01;

    // short quoteEq = cast(short)((chunk ^ quoteMask) - minus);
    // import core.bitop;
    // writeln = bsf(quoteEq & 0x8080) >> 3;


    // writeln = format("%b", quoteEq & 0x8080);
    import core.bitop;
// byte32 res = cast(byte32)_mm256_cmpeq_epi8(cast(long4)chunk,  cast(long4)quoteMask); //results in 0... with +avx2
    byte32 res = chunk == quoteMask; //only compares first 16 bytes if +avx2
    int mask = _mm256_movemask_epi8(cast(long4)(res));
    writeln = res;
    writeln = format("%b", mask);

    writeln = bsf(mask);
}

PS: Using the == operator also results in a 3rd result. Basically, for byte32, it only compares the first 16 bytes

[p0nce]

With what compiler? Cannot repro in unittest, ldc 1.41

[MrcSnm]

Can you also test in LDC 1.41 O3?

[p0nce]

Ah I see, I need to uncomment one line in your repro script. Please provide a crashing test next time.

The issue right now is #153 , the test suite never gets to this unittest.

[p0nce]

Not the same issue as #153 , but #153 masks this on this machine, I'll would need to try on x86 machine.

[p0nce]

OK, still no repro on LDC 1.41, using -mattr=+avx2, -O3 or -O2, x86 or x86_64

[p0nce]

finally repro
To be reproduced, really need to be in an external program and not an unittest.
LDC 1.41 + any -O level + -mattr=+avx or +avx2

[p0nce]

Reproduced without intel-intrinsics or DUB, just ldc2 calls.
Basically what happens is that a library is compiled without -mattr=+avx registers, and linked with something compiled with -mattr=+avx. thus differing ABIs.
_mm256_loadu_si256 returns a truncated register.
repro.zip

[p0nce]

Original repro case with DUB, type dub to repro
repro2.zip
--combined is a workaround.

[MrcSnm]

That is strange. Why would dub fail if dflags are passed to all of them?

[p0nce]

You can check with -v

[MrcSnm]

I don't recall why that happened first. I was testing with redub and that never happened. It only happened when I started using dub which I don't use. But since the issue is the mismatching compiler flags, I'll take notes on that