[llvm-bugs] [Bug 35982] New: [MMX] POSTRAScheduler rearrange emms and mmx instruction

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=35982

            Bug ID: 35982
           Summary: [MMX] POSTRAScheduler rearrange emms and mmx
                    instruction
           Product: new-bugs
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: new bugs
          Assignee: unassignedbugs at nondot.org
          Reporter: ilia.taraban at intel.com
                CC: llvm-bugs at lists.llvm.org

POSTRAScheduler rearrange emms and mmx instruction, so we receive wrong result:

================= main.c ==============
#include <stdio.h>
#include <x86intrin.h>
float sum(__m64);
int main()
{
    float result;
    __m64 x = (_mm_set_pi32(5, 3));
    result = sum(x);
    printf("5 + 3 = %f\n", result);
    _mm_empty();

    return 0;
}

=======================================
================= nice.c ==============
#include <x86intrin.h>

float sum(__m64 x)
{
    double t;
    int x1, x2;

    x1 = _mm_cvtsi64_si32(x);
    x2 = _mm_cvtsi64_si32(_mm_unpackhi_pi32(x, x));

    _mm_empty();

    t = (float)x1 + (float)x2;

    return t;
}
=======================================

>>> clang -v
clang version 7.0.0 (trunk 322555)
Target: x86_64-unknown-linux-gnu
Thread model: posix
...


>>> clang -m32 -O0 -o nice.exe main.c nice.c
>>> ./nice.exe
5 + 3 = 8.000000

>>> clang -m32 -O2 -o nice.exe main.c nice.c
>>> ./nice.exe
5 + 3 = -nan

>>> clang -m32 -O2 -o nice.exe main.c nice.c -mllvm -opt-bisect-limit=194 && ./nice.exe
...
BISECT: running pass (193) Tail Duplication on function (sum)
BISECT: running pass (194) Machine Copy Propagation Pass on function (sum)
BISECT: NOT running pass (195) Post RA top-down list latency scheduler on
function (sum)
BISECT: NOT running pass (196) Branch Probability Basic Block Placement on
function (sum)
...
5 + 3 = 8

>>> clang -m32 -O2 -o nice.exe main.c nice.c -mllvm -opt-bisect-limit=195 && ./nice.exe
...
BISECT: running pass (193) Tail Duplication on function (sum)
BISECT: running pass (194) Machine Copy Propagation Pass on function (sum)
BISECT: running pass (195) Post RA top-down list latency scheduler on function
(sum)
BISECT: NOT running pass (196) Branch Probability Basic Block Placement on
function (sum)
BISECT: NOT running pass (197) X86 Execution Dependency Fix on function (sum)
...
5 + 3 = -nan

Let's look at ASM before and after POSTRAScheduler:
=============== nice-194.s ============
        cvtsi2ssl       %eax, %xmm0
        movq    8(%esp), %mm0
        emms
        punpckhdq       %mm0, %mm0      # mm0 = mm0[1,1]
        movd    %mm0, %ecx
        cvtsi2ssl       %ecx, %xmm1
=======================================

POSTRAScheduler changed order of operations to:

=============== nice-195.s ============
        movq    8(%esp), %mm0
        punpckhdq       %mm0, %mm0      # mm0 = mm0[1,1]
        movd    %mm0, %ecx
        emms
        cvtsi2ssl       %eax, %xmm0
        cvtsi2ssl       %ecx, %xmm1
=======================================

So now emms is placed before mmx operation, and as a result, we receive wrong
answer.

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20180117/7bf1a3f9/attachment.html>