Unsupported vector mode is passed incorrectly

[hjl-tools]

Bugzilla Link	39501
Version	trunk
OS	Linux
CC	@topperc,@echristo,@erichkeane,@iains,@rnk

Extended Description

[hjl@gnu-cfl-1 tmp]$ cat x.c
typedef int __attribute__((mode(QI))) qi;
typedef qi __attribute__((vector_size (32))) v32qi;

v32qi foo (int x)
{
 v32qi y = {'0','1','2','3','4','5','6','7','8','9','a','b','c','d','e','f',
         '0','1','2','3','4','5','6','7','8','9','a','b','c','d','e','f'};
 return y;
}
[hjl@gnu-cfl-1 tmp]$ clang -mno-avx -S -O2 x.c

    movaps  .LCPI0_0(%rip), %xmm0   # xmm0 = [48,49,50,51,52,53,54,55,56,57,97,98,99,100,101,102]
    movaps  %xmm0, %xmm1
    retq

-mno-avx should pass v32qi as before AVX was enabled. For example,

[hjl@gnu-cfl-1 tmp]$ cat z.c
typedef int __attribute__((mode(SI))) si;
typedef si __attribute__((vector_size (128))) v;

v
foo (int x)
{
 v y = {'0','1','2','3','4','5','6','7','8','9','a','b','c','d','e','f',
         '0','1','2','3','4','5','6','7','8','9','a','b','c','d','e','f'};
 return y;
}
[hjl@gnu-cfl-1 tmp]$ clang -S -O2 z.c
...
    movaps  .LCPI0_0(%rip), %xmm0   # xmm0 = [99,100,101,102]
    movaps  %xmm0, 112(%rdi)
    movaps  .LCPI0_1(%rip), %xmm1   # xmm1 = [56,57,97,98]
    movaps  %xmm1, 96(%rdi)
    movaps  .LCPI0_2(%rip), %xmm2   # xmm2 = [52,53,54,55]
    movaps  %xmm2, 80(%rdi)
    movaps  .LCPI0_3(%rip), %xmm3   # xmm3 = [48,49,50,51]
    movaps  %xmm3, 64(%rdi)
    movaps  %xmm0, 48(%rdi)
    movaps  %xmm1, 32(%rdi)
    movaps  %xmm2, 16(%rdi)
    movaps  %xmm3, (%rdi)
    movq    %rdi, %rax
    retq

If there were AVX1024, clang would have passed y in vector register. When AVX1024 were disabled, clang would have passed y in memory and issued an ABI change warning.

[topperc]

As of 2831a317b689c7f005a29f008a8e4c24485c0711 we should at least warn on ABI mismatches. But we still pass large vectors differently than gcc.

[topperc]

Candidate patch https://reviews.llvm.org/D53919

[topperc]

If I'm understanding what's happening right. The frontend is always sending the type directly into IR as the full vector type. The backend will try to fit that type into 4 of whatever vector register width is available. If it doesn't fit it falls back to memory.

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: None (hjl-tools)

| | | | --- | --- | | Bugzilla Link | [39501](https://llvm.org/bz39501) | | Version | trunk | | OS | Linux | | CC | @topperc,@echristo,@erichkeane,@iains,@rnk | ## Extended Description [hjl@gnu-cfl-1 tmp]$ cat x.c typedef int __attribute__((mode(QI))) qi; typedef qi __attribute__((vector_size (32))) v32qi; v32qi foo (int x) { v32qi y = {'0','1','2','3','4','5','6','7','8','9','a','b','c','d','e','f', '0','1','2','3','4','5','6','7','8','9','a','b','c','d','e','f'}; return y; } [hjl@gnu-cfl-1 tmp]$ clang -mno-avx -S -O2 x.c movaps .LCPI0_0(%rip), %xmm0 # xmm0 = [48,49,50,51,52,53,54,55,56,57,97,98,99,100,101,102] movaps %xmm0, %xmm1 retq -mno-avx should pass v32qi as before AVX was enabled. For example, [hjl@gnu-cfl-1 tmp]$ cat z.c typedef int __attribute__((mode(SI))) si; typedef si __attribute__((vector_size (128))) v; v foo (int x) { v y = {'0','1','2','3','4','5','6','7','8','9','a','b','c','d','e','f', '0','1','2','3','4','5','6','7','8','9','a','b','c','d','e','f'}; return y; } [hjl@gnu-cfl-1 tmp]$ clang -S -O2 z.c ... movaps .LCPI0_0(%rip), %xmm0 # xmm0 = [99,100,101,102] movaps %xmm0, 112(%rdi) movaps .LCPI0_1(%rip), %xmm1 # xmm1 = [56,57,97,98] movaps %xmm1, 96(%rdi) movaps .LCPI0_2(%rip), %xmm2 # xmm2 = [52,53,54,55] movaps %xmm2, 80(%rdi) movaps .LCPI0_3(%rip), %xmm3 # xmm3 = [48,49,50,51] movaps %xmm3, 64(%rdi) movaps %xmm0, 48(%rdi) movaps %xmm1, 32(%rdi) movaps %xmm2, 16(%rdi) movaps %xmm3, (%rdi) movq %rdi, %rax retq If there were AVX1024, clang would have passed y in vector register. When AVX1024 were disabled, clang would have passed y in memory and issued an ABI change warning.