Issue 245064
Summary [X86][CostModel] incorrect vector ICMP costs lead to poor codegen
Labels backend:X86, backend:X86 Scheduler+Cost Models, llvm:loopvectorize
Assignees
Reporter Andarwinux
    found from https://github.com/FFmpeg/FFmpeg/blob/master/libavfilter/vf_colordetectdsp.h#L57-L83
https://godbolt.org/z/feE6T1a6e
```c++
#include <stdint.h>
#include <stddef.h>
#define DECL_DETECT_RANGE_IMPL(TYPE, NAME) \
int NAME(const uint8_t* data, ptrdiff_t stride,                  \
                       ptrdiff_t width, ptrdiff_t height,                      \
                       int mpeg_min, int mpeg_max)                             \
{ \
    const TYPE min = mpeg_min;                                                 \
    const TYPE max = mpeg_max;                                                 \
 __builtin_assume(min == mpeg_min); \
    __builtin_assume(max == mpeg_max); \
 \
    while (height--) { \
        const TYPE *row = (const TYPE *) data; \
        uint8_t cond = 0; \
        for (int x = 0; x < width; x++) { \
            const TYPE val = row[x]; \
            cond |= val < min || val > max;                                    \
        } \
        if (cond) \
            return 1; \
        data += stride;                                                        \
    } \
 \
 return 0; \
}

DECL_DETECT_RANGE_IMPL(uint16_t, ff_detect_range16_c)
```
 -O3 -march=znver5 -mllvm -force-vector-width=16
```asm
.LBB0_9:
 vpmovzxwd       zmm8, ymmword ptr [rdi + 2*r10 + 32]
        vpmovzxwd zmm7, ymmword ptr [rdi + 2*r10]
        vpmovzxwd       zmm9, ymmword ptr [rdi + 2*r10 + 64]
        vpmovzxwd       zmm10, ymmword ptr [rdi + 2*r10 + 96]
        add     r10, 64
        vpcmpnleud      k1, zmm0, zmm8
 vpcmpltud       k5, zmm1, zmm8
        vpcmpnleud      k0, zmm0, zmm7
 vpcmpltud       k4, zmm1, zmm7
        vpcmpnleud      k2, zmm0, zmm9
 vpcmpltud       k6, zmm1, zmm9
        vpcmpnleud      k3, zmm0, zmm10
 vpcmpltud       k7, zmm1, zmm10
        vpor    xmm8, xmm6, xmm2
 vpor    xmm7, xmm4, xmm2
        vpor    xmm9, xmm5, xmm2
        korw k1, k1, k5
        korw    k4, k0, k4
        korw    k2, k2, k6
 korw    k3, k3, k7
        vmovdqu8        xmm6 {k1}, xmm8
        vpor xmm8, xmm3, xmm2
        vmovdqu8        xmm4 {k4}, xmm7
        vmovdqu8 xmm5 {k2}, xmm9
        vmovdqu8        xmm3 {k3}, xmm8
        cmp rax, r10
        jne     .LBB0_9
```
-O3 -march=znver5
```asm
.LBB0_11:
        vmovdqu64       zmm8, zmmword ptr [rdi + 2*r11]
        vmovdqu64       zmm9, zmmword ptr [rdi + 2*r11 + 64]
 vmovdqu64       zmm10, zmmword ptr [rdi + 2*r11 + 128]
 vmovdqu64       zmm11, zmmword ptr [rdi + 2*r11 + 192]
        sub     r11, -128
        vextracti64x4   ymm12, zmm8, 1
        vextracti64x4   ymm13, zmm9, 1
        vpmovzxwd       zmm8, ymm8
        vextracti64x4   ymm14, zmm10, 1
        vpmovzxwd       zmm9, ymm9
        vextracti64x4   ymm15, zmm11, 1
        vpmovzxwd       zmm10, ymm10
        vpmovzxwd zmm11, ymm11
        vpmovzxwd       zmm12, ymm12
        vpmovzxwd zmm13, ymm13
        vpmovzxwd       zmm14, ymm14
        vpmovzxwd zmm15, ymm15
        vpcmpnleud      k0, zmm0, zmm8
        vpcmpnleud k1, zmm0, zmm12
        vpcmpnleud      k2, zmm0, zmm13
        vpcmpnleud k3, zmm0, zmm14
        vpcmpnleud      k4, zmm0, zmm15
 vpcmpltud       k5, zmm1, zmm12
        vpcmpltud       k6, zmm1, zmm13
 vpcmpltud       k7, zmm1, zmm14
        kunpckwd        k0, k1, k0
 vpcmpnleud      k1, zmm0, zmm9
        kunpckwd        k1, k2, k1
 vpcmpnleud      k2, zmm0, zmm10
        kmovd   dword ptr [rsp - 4], k0
 vpcmpltud       k0, zmm1, zmm15
        kunpckwd        k2, k3, k2
 vpcmpnleud      k3, zmm0, zmm11
        kunpckwd        k3, k4, k3
 vpcmpltud       k4, zmm1, zmm8
        vpor    ymm8, ymm5, ymm3
 kunpckwd        k4, k5, k4
        vpcmpltud       k5, zmm1, zmm9
 vpor    ymm9, ymm7, ymm3
        kunpckwd        k5, k6, k5
 vpcmpltud       k6, zmm1, zmm10
        vpor    ymm10, ymm6, ymm3
 kunpckwd        k6, k7, k6
        vpcmpltud       k7, zmm1, zmm11
 kord    k1, k1, k5
        kunpckwd        k0, k0, k7
        kmovd   k7, dword ptr [rsp - 4]
        kord    k2, k2, k6
        vmovdqu8        ymm7 {k1}, ymm9
        vpor    ymm9, ymm4, ymm3
        kord    k3, k3, k0
 vmovdqu8        ymm6 {k2}, ymm10
        vmovdqu8        ymm4 {k3}, ymm9
        kord    k4, k7, k4
        vmovdqu8        ymm5 {k4}, ymm8
 cmp     r10, r11
        jne     .LBB0_11
```
```
Cost of 0 for VF 16: IR   %cmp9.us = icmp samesign ugt i32 %mpeg_min, %conv7.us
Cost of 0 for VF 16: IR   %cmp12.us = icmp samesign ult i32 %mpeg_max, %conv7.us
Cost of 0 for VF 32: IR   %cmp9.us = icmp samesign ugt i32 %mpeg_min, %conv7.us
Cost of 0 for VF 32: IR   %cmp12.us = icmp samesign ult i32 %mpeg_max, %conv7.us
```
This is clearly incorrect, v32i32 is illegal.
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs

Reply via email to