Issue 243046
Summary [X86][AVX512] use highway's method for v2i64/v4i64 division (!useAVX512Regs)
Labels enhancement, backend:X86, missed-optimization
Assignees
Reporter Andarwinux
    https://godbolt.org/z/75jWofhKP
```c++
#include "hwy/highway.h"
namespace hn = hwy::HWY_NAMESPACE;

hn::Vec<hn::FixedTag<uint64_t, 4>> udiv(
 hn::Vec<hn::FixedTag<uint64_t, 4>> a,
    hn::Vec<hn::FixedTag<uint64_t, 4>> b) {
    return a / b;
}

hn::Vec<hn::FixedTag<uint64_t, 4>> urem(
 hn::Vec<hn::FixedTag<uint64_t, 4>> a,
    hn::Vec<hn::FixedTag<uint64_t, 4>> b) {
    return a % b;
}

hn::Vec<hn::FixedTag<int64_t, 4>> sdiv(
 hn::Vec<hn::FixedTag<int64_t, 4>> a,
    hn::Vec<hn::FixedTag<int64_t, 4>> b) {
    return a / b;
}

hn::Vec<hn::FixedTag<int64_t, 4>> srem(
 hn::Vec<hn::FixedTag<int64_t, 4>> a,
    hn::Vec<hn::FixedTag<int64_t, 4>> b) {
    return a % b;
}
```
```asm
.LCPI0_0:
        .quad 1
.LCPI0_1:
        .quad 0x3ff0000000000000
udiv(hwy::N_AVX3_DL::Vec256<unsigned long>, hwy::N_AVX3_DL::Vec256<unsigned long>):
        vpmovq2m        k1, ymm1
 vpcmpnltuq      k0, ymm0, ymm1
        vpxor   xmm3, xmm3, xmm3
 vpbroadcastq    ymm1 {k1}, qword ptr [rip + .LCPI0_0]
        vpmovm2q ymm2, k0
        vpsrlq  ymm0 {k1}, ymm2, 63
        vbroadcastsd ymm2, qword ptr [rip + .LCPI0_1]
        vcvtuqq2pd      ymm4, ymm1
 vdivpd  ymm2, ymm2, ymm4
        vcvtuqq2pd      ymm4, ymm0
        vmulpd ymm4, ymm2, ymm4
        vcvttpd2uqq     ymm4, ymm4
        vpmullq ymm5, ymm4, ymm1
        vpsubq  ymm0, ymm0, ymm5
        vcvtqq2pd       ymm5, ymm0
        vmulpd  ymm2, ymm2, ymm5
        vcvttpd2qq      ymm2, ymm2
 vpmullq ymm5, ymm2, ymm1
        vpsubq  ymm0, ymm0, ymm5
 vpcmpnltuq      k0, ymm0, ymm1
        vpcmpnltq       k1, ymm0, ymm3
 vpaddq  ymm0, ymm2, ymm4
        vpmovm2q        ymm1, k0
        vpsubq ymm1 {k1}, ymm3, ymm1
        vpaddq  ymm0, ymm0, ymm1
 ret

.LCPI1_0:
        .quad   1
.LCPI1_1:
        .quad 0x3ff0000000000000
urem(hwy::N_AVX3_DL::Vec256<unsigned long>, hwy::N_AVX3_DL::Vec256<unsigned long>):
        vpcmpnltuq      k0, ymm0, ymm1
        vpmovq2m        k1, ymm1
        vmovdqa ymm3, ymm0
 vbroadcastsd    ymm5, qword ptr [rip + .LCPI1_1]
        vpmovm2q ymm2, k0
        vpsrlq  ymm3 {k1}, ymm2, 63
        vpblendmq       ymm2 {k1}, ymm1, qword ptr [rip + .LCPI1_0]{1to4}
        vcvtuqq2pd      ymm4, ymm2
        vdivpd  ymm4, ymm5, ymm4
        vcvtuqq2pd      ymm5, ymm3
 vmulpd  ymm5, ymm4, ymm5
        vcvttpd2uqq     ymm5, ymm5
 vpmullq ymm6, ymm5, ymm2
        vpsubq  ymm3, ymm3, ymm6
        vcvtqq2pd ymm6, ymm3
        vmulpd  ymm4, ymm4, ymm6
        vxorpd  xmm6, xmm6, xmm6
        vcvttpd2qq      ymm4, ymm4
        vpmullq ymm7, ymm4, ymm2
        vpsubq  ymm3, ymm3, ymm7
        vpcmpnltuq      k0, ymm3, ymm2
        vpcmpnltq       k1, ymm3, ymm6
        vpaddq  ymm3, ymm4, ymm5
        vpmovm2q        ymm2, k0
        vpsubq  ymm2 {k1}, ymm6, ymm2
        vpaddq  ymm2, ymm3, ymm2
        vpmullq ymm1, ymm2, ymm1
 vpsubq  ymm0, ymm0, ymm1
        ret

.LCPI2_0:
        .quad 0x3ff0000000000000
sdiv(hwy::N_AVX3_DL::Vec256<long>, hwy::N_AVX3_DL::Vec256<long>):
        vbroadcastsd    ymm2, qword ptr [rip + .LCPI2_0]
        vcvtqq2pd       ymm3, ymm1
        vpmovq2m        k1, ymm0
        vpxor   xmm6, xmm6, xmm6
        vdivpd  ymm2, ymm2, ymm3
 vcvtqq2pd       ymm3, ymm0
        vmulpd  ymm3, ymm2, ymm3
 vcvttpd2qq      ymm3, ymm3
        vpmullq ymm4, ymm3, ymm1
        vpsubq ymm4, ymm0, ymm4
        vcvtqq2pd       ymm5, ymm4
        vmulpd  ymm2, ymm2, ymm5
        vcvttpd2qq      ymm2, ymm2
        vpmullq ymm5, ymm2, ymm1
        vpsubq  ymm4, ymm4, ymm5
        vpxor   ymm0, ymm4, ymm1
 vpsubq  ymm4 {k1}, ymm6, ymm4
        vpabsq  ymm1, ymm1
 vpcmpnltuq      k0, ymm4, ymm1
        vpcmpnltq       k1, ymm0, ymm6
 vpaddq  ymm0, ymm2, ymm3
        vpmovm2q        ymm1, k0
        vpsubq ymm1 {k1}, ymm6, ymm1
        vpaddq  ymm0, ymm0, ymm1
 ret

.LCPI3_0:
        .quad 0x3ff0000000000000
srem(hwy::N_AVX3_DL::Vec256<long>, hwy::N_AVX3_DL::Vec256<long>):
        vbroadcastsd    ymm3, qword ptr [rip + .LCPI3_0]
        vcvtqq2pd       ymm2, ymm1
        vpmovq2m        k1, ymm0
        vpxor   xmm4, xmm4, xmm4
        vpabsq  ymm7, ymm1
 vdivpd  ymm2, ymm3, ymm2
        vcvtqq2pd       ymm3, ymm0
        vmulpd ymm3, ymm2, ymm3
        vcvttpd2qq      ymm3, ymm3
        vpmullq ymm5, ymm3, ymm1
        vpsubq  ymm5, ymm0, ymm5
        vcvtqq2pd       ymm6, ymm5
        vmulpd  ymm2, ymm2, ymm6
        vcvttpd2qq      ymm2, ymm2
 vpmullq ymm6, ymm2, ymm1
        vpaddq  ymm2, ymm2, ymm3
 vpsubq  ymm5, ymm5, ymm6
        vpxor   ymm6, ymm5, ymm1
        vpsubq ymm5 {k1}, ymm4, ymm5
        vpcmpnltuq      k0, ymm5, ymm7
 vpcmpnltq       k1, ymm6, ymm4
        vpmovm2q        ymm5, k0
 vpsubq  ymm5 {k1}, ymm4, ymm5
        vpaddq  ymm2, ymm2, ymm5
 vpmullq ymm1, ymm2, ymm1
        vpsubq  ymm0, ymm0, ymm1
 ret
```
Compared to sneller's method, highvay's method is slightly less efficient, but since it doesn't need ER, this also works with !useAVX512Regs.


_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs

Reply via email to