https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106081
--- Comment #8 from Jan Hubicka <hubicka at gcc dot gnu.org> --- Imagemagick improved by 17% on zen3 and 11% on altra https://lnt.opensuse.org/db_default/v4/SPEC/37550 https://lnt.opensuse.org/db_default/v4/SPEC/37543 which is cool :) The loop is now optimized as: .L2: vmovdqu16 (%rax), %zmm0 vmovupd (%rdx), %zmm2 addq $64, %rax subq $64, %rdx vpermpd %zmm2, %zmm15, %zmm9 vpermpd %zmm2, %zmm14, %zmm8 vpermpd %zmm2, %zmm13, %zmm7 vpermpd %zmm2, %zmm11, %zmm2 vpshufb %zmm12, %zmm0, %zmm0 vpmovsxwd %ymm0, %zmm1 vextracti64x4 $0x1, %zmm0, %ymm0 vpmovsxwd %ymm0, %zmm0 vcvtdq2pd %ymm1, %zmm10 vextracti32x8 $0x1, %zmm1, %ymm1 vcvtdq2pd %ymm1, %zmm1 vfmadd231pd %zmm2, %zmm10, %zmm6 vfmadd231pd %zmm9, %zmm1, %zmm3 vcvtdq2pd %ymm0, %zmm1 vextracti32x8 $0x1, %zmm0, %ymm0 vcvtdq2pd %ymm0, %zmm0 vfmadd231pd %zmm8, %zmm1, %zmm5 vfmadd231pd %zmm7, %zmm0, %zmm4 cmpq %rax, %rcx jne .L2
