https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118067
--- Comment #9 from Uroš Bizjak <ubizjak at gmail dot com> ---
Unfortunately, the testcase still fails when -mtune=k8 is added to compile
flags:
gcc -O -fno-split-wide-types -mavx512f -mtune=k8
in the same way as reported in Comment #5. The asm dump without -mtune=k8
(generic tuning) suggests that now alternative 9 is used in *movhi_interanl
movq %rdi, %rcx # 2 [c=4 l=3] *movdi_internal/3
vmovdqa64 %zmm0, %zmm5 # 3 [c=4 l=6] movv32hi_internal/3
-> vmovd %xmm0, %eax # 191 [c=4 l=6] *movhi_internal/9
shrw %ax # 9 [c=4 l=3] *lshrhi3_1/0
kmovw %eax, %k0 # 192 [c=4 l=4] *movhi_internal/4
as was previously case with -mtune=intel. OTOH, -mtune=k8 tuning forces
XMM->GPR moves through memory, which causes RA to ICE.
