https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123225
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Priority|P3 |P1
--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
I think this ultimatively requires targets to punt on some cases. On x86 we
get with -march=znver5
foo:
.LFB0:
.cfi_startproc
leaq 2(%rdi), %rcx
movq %rdi, %rax
shrq %rcx
negq %rcx
andl $31, %ecx
je .L2
xorl %edx, %edx
.p2align 5
.p2align 4
.p2align 3
.L4: // alignment loop
addq $2, %rax
cmpw $0, (%rax)
je .L1
incq %rdx
cmpq %rdx, %rcx
jne .L4
.L2:
leaq 2(%rdi,%rcx,2), %rsi
xorl %edx, %edx
.p2align 5
.p2align 4
.p2align 3
.L5: // vector loop
vpxor %xmm0, %xmm0, %xmm0
movq %rdx, %rcx
vpcmpeqw (%rsi,%rdx,2), %zmm0, %k0
addq $32, %rdx
kortestd %k0, %k0
je .L5
leaq (%rax,%rcx,2), %rax
.p2align 4
.p2align 4
.p2align 3
.L6: // early-break epilog
addq $2, %rax
cmpw $0, (%rax)
jne .L6
at least until we fix that there's no epilog required the thing is to
observe that in the alignment loop the alignment check is of comparable
complexity to the original scalar iteration, so that alignment loop
covering up to 63 elements is much slower than the scalar loop.
Possibly only loops with much more actual work per scalar iteration
are worth vectorizing this way (unless we can elide the alignment loop).
