https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92645
--- Comment #13 from Richard Biener <rguenth at gcc dot gnu.org> ---
So with all tricks I arrive at the following for the reduced testcase
f:
.LFB2:
.cfi_startproc
pushq %rbp
.cfi_def_cfa_offset 16
.cfi_offset 6, -16
movl %ecx, %r9d
vpxor %xmm4, %xmm4, %xmm4
movl $255, %eax
shrl $24, %r9d
subl %r9d, %eax
movq %rsp, %rbp
.cfi_def_cfa_register 6
andq $-32, %rsp
movl %eax, %r8d
vmovdqa %ymm4, -32(%rsp)
vmovd %ecx, %xmm4
shrl $7, %eax
vpshufd $0, %xmm4, %xmm4
addl %r8d, %eax
vmovaps %xmm4, -32(%rsp)
vpmovzxbw -32(%rsp), %ymm2
vmovd %eax, %xmm0
vpbroadcastb %xmm0, %xmm0
vpsllw $8, %ymm2, %ymm2
vpaddw .LC1(%rip), %ymm2, %ymm2
testq %rdx, %rdx
je .L10
vpxor %xmm5, %xmm5, %xmm5
vmovdqa .LC0(%rip), %xmm3
xorl %eax, %eax
vmovdqa %ymm5, -32(%rsp)
vmovaps %xmm0, -32(%rsp)
vpmovzxbw -32(%rsp), %ymm4
.p2align 4,,10
.p2align 3
.L7:
vmovdqu (%rsi,%rax), %xmm6
vpxor %xmm5, %xmm5, %xmm5
vmovdqa %ymm5, -32(%rsp)
vmovaps %xmm6, -32(%rsp)
vpmovzxbw -32(%rsp), %ymm0
vpmullw %ymm4, %ymm0, %ymm0
vpaddw %ymm2, %ymm0, %ymm0
vpsrlw $8, %ymm0, %ymm0
vmovdqa %xmm0, %xmm1
vextracti128 $0x1, %ymm0, %xmm0
vpand %xmm1, %xmm3, %xmm1
vpand %xmm0, %xmm3, %xmm0
vpackuswb %xmm0, %xmm1, %xmm0
vmovups %xmm0, (%rdi,%rax)
addq $16, %rax
subq $4, %rdx
jne .L7
.L10:
vzeroupper
leave
.cfi_def_cfa 7, 8
ret
.cfi_endproc
.LFE2:
.size f, .-f
where the stack spills still look bad - shomehow we don't like
_60 = BIT_INSERT_EXPR <{ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 }, _4, 0>;
_61 = [vec_unpack_lo_expr] _60;
which is "widening" _4 to double vector size when we know we'll just need
the lowpart for the VEC_UNPACK_LO_EXPR. This _should_ translate to
a mov %xmm, %ymm but somehow it doesn't.
A small testcase for that is the zxt() function in the reduced testcase.
Using an undef SSA name in place off the { 0, ... } vector doesn't help
either. A simple VIEW_CONVERT isn't valid (it changes size).
