https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79938
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
The situation is slightly better with GCC 7, only two spill/loads are
remaining.
Possibly BIT_INSERT_EXPR helps here. For the testcase you want to add
__attribute__((noinline)) to haddd_epu8 as otherwise eliding the result
vector after inlining into main is desirable (but not done). Then with GCC 7
we get:
haddd_epu8:
.LFB4966:
.cfi_startproc
pextrb $1, %xmm0, %edx
pextrb $0, %xmm0, %eax
addl %edx, %eax
pextrb $2, %xmm0, %edx
addl %edx, %eax
pextrb $3, %xmm0, %edx
addl %edx, %eax
pextrb $5, %xmm0, %edx
movl %eax, -12(%rsp)
pextrb $4, %xmm0, %eax
movd -12(%rsp), %xmm1
addl %edx, %eax
pextrb $6, %xmm0, %edx
addl %edx, %eax
pextrb $7, %xmm0, %edx
addl %edx, %eax
pextrb $9, %xmm0, %edx
pinsrd $1, %eax, %xmm1
pextrb $8, %xmm0, %eax
addl %edx, %eax
pextrb $10, %xmm0, %edx
addl %edx, %eax
pextrb $11, %xmm0, %edx
addl %edx, %eax
pextrb $13, %xmm0, %edx
pinsrd $2, %eax, %xmm1
pextrb $12, %xmm0, %eax
addl %edx, %eax
pextrb $14, %xmm0, %edx
addl %eax, %edx
pextrb $15, %xmm0, %eax
addl %edx, %eax
pinsrd $3, %eax, %xmm1
movdqa %xmm1, %xmm0
ret
which looks optimial to me. The single stack use is because by default
inter-unit moves are disabled. With -mtune=core-avx2 you'd get
haddd_epu8:
.LFB4966:
.cfi_startproc
pextrb $1, %xmm0, %edx
pextrb $0, %xmm0, %eax
addl %edx, %eax
pextrb $2, %xmm0, %edx
addl %edx, %eax
pextrb $3, %xmm0, %edx
addl %edx, %eax
pextrb $5, %xmm0, %edx
movd %eax, %xmm1
pextrb $4, %xmm0, %eax
addl %edx, %eax
pextrb $6, %xmm0, %edx
addl %edx, %eax
pextrb $7, %xmm0, %edx
addl %edx, %eax
pextrb $9, %xmm0, %edx
pinsrd $1, %eax, %xmm1
pextrb $8, %xmm0, %eax
addl %edx, %eax
pextrb $10, %xmm0, %edx
addl %edx, %eax
pextrb $11, %xmm0, %edx
addl %edx, %eax
pextrb $13, %xmm0, %edx
pinsrd $2, %eax, %xmm1
pextrb $12, %xmm0, %eax
addl %edx, %eax
pextrb $14, %xmm0, %edx
addl %eax, %edx
pextrb $15, %xmm0, %eax
addl %edx, %eax
pinsrd $3, %eax, %xmm1
movdqa %xmm1, %xmm0
ret
