https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98399
Bug ID: 98399
Summary: x86: Awful code generation for shifting vectors
Product: gcc
Version: 11.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: gabravier at gmail dot com
Target Milestone: ---
typedef char U __attribute__((vector_size(16)));
U f(U u)
{
return u >> (u & 1);
}
When compiled with -O3, on LLVM, this code generates this :
.LCPI0_0:
.zero 16,1
f(char __vector(16)): # @f(char __vector(16))
movdqa xmm3, xmmword ptr [rip + .LCPI0_0] # xmm3 =
[1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]
pand xmm3, xmm0
punpckhbw xmm1, xmm0 # xmm1 =
xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
psllw xmm3, 5
punpckhbw xmm4, xmm3 # xmm4 =
xmm4[8],xmm3[8],xmm4[9],xmm3[9],xmm4[10],xmm3[10],xmm4[11],xmm3[11],xmm4[12],xmm3[12],xmm4[13],xmm3[13],xmm4[14],xmm3[14],xmm4[15],xmm3[15]
pxor xmm2, xmm2
pxor xmm5, xmm5
pcmpgtw xmm5, xmm4
movdqa xmm6, xmm5
pandn xmm6, xmm1
psraw xmm1, 4
pand xmm1, xmm5
por xmm1, xmm6
paddw xmm4, xmm4
pxor xmm5, xmm5
pcmpgtw xmm5, xmm4
movdqa xmm6, xmm5
pandn xmm6, xmm1
psraw xmm1, 2
pand xmm1, xmm5
por xmm1, xmm6
paddw xmm4, xmm4
pxor xmm5, xmm5
pcmpgtw xmm5, xmm4
movdqa xmm4, xmm5
pandn xmm4, xmm1
psraw xmm1, 1
pand xmm1, xmm5
por xmm1, xmm4
psrlw xmm1, 8
punpcklbw xmm0, xmm0 # xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
punpcklbw xmm3, xmm3 # xmm3 = xmm3[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
pxor xmm4, xmm4
pcmpgtw xmm4, xmm3
movdqa xmm5, xmm4
pandn xmm5, xmm0
psraw xmm0, 4
pand xmm0, xmm4
por xmm0, xmm5
paddw xmm3, xmm3
pxor xmm4, xmm4
pcmpgtw xmm4, xmm3
movdqa xmm5, xmm4
pandn xmm5, xmm0
psraw xmm0, 2
pand xmm0, xmm4
por xmm0, xmm5
paddw xmm3, xmm3
pcmpgtw xmm2, xmm3
movdqa xmm3, xmm2
pandn xmm3, xmm0
psraw xmm0, 1
pand xmm0, xmm2
por xmm0, xmm3
psrlw xmm0, 8
packuswb xmm0, xmm1
ret
Which is rather long, however GCC generates this :
f(char __vector(16)):
push r15
movd edx, xmm0
push r14
push r13
push r12
push rbp
push rbx
sub rsp, 400
movdqa xmm1, XMMWORD PTR .LC0[rip]
movaps XMMWORD PTR [rsp+376], xmm0
movzx ebx, BYTE PTR [rsp+377]
pand xmm1, xmm0
movaps XMMWORD PTR [rsp+344], xmm0
movzx ebp, BYTE PTR [rsp+346]
movd ecx, xmm1
movaps XMMWORD PTR [rsp+360], xmm1
sar dl, cl
movzx ecx, BYTE PTR [rsp+361]
movaps XMMWORD PTR [rsp+328], xmm1
movaps XMMWORD PTR [rsp+312], xmm0
movzx edx, dl
movzx r12d, BYTE PTR [rsp+315]
sar bl, cl
movzx ecx, BYTE PTR [rsp+330]
movaps XMMWORD PTR [rsp+296], xmm1
movaps XMMWORD PTR [rsp+280], xmm0
movzx ebx, bl
movzx r13d, BYTE PTR [rsp+284]
sar bpl, cl
movzx ecx, BYTE PTR [rsp+299]
movaps XMMWORD PTR [rsp+264], xmm1
movaps XMMWORD PTR [rsp+248], xmm0
movzx ebp, bpl
movzx r14d, BYTE PTR [rsp+253]
sar r12b, cl
movzx ecx, BYTE PTR [rsp+268]
movaps XMMWORD PTR [rsp+232], xmm1
movaps XMMWORD PTR [rsp+216], xmm0
movzx r12d, r12b
movzx r15d, BYTE PTR [rsp+222]
sar r13b, cl
movzx ecx, BYTE PTR [rsp+237]
movaps XMMWORD PTR [rsp+200], xmm1
movzx r13d, r13b
sar r14b, cl
movzx ecx, BYTE PTR [rsp+206]
movaps XMMWORD PTR [rsp+184], xmm0
movzx eax, BYTE PTR [rsp+191]
movaps XMMWORD PTR [rsp+168], xmm1
movzx r14d, r14b
sar r15b, cl
movzx ecx, BYTE PTR [rsp+175]
movaps XMMWORD PTR [rsp+120], xmm0
movzx edi, BYTE PTR [rsp+129]
movaps XMMWORD PTR [rsp+152], xmm0
movzx esi, BYTE PTR [rsp+160]
movzx r15d, r15b
sar al, cl
movaps XMMWORD PTR [rsp+136], xmm1
movzx ecx, BYTE PTR [rsp+144]
movaps XMMWORD PTR [rsp+104], xmm1
sar sil, cl
movzx ecx, BYTE PTR [rsp+113]
movaps XMMWORD PTR [rsp+88], xmm0
mov BYTE PTR [rsp-89], sil
sar dil, cl
movaps XMMWORD PTR [rsp+72], xmm1
movzx ecx, BYTE PTR [rsp+82]
movzx esi, dil
movzx edi, BYTE PTR [rsp+98]
movaps XMMWORD PTR [rsp+56], xmm0
movzx r8d, BYTE PTR [rsp+67]
movaps XMMWORD PTR [rsp+40], xmm1
sar dil, cl
movzx ecx, BYTE PTR [rsp+51]
movaps XMMWORD PTR [rsp+24], xmm0
movzx r9d, BYTE PTR [rsp+36]
movaps XMMWORD PTR [rsp+8], xmm1
movzx edi, dil
sar r8b, cl
movzx ecx, BYTE PTR [rsp+20]
movaps XMMWORD PTR [rsp-8], xmm0
movzx r10d, BYTE PTR [rsp+5]
movaps XMMWORD PTR [rsp-24], xmm1
movzx r8d, r8b
sar r9b, cl
movzx ecx, BYTE PTR [rsp-11]
mov BYTE PTR [rsp-120], al
movaps XMMWORD PTR [rsp-40], xmm0
movzx r9d, r9b
sar r10b, cl
movaps XMMWORD PTR [rsp-56], xmm1
movzx ecx, BYTE PTR [rsp-42]
movzx r11d, BYTE PTR [rsp-26]
movaps XMMWORD PTR [rsp-72], xmm0
movzx eax, BYTE PTR [rsp-57]
movzx r10d, r10b
sar r11b, cl
movaps XMMWORD PTR [rsp-88], xmm1
movzx ecx, BYTE PTR [rsp-73]
movzx r11d, r11b
sar al, cl
movzx ecx, al
movzx eax, BYTE PTR [rsp-120]
sal rcx, 8
sal rax, 8
or rcx, r11
or rax, r15
sal rax, 8
or rax, r14
sal rax, 8
or rax, r13
sal rax, 8
or rax, r12
sal rax, 8
or rax, rbp
sal rax, 8
or rax, rbx
movzx ebx, BYTE PTR [rsp-89]
sal rax, 8
sal rcx, 8
or rcx, r10
or rax, rdx
sal rcx, 8
mov QWORD PTR [rsp-120], rax
or rcx, r9
sal rcx, 8
or rcx, r8
sal rcx, 8
or rcx, rdi
sal rcx, 8
or rcx, rsi
sal rcx, 8
or rcx, rbx
mov QWORD PTR [rsp-112], rcx
movdqa xmm0, XMMWORD PTR [rsp-120]
add rsp, 400
pop rbx
pop rbp
pop r12
pop r13
pop r14
pop r15
ret
.LC0:
.byte 1
.byte 1
.byte 1
.byte 1
.byte 1
.byte 1
.byte 1
.byte 1
.byte 1
.byte 1
.byte 1
.byte 1
.byte 1
.byte 1
.byte 1
.byte 1
Using such flags as `-mavx2` seems to marginally improve the situation, but on
LLVM results in far better code generation :
.LCPI0_0:
.zero 16,1
f(char __vector(16)): # @f(char __vector(16))
vpand xmm1, xmm0, xmmword ptr [rip + .LCPI0_0]
vpsllw xmm1, xmm1, 5
vpunpckhbw xmm2, xmm1, xmm1 # xmm2 =
xmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
vpunpckhbw xmm3, xmm0, xmm0 # xmm3 =
xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
vpsraw xmm4, xmm3, 4
vpblendvb xmm3, xmm3, xmm4, xmm2
vpsraw xmm4, xmm3, 2
vpaddw xmm2, xmm2, xmm2
vpblendvb xmm3, xmm3, xmm4, xmm2
vpsraw xmm4, xmm3, 1
vpaddw xmm2, xmm2, xmm2
vpblendvb xmm2, xmm3, xmm4, xmm2
vpsrlw xmm2, xmm2, 8
vpunpcklbw xmm1, xmm1, xmm1 # xmm1 = xmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
vpunpcklbw xmm0, xmm0, xmm0 # xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
vpsraw xmm3, xmm0, 4
vpblendvb xmm0, xmm0, xmm3, xmm1
vpsraw xmm3, xmm0, 2
vpaddw xmm1, xmm1, xmm1
vpblendvb xmm0, xmm0, xmm3, xmm1
vpsraw xmm3, xmm0, 1
vpaddw xmm1, xmm1, xmm1
vpblendvb xmm0, xmm0, xmm3, xmm1
vpsrlw xmm0, xmm0, 8
vpackuswb xmm0, xmm0, xmm2
ret
