https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114162
Bug ID: 114162
Summary: Missing Optimization: Loop is vectorized instead of
removed
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: carnet at student dot ethz.ch
Target Milestone: ---
This simple for-loop is compiled to complex vectorized code instead of
optmizing the loop like clang is able to.
https://godbolt.org/z/c3fjWbWEz
Source:
unsigned short a;
int main() {
int counter = 0;
// n and a are mutliples of 4
int n = 4;
a *= n;
for (; a > 0; a += n)
counter++;
return counter;
}
x86 -O3 Assembly:
main:
movzwl a(%rip), %edx
salw $2, %dx
je .L24
movl $-4, %ecx
subl %edx, %ecx
cmpw $52, %cx
jbe .L15
shrw $2, %cx
movd %edx, %xmm0
movl $8, %edi
xorl %eax, %eax
addl $1, %ecx
punpcklwd %xmm0, %xmm0
movd %edi, %xmm3
movdqa .LC0(%rip), %xmm1
movl %ecx, %esi
pshufd $0, %xmm0, %xmm0
movl $2097184, %edi
pshufd $0, %xmm3, %xmm3
shrw $3, %si
paddw .LC1(%rip), %xmm0
movd %edi, %xmm2
movzwl %si, %esi
pshufd $0, %xmm2, %xmm2
.L5:
addl $1, %eax
movdqa %xmm1, %xmm4
movdqa %xmm0, %xmm5
paddd %xmm3, %xmm1
paddw %xmm2, %xmm0
cmpl %eax, %esi
jne .L5
movl %ecx, %esi
movl %ecx, %eax
andl $-8, %esi
andl $32760, %eax
andl $7, %ecx
leal (%rdx,%rsi,4), %edx
je .L6
cmpw $-4, %dx
je .L25
cmpw $-8, %dx
je .L26
cmpw $-12, %dx
je .L27
cmpw $-16, %dx
je .L28
cmpw $-20, %dx
je .L29
cmpw $-24, %dx
je .L30
addl $7, %eax
addl $28, %edx
.L14:
movw %dx, a(%rip)
ret
.L26:
addl $2, %eax
xorl %edx, %edx
jmp .L14
.L24:
movw $0, a(%rip)
xorl %eax, %eax
ret
.L15:
xorl %eax, %eax
andb $4, %cl
jne .L4
movl $1, %eax
addw $4, %dx
je .L14
.L4:
addl $2, %eax
addw $8, %dx
jne .L4
jmp .L14
.L6:
paddd .LC5(%rip), %xmm4
paddw .LC7(%rip), %xmm5
pshufd $255, %xmm4, %xmm0
pextrw $7, %xmm5, %edx
movd %xmm0, %eax
jmp .L14
.L25:
addl $1, %eax
xorl %edx, %edx
jmp .L14
.L27:
addl $3, %eax
xorl %edx, %edx
jmp .L14
.L28:
addl $4, %eax
xorl %edx, %edx
jmp .L14
.L29:
addl $5, %eax
xorl %edx, %edx
jmp .L14
.L30:
addl $6, %eax
xorl %edx, %edx
jmp .L14
a:
.zero 2
.LC0:
.long 0
.long 1
.long 2
.long 3
.LC1:
.value 0
.value 4
.value 8
.value 12
.value 16
.value 20
.value 24
.value 28
.LC5:
.long 5
.long 5
.long 5
.long 5
.LC7:
.long 262148
.long 262148
.long 262148
.long 262148
