https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107093
Bug ID: 107093
Summary: AVX512 mask operations not simplified in fully masked
loop
Product: gcc
Version: 13.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: rguenth at gcc dot gnu.org
Target Milestone: ---
Trying to implement WHILE_ULT for AVX512 I run into optimization issues.
Consider
double a[1024], b[1024];
void foo (int n)
{
for (int i = 0; i < n; ++i)
a[i] = b[i] * 3.;
}
compiled with -O3 -march=cascadelake --param vect-partial-vector-usage=2
I get snippets like
kxnorb %k1, %k1, %k1
kortestb %k1, %k1
je .L11
or
kxorb %k1, %k1, %k1
kxnorb %k1, %k1, %k1
where we fail to simplify the operations. Looking at the RTL it looks
like missed jump threading, but I do see the ops being
(insn 18 72 74 5 (parallel [
(set (reg:QI 69 k1 [orig:86 loop_mask_15 ] [86])
(not:QI (xor:QI (reg:QI 69 k1 [orig:86 loop_mask_15 ] [86])
(reg:QI 69 k1 [orig:86 loop_mask_15 ] [86]))))
(unspec [
(const_int 0 [0])
] UNSPEC_MASKOP)
]) 1912 {kxnorqi}
(expr_list:REG_EQUAL (const_int -1 [0xffffffffffffffff])
(nil)))
thus having an UNSPEC in them. When emitting a SET from constm1 I end up
with mask<->GPR moves and if-converted code which isn't optimal either.
When doing -fno-if-conversion I get
.L7:
vmovapd b(%rax), %ymm1{%k1}
addl $4, %ecx
movl %edi, %edx
vmulpd %ymm2, %ymm1, %ymm0
subl %ecx, %edx
vmovapd %ymm0, a(%rax){%k1}
kxnorb %k1, %k1, %k1
cmpl $4, %edx
jge .L5
vpbroadcastd %edx, %xmm0
vpcmpd $1, %xmm0, %xmm3, %k1
.L5:
addq $32, %rax
kortestb %k1, %k1
jne .L7
which also doesn't have the desired short-cut from the cmpl $4, %edx.