https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107257
--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
With -fno-tree-slp-vectorize the failure goes away. diff in optimized:
__attribute__((noipa, noinline, noclone, no_icf))
void q ()
{
- <bb 2> [local count: 1073742492]:
- MEM <unsigned long> [(int *)&d] = 18380021091030725580;
+ int b_lsm.15;
+ int f;
+ _Bool _5;
+
+ <bb 2> [local count: 1073741824]:
+ MEM <vector(2) int> [(int *)&d] = { -15535156, -15535156 };
d[2] = -15535156;
+
+ <bb 3> [local count: 8687551919]:
+ # f_30 = PHI <f_25(5), 0(2)>
+ if (f_30 == 0)
+ goto <bb 5>; [1.43%]
+ else
+ goto <bb 4>; [98.57%]
+
+ <bb 4> [local count: 4698759900]:
+ if (f_30 != 2)
+ goto <bb 5>; [79.66%]
+ else
+ goto <bb 6>; [20.34%]
+
+ <bb 5> [local count: 7731921808]:
+ # f_25 = PHI <2(4), 1(3)>
+ goto <bb 3>; [100.00%]
+
+ <bb 6> [local count: 1073742492]:
a = 2;
- b = -1;
+ _5 = f_30 == 0;
+ b_lsm.15_33 = _5 ? 2 : -1;
+ b = b_lsm.15_33;
return;
}
both variants look OK so instead RTL expansion looks fishy(?)
;; b = b_lsm.15_33;
(insn 28 26 29 (parallel [
(set (reg:CCC 17 flags)
(ne:CCC (reg/v:SI 83 [ f ])
(const_int 0 [0])))
(set (reg:SI 89)
(neg:SI (reg/v:SI 83 [ f ])))
]) -1
(nil))
(insn 29 28 30 (parallel [
(set (reg:SI 90)
(neg:SI (ltu:SI (reg:CCC 17 flags)
(const_int 0 [0]))))
(clobber (reg:CC 17 flags))
]) -1
(nil))
(insn 30 29 31 (parallel [
(set (reg:SI 90)
(ior:SI (reg:SI 90)
(const_int 2 [0x2])))
(clobber (reg:CC 17 flags))
]) -1
(nil))
(insn 31 30 32 (set (reg:SI 88 [ b_lsm.15 ])
(reg:SI 90)) -1
(nil))
(insn 32 31 0 (set (mem/c:SI (symbol_ref:DI ("b") [flags 0x2] <var_decl
0x7ffff6526d80 b>) [1 b+0 S4 A32])
(reg:SI 88 [ b_lsm.15 ])) -1
(nil))
and we eventually optimize this to
q:
.LFB2:
.cfi_startproc
movq .LC0(%rip), %rax
movl $-15535156, d+8(%rip)
movl $2, a(%rip)
movq %rax, d(%rip)
movl $2, b(%rip)
ret
