https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96831
Bug ID: 96831
Summary: gcc.dg/tree-ssa/scev-[345].c FAIL on i?86 and arm
Product: gcc
Version: 11.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: rguenth at gcc dot gnu.org
Target Milestone: ---
The testcases now FAIL after dumping TARGET_MEM_REF bases as &a instead of
symbol: a and thus show that IVOPTs doesn't consider using a pointer IV:
For lp64 (x86_64) I see
<bb 4> :
_10 = (sizetype) k_4(D);
_8 = _10 * 8;
_6 = (sizetype) k_4(D);
_3 = _6 * 8;
_2 = _3 + 4;
_1 = &a + _2;
ivtmp.5_7 = (unsigned long) _1;
<bb 5> :
# i_12 = PHI <k_4(D)(4), i_9(6)>
# ivtmp.5_5 = PHI <ivtmp.5_7(4), ivtmp.5_11(6)>
_1_16 = (int *) ivtmp.5_5;
a_p = _1_16;
_19 = (void *) ivtmp.5_5;
MEM[(int *)_19] = 100;
i_9 = k_4(D) + i_12;
ivtmp.5_11 = ivtmp.5_5 + _8;
if (i_9 <= 999)
so the ADDRESS use and the dereference use use the same IV. But with ilp32
we end up with
<bb 5> :
# i_12 = PHI <k_4(D)(4), i_9(6)>
_5 = (unsigned int) &a;
_11 = (unsigned int) i_12;
_10 = _11 * 8;
_8 = _5 + _10;
_7 = _8 + 4;
_6 = (int *) _7;
_1_16 = _6;
a_p = _1_16;
_3 = (sizetype) i_12;
MEM[(int *)&a + _3 * 8] = 100;
i_9 = k_4(D) + i_12;
if (i_9 <= 999)
(scev-4.c) and eventually
.L2:
movl $100, a+4(,%eax,8)
leal a+4(,%eax,8), %ecx
addl %edx, %eax
cmpl $999, %eax
jle .L2
movl %ecx, a_p
instead of
.L3:
addl %edi, %ecx
movl $100, (%rax)
movq %rax, %rsi
addq %rdx, %rax
cmpl $999, %ecx
jle .L3
movq %rsi, a_p(%rip)
it's not entirely clear whether based on costs one is better than the other
(we are not able to replace the COMPARE use and thus the tested for variant
uses one more IV).
