https://gcc.gnu.org/bugzilla/show_bug.cgi?id=125055
Bug ID: 125055
Summary: Wrong code vectorizing negation with -O2 -ftrapv
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: acoplan at gcc dot gnu.org
Target Milestone: ---
Consider the following testcase:
int a[8];
int b[8];
void foo (void) {
for (unsigned i = 0; i != 8; ++i)
b[i] = -a[i];
}
compiled with -O2 -ftrapv on aarch64, we get:
foo:
adrp x0, .LANCHOR0
add x0, x0, :lo12:.LANCHOR0
ldp q30, q31, [x0, 32]
neg v30.4s, v30.4s
neg v31.4s, v31.4s
stp q30, q31, [x0]
ret
but this is wrong since if any a[i] is INT_MIN, the program should trap due to
the presence of -ftrapv. If we add -fno-tree-vectorize, we instead get:
foo:
stp x29, x30, [sp, -48]!
mov x29, sp
stp x19, x20, [sp, 16]
adrp x20, .LANCHOR0
add x20, x20, :lo12:.LANCHOR0
str x21, [sp, 32]
add x21, x20, 32
mov x19, 0
.L2:
ldr w0, [x19, x21]
bl __negvsi2
str w0, [x20, x19]
add x19, x19, 4
cmp x19, 32
bne .L2
ldr x21, [sp, 32]
ldp x19, x20, [sp, 16]
ldp x29, x30, [sp], 48
ret
which looks correct; it calls a libgcc helper to handle the possible trap. I
think there are many related testcases where -ftrapv isn't respected by the
vectorizer, e.g. even a simple addition loop such as:
int a[8];
int b[8];
void foo (void) {
for (unsigned i = 0; i != 8; ++i)
a[i] += b[i];
}
is incorrectly vectorized in the presence of -ftrapv (the addition can of
course overflow), we don't check for it, instead generating:
foo:
adrp x0, .LANCHOR0
add x0, x0, :lo12:.LANCHOR0
ldp q28, q31, [x0]
ldp q29, q30, [x0, 32]
add v28.4s, v29.4s, v28.4s
add v30.4s, v31.4s, v30.4s
stp q28, q30, [x0]
ret
