https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98581
Bug ID: 98581
Summary: unexpected reassociation for umin/umax ?
Product: gcc
Version: unknown
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: ktkachov at gcc dot gnu.org
Target Milestone: ---
typedef signed int *__restrict__ pSINT;
typedef unsigned int *__restrict__ pUINT;
#define MIN(a, b) ((a) < (b) ? (a) : (b))
#define MAX(a, b) ((a) > (b) ? (a) : (b))
void saba_s (pSINT a, pSINT b, pSINT c)
{
int i;
for (i = 0; i < 4; i++)
c[i] += (MAX (a[i], b[i]) - MIN (a[i], b[i]));
}
void saba_u (pUINT a, pUINT b, pUINT c)
{
int i;
for (i = 0; i < 4; i++)
c[i] += (MAX (a[i], b[i]) - MIN (a[i], b[i]));
}
On aarch64 at -O3 generates:
saba_s:
ldr q0, [x0]
ldr q1, [x1]
ldr q2, [x2]
sabd v0.4s, v0.4s, v1.4s
add v0.4s, v0.4s, v2.4s
str q0, [x2]
ret
saba_u:
ldr q1, [x0]
ldr q2, [x1]
ldr q3, [x2]
umax v0.4s, v1.4s, v2.4s
umin v1.4s, v1.4s, v2.4s
add v0.4s, v0.4s, v3.4s
sub v0.4s, v0.4s, v1.4s
str q0, [x2]
ret
I would expect the (MAX (a[i], b[i]) - MIN (a[i], b[i])) part to match a uabd
instruction for the unsigned case, but it looks like the add and sub operations
are swapped which prevents the RTL pattern matching the operation.
This comes out this way out of GIMPLE. At expand the signed version is:
vect__4.6_40 = MEM <vector(4) int> [(int *)c_16(D)];
vect__6.9_37 = MEM <vector(4) int> [(int *)b_17(D)];
vect__8.12_34 = MEM <vector(4) int> [(int *)a_18(D)];
vect__9.13_33 = MAX_EXPR <vect__8.12_34, vect__6.9_37>;
vect__10.14_32 = MIN_EXPR <vect__8.12_34, vect__6.9_37>;
vect__11.15_31 = vect__9.13_33 - vect__10.14_32;
vect__12.16_30 = vect__11.15_31 + vect__4.6_40;
MEM <vector(4) int> [(int *)c_16(D)] = vect__12.16_30;
return;
the unsigned is:
vect__4.25_38 = MEM <vector(4) unsigned int> [(unsigned int *)c_16(D)];
vect__6.28_35 = MEM <vector(4) unsigned int> [(unsigned int *)b_17(D)];
vect__8.31_32 = MEM <vector(4) unsigned int> [(unsigned int *)a_18(D)];
vect__9.32_31 = MAX_EXPR <vect__8.31_32, vect__6.28_35>;
vect__10.33_30 = MIN_EXPR <vect__8.31_32, vect__6.28_35>;
vect__13.34_29 = vect__9.32_31 + vect__4.25_38;
vect__12.35_28 = vect__13.34_29 - vect__10.33_30;
MEM <vector(4) unsigned int> [(unsigned int *)c_16(D)] = vect__12.35_28;
return;