https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122189
Bug ID: 122189
Summary: [MVE] carry-in incorrectly taken into account by vadcq
Product: gcc
Version: 13.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: clyon at gcc dot gnu.org
Target Milestone: ---
Target: arm
#include <arm_mve.h>
#include <inttypes.h>
#include <stdio.h>
__attribute((noinline)) void print_uint32x4_t(const char *name,
uint32x4_t val) {
printf("%s: %u, %u, %u, %u\n", name, vgetq_lane_u32(val, 0),
vgetq_lane_u32(val, 1), vgetq_lane_u32(val, 2),
vgetq_lane_u32(val, 3));
}
void __attribute__ ((noinline)) test_2(void) {
uint32x4_t v12 = vdupq_n_u32(1);
unsigned v17 = 0;
uint32x4_t v18 = vadcq_u32(v12, v12, &v17);
v17 = 1;
uint32x4_t v108 = vadcq_u32(v12, v12, &v17);
print_uint32x4_t("v108", v108);
}
int main() { test_2(); }
Compiled with -O2 -march=armv8.1-m.main+mve.fp -mfloat-abi=hard
prints
v108: 2, 2, 2, 2
instead of the expected
v108: 3, 2, 2, 2
generates:
test_2:
@ args = 0, pretend = 0, frame = 8
@ frame_needed = 0, uses_anonymous_args = 0
push {lr}
sub sp, sp, #12
vmov.i32 q3, #0x1 @ v4si
vmrs r3, FPSCR_nzcvqc
bic r3, r3, #536870912
vmsr FPSCR_nzcvqc, r3
vadc.i32 q0, q3, q3
vmrs r3, FPSCR_nzcvqc
vmrs r3, FPSCR_nzcvqc
bic r3, r3, #536870912
orr r3, r3, #536870912
vmsr FPSCR_nzcvqc, r3
vadc.i32 q3, q3, q3
vmrs r3, FPSCR_nzcvqc
ubfx r3, r3, #29, #1
str r3, [sp, #4]
ldr r0, .L7
bl print_uint32x4_t
add sp, sp, #12
@ sp needed
ldr pc, [sp], #4
.L8:
.align 2
.L7:
.word .LC1
.size test_2, .-test_2
The problem is that the vadcq pattern does not mention vfpcc as input in the
"add" part. As a result, the compiler sees 2 additions with the same inputs and
concludes both give the same results, thus using the 1st result as input of the
print routine.
In addition, the "set carry" part of the pattern should use a different unspec
code.
