https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119554
Bug ID: 119554
Summary: [risc-v][bug] Unusual Behavior Observed with RISC-V
Vector Extension (RVV)
Product: gcc
Version: 14.2.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: akhilesh.k at samsung dot com
Target Milestone: ---
Hello
During the RISC-V VectoSC-V Vector Extension (RVV) performance test, I observed
some unusual behavior. It appears that the vector instructions vslideup.vi are
not functioning as expected on QEMU (when running with rvv_ta_all_1s=true, as
the default seems to be disabled) as well as on the target with GCC 14.2.
sharing a sample test application to reproduce this issue.
#include<stdio.h>
void test(char * counter)
{
unsigned int *ctr = (unsigned int *)counter;
unsigned int c_in = 0;
unsigned int c_out = 0;
for(int i = 0; i < 4; i++) {
c_out = ctr[i] >> 31;
ctr[i] <<= 1;
ctr[i] |= c_in;
c_in = c_out;
}
if(c_out) {
ctr[0] ^= 0x00000087;
}
}
int main () {
unsigned int counter[] = {0xe826d5bb, 0xc744010d, 0xccd28c99,
0xe3781d64};
test((char *)counter);
printf("test result %x %x %x %x \n", counter[0], counter[1],
counter[2], counter[3]);
return 0;
}
akhilesh.k@$ ./output/host/bin/riscv64-buildroot-linux-gnu-gcc -march=rv64gcv
-g -o test ~/Activity/rvv/test.c --static -O3
akhilesh.k@$
##### test results with rvv_ta_all_1s option #########
QEMU_CPU=rv64,vlen=128,rvv_ta_all_1s=true,rvv_ma_all_1s=true,v=true,vext_spec=v1.0
~/qemu_bin/bin/qemu-riscv64 test
test result d04dabf1 ffffffff ffffffff ffffffff
akhilesh.k@$
####### Default test results ###############
akhilesh.k$ ~/qemu_bin/bin/qemu-riscv64 test
test result d04dabf1 8e88021b 99a51933 c6f03ac9
akhilesh.k$
It seems that "Vector tail agnostic" bit set by the compiler incorrectly .
(gdb) c
Continuing.
Breakpoint 2.2, test (counter=0x7f1d2898afb8 "v\253M\320", '\377' <repeats 12
times>) at /home/akhilesh.k/Activity/rvv/test.c:30
30 ctr[0] ^= 0x00000087;
(gdb) info reg v
v0 {q = {0xffffffffffffffffffffffffffffff08}, l =
{0xffffffffffffff08, 0xffffffffffffffff}, w = {0xffffff08, 0xffffffff,
0xffffffff, 0xffffffff}, s = {0xff08, 0xffff,
0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff}, b = {0x8, 0xff <repeats 15
times>}}
v1 {q = {0xffffffffffffffffffffffffd04dab76}, l =
{0xffffffffd04dab76, 0xffffffffffffffff}, w = {0xd04dab76, 0xffffffff,
0xffffffff, 0xffffffff}, s = {0xab76, 0xd04d,
0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff}, b = {0x76, 0xab, 0x4d,
0xd0, 0xff <repeats 12 times>}}
v2 {q = {0xc6f03ac899a519328e88021ad04dab76}, l =
{0x8e88021ad04dab76, 0xc6f03ac899a51932}, w = {0xd04dab76, 0x8e88021a,
0x99a51932, 0xc6f03ac8}, s = {0xab76, 0xd04d,
0x21a, 0x8e88, 0x1932, 0x99a5, 0x3ac8, 0xc6f0}, b = {0x76, 0xab, 0x4d,
0xd0, 0x1a, 0x2, 0x88, 0x8e, 0x32, 0x19, 0xa5, 0x99, 0xc8, 0x3a, 0xf0, 0xc6}}
v3 {q = {0x1}, l = {0x1, 0x0}, w = {0x1, 0x0, 0x0, 0x0}, s = {0x1,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x1, 0x0 <repeats 15 times>}}
v4 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v5 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v6 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v7 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v8 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v9 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v10 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v11 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v12 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v13 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v14 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v15 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v16 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v17 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v18 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v19 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v20 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v21 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v22 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v23 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v24 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v25 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v26 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v27 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v28 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v29 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v30 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
v31 {q = {0x0}, l = {0x0, 0x0}, w = {0x0, 0x0, 0x0, 0x0}, s = {0x0,
0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}, b = {0x0 <repeats 16 times>}}
(gdb) p/x $vtype
$1 = 0xd0
(gdb)
