1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
|
instruction: 0.968
graphic: 0.919
device: 0.848
assembly: 0.842
socket: 0.826
vnc: 0.774
network: 0.701
semantic: 0.687
boot: 0.686
mistranslation: 0.642
other: 0.581
KVM: 0.346
qemu-aarch64: Incorrect result for ssra instruction when using vector lengths of 1024-bit or higher.
Description of problem:
```
#include <arm_sve.h>
#include <stdio.h>
#define SZ 32
int main(int argc, char* argv[]) {
svbool_t pg = svptrue_b64();
uint64_t VL = svcntd();
fprintf(stderr, "One SVE vector can hold %li uint64_ts\n", VL);
int64_t sr[SZ], sx[SZ], sy[SZ];
uint64_t ur[SZ], ux[SZ], uy[SZ];
for (uint64_t i = 0; i < SZ; ++i) {
sx[i] = ux[i] = 0;
sy[i] = uy[i] = 1024;
}
for (uint64_t i = 0; i < SZ; i+=VL) {
fprintf(stderr, "Processing elements %li - %li\n", i, i + VL - 1);
svint64_t SX = svld1(pg, sx + i);
svint64_t SY = svld1(pg, sy + i);
svint64_t SR = svsra(SX, SY, 4);
svst1(pg, sr + i, SR);
svuint64_t UX = svld1(pg, ux + i);
svuint64_t UY = svld1(pg, uy + i);
svuint64_t UR = svsra(UX, UY, 4);
svst1(pg, ur + i, UR);
}
for (uint64_t i = 0; i < SZ; ++i) {
fprintf(stderr, "sr[%li]=%li, ur[%li]\n", i, sr[i], ur[i]);
}
return 0;
}
```
Steps to reproduce:
1. Build the above C source using "gcc -march=armv9-a -O1 ssra.c", can also use clang.
2. Run with "qemu-aarch64 -cpu max,sve-default-vector-length=64 ./a.out" and you'll see the expected result of 64 (signed and unsigned)
3. Run with "qemu-aarch64 -cpu max,sve-default-vector-length=128 ./a.out" and you'll see the expected result of 64 for unsigned but the signed result is 0. This suggests the emulation of SVE2 ssra instruction is incorrect for this and bigger vector lengths.
Additional information:
|