diff options
Diffstat (limited to 'results/classifier/accel-gemma3:12b/tcg/1245543')
| -rw-r--r-- | results/classifier/accel-gemma3:12b/tcg/1245543 | 24 |
1 files changed, 24 insertions, 0 deletions
diff --git a/results/classifier/accel-gemma3:12b/tcg/1245543 b/results/classifier/accel-gemma3:12b/tcg/1245543 new file mode 100644 index 00000000..fdd7f401 --- /dev/null +++ b/results/classifier/accel-gemma3:12b/tcg/1245543 @@ -0,0 +1,24 @@ + +Wrong implementation of SSE4.1 pmovzxbw and similar instructions + +QEMU 1.5.0 (and git version, as far as I can tell from the source code) has incorrect implementation of pmovzxbw and similar SSE4.1 instructions. The instruction zero-extends the first 8 8-bit elements of a vector to 16bit vector and puts them to another vector. The current implementation applies this operation only to the first element and zeros out the rest. + +To verify, compile the attached program for SSE4.1 (g++ -msse4.1 cvtint.cc). On real hardware, it produces the following output: + +$ ./a.out +1 0 2 0 3 0 4 0 5 0 6 0 7 0 8 0 + +On QEMU, the output is as follows: + +$ ./a.out +1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 + +QEMU is invoked as: + +qemu-system-x86_64 \ + -M pc -cpu Haswell,+sse4.1,+avx,+avx2,+fma,enforce -m 512 \ + -serial stdio -no-reboot \ + -kernel vmlinuz -initrd initrd.img \ + -netdev user,id=user.0 -device rtl8139,netdev=user.0 -redir tcp:2222::22 \ + -hda ubuntu-amd64.ext3 \ + --append "rw console=tty root=/dev/sda" \ No newline at end of file |