add deepseek and gemma results

1 files changed, 53 insertions, 0 deletions

diff --git a/results/classifier/gemma3:12b/mistranslation/2865 b/results/classifier/gemma3:12b/mistranslation/2865
new file mode 100644
index 00000000..c2b6687b
--- /dev/null
+++ b/results/classifier/gemma3:12b/mistranslation/2865
@@ -0,0 +1,53 @@
+
+loongarch64: wrong implementation of `xvldi` instruction
+Description of problem:
+Consider this sample program.
+
+```c++
+#include <cstdio>
+#include <cstdint>
+#include <lsxintrin.h>
+#include <lasxintrin.h>
+
+void dump_u32(__m256i x) {
+    uint32_t tmp[32/4];
+    __lasx_xvst(x, tmp, 0);
+    putchar('[');
+    for (int i=0; i < 32/4; i++) {
+        if (i > 0) {
+            putchar(' ');
+        }
+
+        printf("%08x", tmp[i]);
+    }
+    puts("]");
+}
+
+int main() {
+    __m256i const1 = __lasx_xvldi(-3832);
+    dump_u32(const1);
+}
+```
+
+The magic constants here means: replicate in 32-bit words a byte (0x4) shifted left by 8. We should have a vector of words 0x800, and indeed, the program run on a real hardware prints expected:
+
+```
+[00000800 00000800 00000800 00000800 00000800 00000800 00000800 00000800]
+```
+
+The same program run under Qemu prints:
+
+```
+[08000800 00000000 08000800 00000000 08000800 00000000 08000800 00000000]
+```
+Additional information:
+I grabbed the latest sources, it seems there's bug in `target/loongarch/tcg/insn_trans/trans_vec.c.inc`, in function `vldi_get_value`.
+
+```c
+    case 1:
+        /* data: {2{16'0, imm[7:0], 8'0}} */
+        data = (t << 24) | (t << 8);
+        break;
+```
+
+There should be `(t << (8+32)) | t << 8`.