1 files changed, 29 insertions, 0 deletions
diff --git a/results/classifier/semantic-bugs/826 b/results/classifier/semantic-bugs/826
new file mode 100644
index 000000000..3ef3962e4
--- /dev/null
+++ b/results/classifier/semantic-bugs/826
@@ -0,0 +1,29 @@
+instruction: 0.989
+graphic: 0.822
+device: 0.639
+mistranslation: 0.592
+network: 0.489
+other: 0.438
+assembly: 0.432
+semantic: 0.404
+vnc: 0.261
+socket: 0.178
+boot: 0.132
+KVM: 0.015
+
+AArch64 SVE2 LDNT1SB (vector plus scalar) load address incorrectly calculated
+Description of problem:
+During execution of the following SVE2 instruction:
+`ldnt1sb {z6.d}, p3/z, [z14.d, x9]`
+with the following register state:
+```
+(gdb) p $p3
+$1 = {0x7, 0x0, 0x74, 0x0, 0x43, 0x0, 0x29, 0x0, 0x47, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x47, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x66, 0xe4, 0x64, 0x0, 0x0, 0x0, 0x0, 0x0, 0x20, 0x11, 0x31, 0x1, 0x0, 0x0, 0x0, 0x0, 0x20, 0x11, 0x31, 0x1, 0x0, 0x0, 0x0, 0x0, 0xb0, 0x8b, 0x49, 0x34, 0xfc, 0x7f, 0x0, 0x0, 0xe0, 0x71, 0x30, 0x1, 0x0, 0x0, 0x0, 0x0}
+(gdb) p $z14.d.u
+$2 = {0x3bdeaa30, 0x3bdeaa33, 0x3bdeaa36, 0x3bdeaa39, 0x3bdeaa3c, 0x3bdeaa3f, 0x3bdeaa42, 0x3bdeaa45}
+(gdb) p $x9
+$3 = 0x0
+```
+QEMU produces a data abort due to an address fault on address `0x5EE45E4E`, which it clearly should not have tried to load.
+Additional information:
+A quick look at the implementation of the LDNT1SB instruction in QEMU points to the following commit: https://gitlab.com/qemu-project/qemu/-/commit/cf327449816d5643106445420a0b06b0f38d4f01 which simply redirects to SVE's LD1SB handler. As these instructions use a new flavor of SVE scatter/gather loads (vector plus scalar) which SVE LD1SB does not support, I wonder if the LD1SB handler simply decodes it as the wrong instruction and treats it as a (scalar plus vector) instruction, which LD1SB does support, but whose address calculation is completely different.