summary refs log tree commit diff stats
path: root/results/classifier/semantic-bugs-usermode/test/1820686
diff options
context:
space:
mode:
Diffstat (limited to 'results/classifier/semantic-bugs-usermode/test/1820686')
-rw-r--r--results/classifier/semantic-bugs-usermode/test/182068625
1 files changed, 25 insertions, 0 deletions
diff --git a/results/classifier/semantic-bugs-usermode/test/1820686 b/results/classifier/semantic-bugs-usermode/test/1820686
new file mode 100644
index 000000000..f7fd19f6d
--- /dev/null
+++ b/results/classifier/semantic-bugs-usermode/test/1820686
@@ -0,0 +1,25 @@
+instruction: 0.949
+device: 0.824
+network: 0.809
+other: 0.742
+socket: 0.724
+semantic: 0.691
+vnc: 0.683
+mistranslation: 0.678
+graphic: 0.598
+boot: 0.498
+KVM: 0.321
+assembly: 0.305
+
+risc-v: 'c.unimp' instruction decoded as 'c.addi4spn fp, 0'
+
+QEMU 3.1 incorrectly decodes the "c.unimp" instruction (opcode 0x0000) as an "addi4spn fp, 0" when either of the two following bytes are non-zero. This is because the ctx->opcode value used when decoding the instruction is actually filled with a 32-bit load (to handle normal uncompressed instructions) but when a compressed instruction is found only the low 16 bits are valid. Other reserved/illegal bit patterns with the addi4spn opcode are also incorrectly decoded.
+
+I believe that the switch to decodetree on master happened to fix this issue, but hopefully it is helpful to have this recorded somewhere. I've included a simple one line patch if anyone wants to backport this.
+
+
+
+Thanks.  If you spin a full patch (ie, "git commit -s" and then "git show") I can drop it on riscv-qemu-3.1, our backports branch.  Otherwise hopefully we got the bug via the decodetree conversion.
+
+Since this bug isn't present in the decodetree version of the riscv decoder, we might as well just close this as fix-released; we won't be doing more point-releases of QEMU versions as old as 3.1.
+