summary refs log tree commit diff stats
path: root/results/classifier/phi4:14b/output/manual-review/1574346
diff options
context:
space:
mode:
Diffstat (limited to 'results/classifier/phi4:14b/output/manual-review/1574346')
-rw-r--r--results/classifier/phi4:14b/output/manual-review/157434615
1 files changed, 15 insertions, 0 deletions
diff --git a/results/classifier/phi4:14b/output/manual-review/1574346 b/results/classifier/phi4:14b/output/manual-review/1574346
new file mode 100644
index 00000000..59fce0e2
--- /dev/null
+++ b/results/classifier/phi4:14b/output/manual-review/1574346
@@ -0,0 +1,15 @@
+
+
+
+TCG: mov to segment register is incorrectly emulated for AMD CPUs
+
+In TCG mode, the effect of:
+
+xorl %eax, %eax
+movl %eax, %gs
+
+is to mark the GS segment unusable and set its base to zero.  After doing this, reading MSR_GS_BASE will return zero and using a GS prefix in long mode will treat the GS base as zero.
+
+This is correct for Intel CPUs but is incorrect for AMD CPUs.  On an AMD CPU, writing 0 to %gs using mov, pop, or (I think) lgs will leave the base unchanged.
+
+To make it easier to use TCG to validate behavior on different CPUs, please consider changing the TCG behavior to match actual CPU behavior when emulating an AMD CPU.
\ No newline at end of file