diff mbox series

[2/2] Add testcase for backward threading from never-executed entry edges [PR105679]

Message ID 20260902120747.2705241-4-aldy@quesejoda.com
State New
Headers show
Series [1/2] Add testcases for backward threading to never-executed edges [PR106495] | expand

Commit Message

Aldy Hernandez Sept. 2, 2026, 12:07 p.m. UTC 
r13-1891-g49ba4fdeb648c1 made the backward threader reject paths
whose entry edge is probably never executed, but landed without a
testcase.  Add one retroactively.

[Also AI generated.]

Assisted-by: Claude Fable 5 (Anthropic)

OK?

gcc/testsuite/ChangeLog:

	PR tree-optimization/105679
	PR tree-optimization/126906
	* gcc.dg/tree-ssa/pr105679.c: New test.
---
 gcc/testsuite/gcc.dg/tree-ssa/pr105679.c | 33 ++++++++++++++++++++++++
 1 file changed, 33 insertions(+)
 create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/pr105679.c

Comments

Jeff Law Sept. 2, 2026, 12:48 p.m. UTC | #1 
On 9/2/26 6:07 AM, Aldy Hernandez wrote:
> r13-1891-g49ba4fdeb648c1 made the backward threader reject paths
> whose entry edge is probably never executed, but landed without a
> testcase.  Add one retroactively.
>
> [Also AI generated.]
>
> Assisted-by: Claude Fable 5 (Anthropic)
>
> OK?
>
> gcc/testsuite/ChangeLog:
>
> 	PR tree-optimization/105679
> 	PR tree-optimization/126906
> 	* gcc.dg/tree-ssa/pr105679.c: New test.
It took me a minute to realize that the sanitizer is what's inserting 
the irq > 31 check.  I might have tried to guide towards exposing the 
code generated by sanitization to make the testcase clearer, but it's 
not worth it IMHO to do it now.

OK.

jeff
diff mbox series

Patch

diff --git a/gcc/testsuite/gcc.dg/tree-ssa/pr105679.c b/gcc/testsuite/gcc.dg/tree-ssa/pr105679.c
new file mode 100644
index 00000000000..5f7b6cf0d1d
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/tree-ssa/pr105679.c
@@ -0,0 +1,33 @@ 
+/* PR tree-optimization/105679 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -Warray-bounds -fno-strict-overflow -fsanitize=shift -fsanitize-coverage=trace-pc -fdump-tree-threadfull1-details" } */
+
+/* Reduced with cvise from attachment 53010: the Linux kernel hit spurious
+   -Warray-bounds warnings with -fsanitize=shift.  The sanitizer's
+   shift-out-of-bounds check for 1 << irq gives the irq > 31 branch a precise
+   never-executed count; the backward threader nevertheless used that block as
+   a thread entry, isolating a never-executed path on which irq > 31 indexes
+   the 2-element array and drawing the bogus warning.  The bare shift statement
+   only exists to feed the sanitizer.
+
+   r13-1891 (the PR105679 fix) rejects paths whose entry edge is probably never
+   executed.  This test FAILs before that commit and PASSes with it.  */
+
+struct many_objects
+{
+  int array[2];
+} instance;
+
+int entry;
+
+int
+work (unsigned int irq)
+{
+  1 << irq;
+  entry = instance.array[irq];	/* { dg-bogus "above array bounds" } */
+  if (irq)
+    return 0;
+  return 1;
+}
+
+/* { dg-final { scan-tree-dump "path entry is probably never executed" "threadfull1" } } */