Hi!
The following testcase is rejected, because IPA-SRA decides to
turn bar.constprop call into bar.constprop.isra which returns void.
While there is no explicit lhs on the call, as it is a musttail call
the tailc pass checks if IPA-VRP returns singleton from that function
and the function returns the same value and in that case it still turns
it into a tail call. This can't work with IPA-SRA changing it into
void returning function though.
The following patch fixes this by forcing returning the original type
if there are musttail calls.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2025-03-27 Jakub Jelinek <ja...@redhat.com>
PR ipa/119484
* ipa-sra.cc (isra_analyze_call): Don't set m_return_ignored if
gimple_call_must_tail_p even if it doesn't have lhs.
* c-c++-common/pr119484.c: New test.
--- gcc/ipa-sra.cc.jj 2025-03-01 09:13:17.736075042 +0100
+++ gcc/ipa-sra.cc 2025-03-27 14:23:43.884552767 +0100
@@ -2242,7 +2242,11 @@ isra_analyze_call (cgraph_edge *cs)
BITMAP_FREE (analyzed);
}
}
- else
+ /* Don't set m_return_ignored for musttail calls. The tailc/musttail passes
+ compare the returned value against the IPA-VRP return value range if
+ it is a singleton, but if the call is changed to something which doesn't
+ return anything, it will always fail. */
+ else if (!gimple_call_must_tail_p (call_stmt))
csum->m_return_ignored = true;
}
--- gcc/testsuite/c-c++-common/pr119484.c.jj 2025-03-27 14:30:32.988976396
+0100
+++ gcc/testsuite/c-c++-common/pr119484.c 2025-03-27 14:31:17.074375485
+0100
@@ -0,0 +1,21 @@
+/* PR ipa/119484 */
+/* { dg-do compile { target musttail } } */
+/* { dg-options "-O2 -fdump-tree-optimized" } */
+/* { dg-final { scan-tree-dump-times "bar\[.a-z0-9]* \\\(\[^\n\r]*\\\);
\\\[tail call\\\] \\\[must tail call\\\]" 1 "optimized" } } */
+
+void foo (int);
+
+[[gnu::noinline]] static int
+bar (int x)
+{
+ foo (x);
+ return 0;
+}
+
+int
+baz (int x)
+{
+ if (x == 1)
+ [[gnu::musttail]] return bar (x);
+ return 0;
+}
Jakub
