The xpassing change in generated code was as follows, at
r14-9788-gb7bd2ec73d66f7 (where I locally applied a revert
to verify that this suspect was the cause). That was so
much of an improvement that I had to share it! Worth the
testsuite churn anyway. :)
Segher, if you end up reverting r14-9692-g839bc42772ba7a (as
unfortunately seems not unlikely), then please also revert this
commit: r14-9799-g4c8b3600c4856f7915281ae3ff4d97271c83a540.
--- pr93372-2.s.pre 2024-04-05 01:49:47.985685902 +0200
+++ pr93372-2.s.post 2024-04-05 01:42:02.296489730 +0200
@@ -5,12 +5,9 @@
.global _f
.type _f, @function
_f:
- move.d $r10,$r9
- sub.d $r11,$r9
- cmp.d $r11,$r10
- seq $r10
- move.d $r10,[$r12]
- cmpq 0,$r9
+ sub.d $r11,$r10
+ seq $r9
+ move.d $r9,[$r12]
ret
sge $r10
-- >8 --
After r14-9692-g839bc42772ba7a, a sequence that actually
looks optimal is now emitted, observed at
r14-9788-gb7bd2ec73d66f7. This caused an XPASS for this
test. While adjusting the test, better also guard it
against regressions by checking that there are no redundant
move insns.
That's the only test that's improved to the point of
affecting test-patterns. E.g. pr93372-5.c (which references
pr93372-2.c) is also improved, though it retains a redundant
compare insn. (PR 93372 was about regressions from the cc0
representation; not further improvement like here, thus it's
not tagged. Though, I did not double-check whether this
actually *was* a regression from cc0.)
* gcc.target/cris/pr93372-2.c: Tweak scan-assembler
checks to cover recent combine improvement.
---
gcc/testsuite/gcc.target/cris/pr93372-2.c | 15 ++++++++-------
1 file changed, 8 insertions(+), 7 deletions(-)
diff --git a/gcc/testsuite/gcc.target/cris/pr93372-2.c
b/gcc/testsuite/gcc.target/cris/pr93372-2.c
index 912069c018d5..2ef6471a990b 100644
--- a/gcc/testsuite/gcc.target/cris/pr93372-2.c
+++ b/gcc/testsuite/gcc.target/cris/pr93372-2.c
@@ -1,19 +1,20 @@
/* Check that eliminable compare-instructions are eliminated. */
/* { dg-do compile } */
/* { dg-options "-O2" } */
-/* { dg-final { scan-assembler-not "\tcmp|\ttest" { xfail *-*-* } } } */
-/* { dg-final { scan-assembler-not "\tnot" { xfail cc0 } } } */
-/* { dg-final { scan-assembler-not "\tlsr" { xfail cc0 } } } */
+/* { dg-final { scan-assembler-not "\tcmp|\ttest" } } */
+/* { dg-final { scan-assembler-not "\tnot" } } */
+/* { dg-final { scan-assembler-not "\tlsr" } } */
+/* We should get just one move, storing the result into *d. */
+/* { dg-final { scan-assembler-times "\tmove" 1 } } */
int f(int a, int b, int *d)
{
int c = a - b;
- /* Whoops! We get a cmp.d with the original operands here. */
+ /* We used to get a cmp.d with the original operands here. */
*d = (c == 0);
- /* Whoops! While we don't get a test.d for the result here for cc0,
- we get a sequence of insns: a move, a "not" and a shift of the
- subtraction-result, where a simple "spl" would have done. */
+ /* We used to get a suboptimal sequence, but now we get the optimal "sge"
+ (a.k.a "spl") re-using flags from the subtraction. */
return c >= 0;
}
--
2.30.2
