https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113126
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |crazylht at gmail dot com
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
diff --git a/gcc/match.pd b/gcc/match.pd
index 876a9d1c06e..abbd03742f9 100644
--- a/gcc/match.pd
+++ b/gcc/match.pd
@@ -6792,11 +6792,12 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
&& exact_real_truncate (TYPE_MODE (double_type_node), &orig))
type1 = double_type_node;
}
- tree newtype
- = (element_precision (TREE_TYPE (@00)) > element_precision (type1)
- ? TREE_TYPE (@00) : type1);
+ tree newtype
+ = (element_precision (TREE_TYPE (@00)) > element_precision (type1)
+ ? TREE_TYPE (@00) : type1);
}
- (if (element_precision (TREE_TYPE (@0)) > element_precision (newtype))
+ (if (element_precision (TREE_TYPE (@0)) > element_precision (newtype)
+ && is_truth_type_for (newtype, type))
(cmp (convert:newtype @00) (convert:newtype @10))))))))
as I thought with AVX512VL we can handle this, but for V2SFmode we get
V2SImode as mask_mode. And changing the test to V4SF/V4DFmode reveals
that we don't use float-extend (I'd have to trace vector lowering and
forwprop). There might be an opportunity to improve what we do
for convertvector.
But it fixes the testcase.
