There may be a slight imprecision in expand_compound_operation.
When it encounters a SIGN_EXTEND where it's already known that the
sign bit is zero, it may replace that with a ZERO_EXTEND (and
tries to simplify that further). However, the pattern is only
replaced if the new set_src_cost() is _lower_ than the old cost.
The patch changes that to "not higher than", assuming that the
ZERO_EXTEND form is generally preferrable unless there is a reason
to believe it's not (i.e. its cost is higher). The comment atop
this code block seems to support this:
/* Convert sign extension to zero extension, if we know that the high
bit is not set, as this is easier to optimize. It will be converted
back to cheaper alternative in make_extraction. */
On s390[x] this gets rid of some SIGN_EXTENDs completely.
(The patched code uses the cheaper of both replacement patterns.)
--
The patch hasn't got a lot of testing yet as I'd like to hear your
opinion on the patch first.
Ciao
Dominik ^_^ ^_^
--
Dominik Vogt
IBM Germany
gcc/ChangeLog-signextend-1
* combine.c (expand_compound_operation): Substitute ZERO_EXTEND for
SIGN_EXTEND if the costs are equal or lower.
Choose the cheapest replacement.
>From a1fca26fcd4df673a6cab0f72dd856acdfeac6d1 Mon Sep 17 00:00:00 2001
From: Dominik Vogt <v...@linux.vnet.ibm.com>
Date: Wed, 14 Dec 2016 13:05:30 +0100
Subject: [PATCH] combine: Replace sign_extend with zero_extend more often.
1)
expand_compound_operation() replaces sign_extend with zero_extend if it can
prove that the sign bit is zero. However, it does that only if the rtx cost
of the result is lower than that of the original expression. Allow the
substitution for equal costs too.
2)
Choose the cheaper replacement pattern instead of just the first one that is
cheaper than the original pattern.
---
gcc/combine.c | 13 ++++++-------
1 file changed, 6 insertions(+), 7 deletions(-)
diff --git a/gcc/combine.c b/gcc/combine.c
index 1456290..4ecfb4b 100644
--- a/gcc/combine.c
+++ b/gcc/combine.c
@@ -7090,13 +7090,12 @@ expand_compound_operation (rtx x)
/* Make sure this is a profitable operation. */
if (set_src_cost (x, mode, optimize_this_for_speed_p)
- > set_src_cost (temp2, mode, optimize_this_for_speed_p))
- return temp2;
- else if (set_src_cost (x, mode, optimize_this_for_speed_p)
- > set_src_cost (temp, mode, optimize_this_for_speed_p))
- return temp;
- else
- return x;
+ >= set_src_cost (temp2, mode, optimize_this_for_speed_p))
+ x = temp2;
+ if (set_src_cost (x, mode, optimize_this_for_speed_p)
+ >= set_src_cost (temp, mode, optimize_this_for_speed_p))
+ x = temp;
+ return x;
}
/* We can optimize some special cases of ZERO_EXTEND. */
--
2.3.0
