https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70359
--- Comment #42 from Aldy Hernandez <aldyh at gcc dot gnu.org> ---
(In reply to Jeffrey A. Law from comment #36)
> WRT the division removal. That seems so profitable that a slight increase
> in codesize is warranted. So if we fix the other issue and the source of
> the remaining codesize regressions is the removal of the division, I would
> consider this BZ resolved.
Richi. Jeff.
Limiting the single_use with optimize_size as well may give us the best of both
worlds. Would you like me to post the [untested] patch below upstream? With
this patch code size is even smaller than GCC 5.3.
I really don't care. Actually, I'd prefer to do nothing and close the PR ;-).
Up to y'all.
diff --git a/gcc/match.pd b/gcc/match.pd
index f61c4d7440a..5d29bf62dc9 100644
--- a/gcc/match.pd
+++ b/gcc/match.pd
@@ -1290,11 +1290,12 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
/* X / C1 op C2 into a simple range test. */
(for cmp (simple_comparison)
(simplify
- (cmp (trunc_div:s @0 INTEGER_CST@1) INTEGER_CST@2)
+ (cmp (trunc_div:s@3 @0 INTEGER_CST@1) INTEGER_CST@2)
(if (INTEGRAL_TYPE_P (TREE_TYPE (@0))
&& integer_nonzerop (@1)
&& !TREE_OVERFLOW (@1)
- && !TREE_OVERFLOW (@2))
+ && !TREE_OVERFLOW (@2)
+ && (!optimize_size || single_use (@3)))
(with { tree lo, hi; bool neg_overflow;
enum tree_code code = fold_div_compare (cmp, @1, @2, &lo, &hi,
&neg_overflow); }
@@ -1456,9 +1457,10 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
(for cmp (eq ne)
ocmp (lt ge)
(simplify
- (cmp (trunc_div @0 @1) integer_zerop)
+ (cmp (trunc_div@2 @0 @1) integer_zerop)
(if (TYPE_UNSIGNED (TREE_TYPE (@0))
- && (VECTOR_TYPE_P (type) || !VECTOR_TYPE_P (TREE_TYPE (@0))))
+ && (VECTOR_TYPE_P (type) || !VECTOR_TYPE_P (TREE_TYPE (@0)))
+ && (!optimize_size || single_use (@2)))
(ocmp @0 @1))))
/* X == C - X can never be true if C is odd. */
