The patch well fix the adobe_cpp performance regression on the int8_t
type. But the same degradation exists on uint8_t type, which is not
fixed by the patch referenced in PR53676.
With the signed version, the code:
result_5 = (signed char) ((int) result_2 + 2)
is now well narrowed to:
result_5 = (signed char) ((unsigned char) result_2 + 2)
But with the unsigned version:
result_5 = (unsigned char) ((int) result_2 + 2)
is not narrowed to:
result_5 = (unsigned char) ((unsigned char) result_2 + 2)
As a consequence, result_5 is not detected as polynomial, and the loop
is not removed.
Tested on SH target. bootstrap + regression tests on
x86_64-unknown-linux-gnu in progress.
Regards,
Laurent
--- gcc.orig/gcc/tree-chrec.c 2013-04-22 13:20:03.0 +0200
+++ gcc/gcc/tree-chrec.c 2013-04-22 13:22:51.0 +0200
@@ -1346,13 +1346,12 @@
res = fold_build2 (TREE_CODE (chrec), type,
fold_convert (type, TREE_OPERAND (chrec, 0)),
fold_convert (type, TREE_OPERAND (chrec, 1)));
- /* Similar perform the trick that (signed char)((int)x + 2) can be
- narrowed to (signed char)((unsigned char)x + 2). */
+ /* Similar perform the trick that ([un]signed char)((int)x + 2) can be
+ narrowed to ([un]signed char)((unsigned char)x + 2). */
else if (use_overflow_semantics
&& TREE_CODE (chrec) == POLYNOMIAL_CHREC
&& TREE_CODE (ct) == INTEGER_TYPE
&& TREE_CODE (type) == INTEGER_TYPE
- && TYPE_OVERFLOW_UNDEFINED (type)
&& TYPE_PRECISION (type) < TYPE_PRECISION (ct))
{
tree utype = unsigned_type_for (type);
--- gcc.orig/gcc/tree-scalar-evolution.c 2013-04-22 13:20:12.0 +0200
+++ gcc/gcc/tree-scalar-evolution.c 2013-04-22 13:21:43.0 +0200
@@ -1767,7 +1767,6 @@
if (TREE_CODE (type) == INTEGER_TYPE
&& TREE_CODE (TREE_TYPE (rhs1)) == INTEGER_TYPE
&& TYPE_PRECISION (type) < TYPE_PRECISION (TREE_TYPE (rhs1))
- && TYPE_OVERFLOW_UNDEFINED (type)
&& TREE_CODE (rhs1) == SSA_NAME
&& (def = SSA_NAME_DEF_STMT (rhs1))
&& is_gimple_assign (def)