This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[PATCH] Address one thing in PR32698
- From: Richard Guenther <rguenther at suse dot de>
- To: gcc-patches at gcc dot gnu dot org
- Date: Mon, 9 Jul 2007 21:41:45 +0200 (CEST)
- Subject: [PATCH] Address one thing in PR32698
This patch makes us consistently canonicalize A * C +- B * C
instead of depending on the actual constant. So for array
accesses like
p[i + 1] + p[i + 2] + p[i + 3]
we produced a mix like
(*(p + (D.1660 + 8)) + *(p + (i + 1) * 4)) + *(p + (D.1660 + 12))
now we match the original input and create
(*(p + (i + 2) * 4) + *(p + (i + 1) * 4)) + *(p + (i + 3) * 4)
it doesn't change code generation on x86_64, on i686 we use
one register less.
As this patch only changes fold it doesn't address any deficiency
selecting a proper addressing mode (which shouldn't be done
out of context and is done later by both ivopts (if in loops) and
fwprop at rtl). It merely corrects a mistake I made at the
point I invented the function doing the canonicalization.
Bootstrapped and tested on x86_64-unknown-linux-gnu, applied to
mainline.
Richard.
2007-07-09 Richard Guenther <rguenther@suse.de>
PR middle-end/32698
* fold-const.c (fold_plusminus_mult_expr): Move constant
arguments second to allow decomposing.
Index: fold-const.c
===================================================================
*** fold-const.c (revision 126376)
--- fold-const.c (working copy)
*************** fold_plusminus_mult_expr (enum tree_code
*** 7122,7127 ****
--- 7122,7132 ----
arg00 = TREE_OPERAND (arg0, 0);
arg01 = TREE_OPERAND (arg0, 1);
}
+ else if (TREE_CODE (arg0) == INTEGER_CST)
+ {
+ arg00 = build_one_cst (type);
+ arg01 = arg0;
+ }
else
{
arg00 = arg0;
*************** fold_plusminus_mult_expr (enum tree_code
*** 7132,7137 ****
--- 7137,7147 ----
arg10 = TREE_OPERAND (arg1, 0);
arg11 = TREE_OPERAND (arg1, 1);
}
+ else if (TREE_CODE (arg1) == INTEGER_CST)
+ {
+ arg10 = build_one_cst (type);
+ arg11 = arg1;
+ }
else
{
arg10 = arg1;