On Sun, Jul 22, 2012 at 5:39 PM, Uros Bizjak <ubiz...@gmail.com> wrote:
> The core of the problem was in fact the possibility of invalid
> addresses to sneak through various LEA patterns.
>
> Attached patch adds handling of SImode SUBREGs of addresses to generic
> code. This way, we can use one LEA pattern that accepts all valid
> address RTXes (modulo segment prefixed ones).
>
> 2012-07-22 Uros Bizjak <ubiz...@gmail.com>
>
> PR target/53961
> * config/i386/i386.md (*lea): New insn pattern.
> (*lea_1): Remove.
> (*lea<mode>_2): Ditto.
> (*lea_{3,4,5,6}_zext): Ditto.
> * config/i386/constraints.md (j): Remove address constraint.
> * config/i386/i386.c (ix86_decompose_address): Allow SImode subreg
> of an address.
> (ix86_print_operand_address): Handle SImode subreg of an address.
> (ix86_avoid_lea_for_add): Reject zero-extended addresses for now.
Following patch fixes an overlook ... we have to prevent zero extended
addresses in ix86_avoid_lea_for_addr, not ix86_avoid_lea_for_add. Spot
the difference in function names. ;)
Uros.
Index: config/i386/i386.c
===================================================================
--- config/i386/i386.c (revision 189767)
+++ config/i386/i386.c (working copy)
@@ -16864,11 +16864,6 @@ ix86_avoid_lea_for_add (rtx insn, rtx operands[])
unsigned int regno1 = true_regnum (operands[1]);
unsigned int regno2 = true_regnum (operands[2]);
- /* FIXME: Handle zero-extended addresses. */
- if (GET_CODE (operands[1]) == ZERO_EXTEND
- || GET_CODE (operands[1]) == AND)
- return false;
-
/* Check if we need to optimize. */
if (!TARGET_OPT_AGU || optimize_function_for_size_p (cfun))
return false;
@@ -16921,6 +16916,11 @@ ix86_avoid_lea_for_addr (rtx insn, rtx operands[])
struct ix86_address parts;
int ok;
+ /* FIXME: Handle zero-extended addresses. */
+ if (GET_CODE (operands[1]) == ZERO_EXTEND
+ || GET_CODE (operands[1]) == AND)
+ return false;
+
/* Check we need to optimize. */
if (!TARGET_OPT_AGU || optimize_function_for_size_p (cfun))
return false;
