The following patch fixes a bug in LRA inheritance found on compiling a
big file on s390x. Unfortunately, the bug is very hard to reproduce and
the test is too big and can not be included. But even if it were
included, checking a correct code generation would be a problem too.
The bug was in undoing inheritance transformation. Not all inheritance
pseudo occurrences were removed which resulted in reading a garbage from
memory.
The patch was successfully bootstrapped and tested on x86-64.
Committed to the trunk as rev. 212464.
2014-07-11 Vladimir Makarov <vmaka...@redhat.com>
* lra-constraints.c (remove_inheritance_pseudos): Process
destination pseudo too.
Index: lra-constraints.c
===================================================================
--- lra-constraints.c (revision 212460)
+++ lra-constraints.c (working copy)
@@ -5697,6 +5697,20 @@ remove_inheritance_pseudos (bitmap remov
SUBREG_REG (SET_SRC (set)) = SET_SRC (prev_set);
else
SET_SRC (set) = SET_SRC (prev_set);
+ /* As we are finishing with processing the insn
+ here, check the destination too as it might
+ inheritance pseudo for another pseudo. */
+ if (bitmap_bit_p (remove_pseudos, dregno)
+ && bitmap_bit_p (&lra_inheritance_pseudos, dregno)
+ && (restore_regno
+ = lra_reg_info[dregno].restore_regno) >= 0)
+ {
+ if (GET_CODE (SET_DEST (set)) == SUBREG)
+ SUBREG_REG (SET_DEST (set))
+ = regno_reg_rtx[restore_regno];
+ else
+ SET_DEST (set) = regno_reg_rtx[restore_regno];
+ }
lra_push_insn_and_update_insn_regno_info (curr_insn);
lra_set_used_insn_alternative_by_uid
(INSN_UID (curr_insn), -1);
