------- Comment #2 from krebbel at gcc dot gnu dot org 2008-10-06 14:07 -------
Just to check whether the propagation of the conflicting hard regs in
ira_flatting really is the main problem I've tried the following patch. With
that patch the ira branch bootstraps on s390x.
Index: gcc/ira-build.c
===================================================================
--- gcc/ira-build.c.orig 2008-10-06 11:16:39.000000000 +0200
+++ gcc/ira-build.c 2008-10-06 14:44:57.000000000 +0200
@@ -2147,7 +2147,7 @@ ira_flattening (int max_regno_before_emi
ira_assert (ALLOCNO_CAP_MEMBER (parent_a) == NULL);
if (ALLOCNO_MEM_OPTIMIZED_DEST (a) != NULL)
mem_dest_p = true;
- if (propagate_p)
+ /* if (propagate_p)*/
{
if (!allocno_propagated_p [ALLOCNO_NUM (parent_a)])
COPY_HARD_REG_SET (ALLOCNO_TOTAL_CONFLICT_HARD_REGS (parent_a),
I think one reason why this problem does not occur more often and not on other
targets is that on S/390 r6 is used as argument register but is also call
saved!
The conflict between r52 and hard reg r6 is recorded for an instruction which
loads the argument for a function call. Since such an INSN is most likely more
or less directly followed by a call instruction the missing propagation of
conflicting hard regs is papered over by ira_build_conflicts. This function
always adds the call clobbered registers to the conflict sets of pseudos which
are live across function calls. For r6 on S/390 this does not happen what - at
least to my understanding - reveals the bug in ira_flattening.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37674
