Frederic Riss sent me IRA dump files for an in-house GCC back-end which
helped to find a bug in coloring in presence of multi-register
allocnos. The bug resulted in making allocno colorable too earlier and
as a consequence in worse code. The problem was in preventing
processing of all conflicting objects of an allocno although the impact
of each object was taken as 1. It was my mistake in merging code for
IRA coloring without cover classes and Bernd's work on multi-object
allocnos.
The following patch changes SPEC2000 code on x86 (32-bit mode) for vpr,
gcc, crafty, parser, and vortex and improves scores (up to 0.7% for
crafty) and decrease code size (up to 0.5% for crafty) for these
benchmarks. Frederic Riss also reported considerable improvements for
his back-end.
The patch was successfully bootstrapped on x86/x86-64 and committed as
rev. 185971.
2012-03-29 Vladimir Makarov <vmaka...@redhat.com>
* ira-color.c (setup_left_conflict_sizes_p): Process all
conflicting objects.
Index: ira-color.c
===================================================================
--- ira-color.c (revision 185967)
+++ ira-color.c (working copy)
@@ -821,7 +821,6 @@ setup_left_conflict_sizes_p (ira_allocno
node_preorder_num = node->preorder_num;
COPY_HARD_REG_SET (node_set, node->hard_regs->set);
node_check_tick++;
- curr_allocno_process++;
for (k = 0; k < nobj; k++)
{
ira_object_t obj = ALLOCNO_OBJECT (a, k);
@@ -838,12 +837,10 @@ setup_left_conflict_sizes_p (ira_allocno
conflict_data = ALLOCNO_COLOR_DATA (conflict_a);
if (! ALLOCNO_COLOR_DATA (conflict_a)->in_graph_p
- || conflict_data->last_process == curr_allocno_process
|| ! hard_reg_set_intersect_p (profitable_hard_regs,
conflict_data
->profitable_hard_regs))
continue;
- conflict_data->last_process = curr_allocno_process;
conflict_node = conflict_data->hard_regs_node;
COPY_HARD_REG_SET (conflict_node_set, conflict_node->hard_regs->set);
if (hard_reg_set_subset_p (node_set, conflict_node_set))
