Working on treeprop, I noticed that sometimes the SciMark2 score
got worse, but only on amd64.
Looking at the generated code, I found that the problem was that
a variable heavily used in deep nesting BBs (level 3) was left on
the stack, while variables used at level 1 were put in registers.
So I decided to tune is a bit.
With the attached patch, I get about a 25% speedup on amd64, but
also a 2 or 3% decrease on x86...
Using a less aggressive approach (put "(1 * (bb->nesting))" in the
increment expression), the amd64 speedup is only 10 or 15%, but
at least x86 does not get worse.
I'll go on looking at the generated code on both CPUs to find a
better tradeoff, but in the meanwhile if somebody wants to have a
look at what happens with other architectures...
A (very rough) script to automate the process is here:
http://primates.ximian.com/~massi/BENCHMARKS/benchmark.pl
and some explanation for it is here:
http://primates.ximian.com/~massi/blog/archive/2006/Mar-13.html
Ciao,
Massi
Index: mono/mono/mini/liveness.c
===================================================================
--- mono/mono/mini/liveness.c (revision 57888)
+++ mono/mono/mini/liveness.c (working copy)
@@ -11,6 +11,10 @@
#include "inssel.h"
#include "aliasing.h"
+//#define SPILL_COST_INCREMENT (1 << (bb->nesting << 1))
+#define SPILL_COST_INCREMENT (1 + (bb->nesting * 2))
+
+
//#define DEBUG_LIVENESS
#if SIZEOF_VOID_P == 8
@@ -116,7 +120,7 @@
if (!mono_bitset_test_fast (bb->kill_set, idx))
mono_bitset_set_fast (bb->gen_set, idx);
if (inst->ssa_op == MONO_SSA_LOAD)
- vi->spill_costs += 1 + (bb->nesting * 2);
+ vi->spill_costs += SPILL_COST_INCREMENT;
affected_variable = affected_variable->next;
}
@@ -138,7 +142,7 @@
update_live_range (cfg, idx, bb->dfn, inst_num);
mono_bitset_set_fast (bb->kill_set, idx);
if (inst->ssa_op == MONO_SSA_STORE)
- vi->spill_costs += 1 + (bb->nesting * 2);
+ vi->spill_costs += SPILL_COST_INCREMENT;
affected_variable = affected_variable->next;
}
Index: mono/mono/mini/linear-scan.c
===================================================================
--- mono/mono/mini/linear-scan.c (revision 57888)
+++ mono/mono/mini/linear-scan.c (working copy)
@@ -218,7 +218,7 @@
}
else {
if (cfg->verbose_level > 2)
- printf ("NOT REGVAR: %d\n", vmv->idx);
+ printf ("NOT REGVAR: %d (C%d)\n", vmv->idx, vmv->spill_costs);
}
}
}
_______________________________________________
Mono-devel-list mailing list
Mono-devel-list@lists.ximian.com
http://lists.ximian.com/mailman/listinfo/mono-devel-list
