Hi,
With all the work that has recently gone in to ifcvt, I thought I'd revisit
the branch cost settings for Cortex-A57. After a run of experiments [1],
I found {1, 3} to be the sweet spot, giving a small set of performance
improvements across some popular benchmarks.
I'd therefore like to propose changing the branch cost to those numbers.
Patch bootstrapped tuning for Cortex-A57 with no issues. I'll revisit
the same for Cortex-A53.
OK?
Thanks,
James
[1]: {2, 2}, {2, 3}, {2, 4},
{1, 2}, {1, 3}, {1, 4},
{3, 3}, {3, 4},
{4, 4}, {4, 5},
{5, 5}, {5, 6}
---
2015-11-12 James Greenhalgh <[email protected]>
* config/aarch64/aarch64.c (cortexa57_branch_costs): New.
(cortexa57_tunings): Use it.
diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c
index 5ec7f08..96df9ed 100644
--- a/gcc/config/aarch64/aarch64.c
+++ b/gcc/config/aarch64/aarch64.c
@@ -338,6 +338,13 @@ static const struct cpu_branch_cost generic_branch_cost =
2 /* Unpredictable. */
};
+/* Branch costs for Cortex-A57. */
+static const struct cpu_branch_cost cortexa57_branch_cost =
+{
+ 1, /* Predictable. */
+ 3 /* Unpredictable. */
+};
+
static const struct tune_params generic_tunings =
{
&cortexa57_extra_costs,
@@ -393,7 +400,7 @@ static const struct tune_params cortexa57_tunings =
&cortexa57_addrcost_table,
&cortexa57_regmove_cost,
&cortexa57_vector_cost,
- &generic_branch_cost,
+ &cortexa57_branch_cost,
4, /* memmov_cost */
3, /* issue_rate */
(AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD
