Igor noticed in https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64043#c13
a regression caused by the option streaming patch on LTO SPEC scores.
While looking for posisble causes I noticed that x86_tune_cost is not properly
restored at cfun change. Fixed by the following patch.
Igor, does it fix the regression by any chance?
Bootstrapped/regtested x86_64-linux, comitted.
* i386.opt (prefetch_sse): New targetsave.
* i386.c (ix86_function_specific_save): Save prefetch_sse.
(ix86_function_specific_restore): Restore prefetch_sse and initialize
ix86_cost/ix86_tune_cost.
Index: config/i386/i386.opt
===================================================================
--- config/i386/i386.opt (revision 219969)
+++ config/i386/i386.opt (working copy)
@@ -60,6 +60,10 @@ const char *x_ix86_tune_string
TargetSave
unsigned char schedule
+;; True if processor has SSE prefetch instruction.
+TargetSave
+unsigned char prefetch_sse
+
;; branch cost
TargetSave
unsigned char branch_cost
Index: config/i386/i386.c
===================================================================
--- config/i386/i386.c (revision 219969)
+++ config/i386/i386.c (working copy)
@@ -4466,6 +4466,7 @@ ix86_function_specific_save (struct cl_t
{
ptr->arch = ix86_arch;
ptr->schedule = ix86_schedule;
+ ptr->prefetch_sse = x86_prefetch_sse;
ptr->tune = ix86_tune;
ptr->branch_cost = ix86_branch_cost;
ptr->tune_defaulted = ix86_tune_defaulted;
@@ -4523,6 +4524,7 @@ ix86_function_specific_restore (struct g
ix86_arch = (enum processor_type) ptr->arch;
ix86_schedule = (enum attr_cpu) ptr->schedule;
ix86_tune = (enum processor_type) ptr->tune;
+ x86_prefetch_sse = ptr->prefetch_sse;
opts->x_ix86_branch_cost = ptr->branch_cost;
ix86_tune_defaulted = ptr->tune_defaulted;
ix86_arch_specified = ptr->arch_specified;
@@ -4553,6 +4555,13 @@ ix86_function_specific_restore (struct g
opts->x_ix86_tune_memset_strategy = ptr->x_ix86_tune_memset_strategy;
opts->x_ix86_tune_no_default = ptr->x_ix86_tune_no_default;
opts->x_ix86_veclibabi_type = ptr->x_ix86_veclibabi_type;
+ ix86_tune_cost = processor_target_table[ix86_tune].cost;
+ /* TODO: ix86_cost should be chosen at instruction or function granuality
+ so for cold code we use size_cost even in !optimize_size compilation. */
+ if (opts->x_optimize_size)
+ ix86_cost = &ix86_size_cost;
+ else
+ ix86_cost = ix86_tune_cost;
/* Recreate the arch feature tests if the arch changed */
if (old_arch != ix86_arch)
----- End forwarded message -----
