aboutsummaryrefslogtreecommitdiff
path: root/mesalib/src/mesa/program/register_allocate.c
diff options
context:
space:
mode:
authormarha <marha@users.sourceforge.net>2013-04-08 08:20:47 +0200
committermarha <marha@users.sourceforge.net>2013-04-08 08:20:47 +0200
commita4505330e3469b6956348812ff2d6c24f812a8ba (patch)
treecf2b88c771188580ab585f763d2e7a0309ecab26 /mesalib/src/mesa/program/register_allocate.c
parent44fb3db38c5148666f62c78e10fc37bb20ed5c5c (diff)
parent95fb19d661154ba8cfc6c793a0daa25657294b3b (diff)
downloadvcxsrv-a4505330e3469b6956348812ff2d6c24f812a8ba.tar.gz
vcxsrv-a4505330e3469b6956348812ff2d6c24f812a8ba.tar.bz2
vcxsrv-a4505330e3469b6956348812ff2d6c24f812a8ba.zip
Merge remote-tracking branch 'origin/released'
* origin/released: fontconfig libXau mesa xserver xkeyboard-config git update 8 Apr 2013
Diffstat (limited to 'mesalib/src/mesa/program/register_allocate.c')
-rw-r--r--mesalib/src/mesa/program/register_allocate.c31
1 files changed, 28 insertions, 3 deletions
diff --git a/mesalib/src/mesa/program/register_allocate.c b/mesalib/src/mesa/program/register_allocate.c
index a9064c38c..2c826fc66 100644
--- a/mesalib/src/mesa/program/register_allocate.c
+++ b/mesalib/src/mesa/program/register_allocate.c
@@ -70,6 +70,7 @@
* this during ra_set_finalize().
*/
+#include <stdbool.h>
#include <ralloc.h>
#include "main/imports.h"
@@ -93,6 +94,8 @@ struct ra_regs {
struct ra_class **classes;
unsigned int class_count;
+
+ bool round_robin;
};
struct ra_class {
@@ -185,6 +188,22 @@ ra_alloc_reg_set(void *mem_ctx, unsigned int count)
return regs;
}
+/**
+ * The register allocator by default prefers to allocate low register numbers,
+ * since it was written for hardware (gen4/5 Intel) that is limited in its
+ * multithreadedness by the number of registers used in a given shader.
+ *
+ * However, for hardware without that restriction, densely packed register
+ * allocation can put serious constraints on instruction scheduling. This
+ * function tells the allocator to rotate around the registers if possible as
+ * it allocates the nodes.
+ */
+void
+ra_set_allocate_round_robin(struct ra_regs *regs)
+{
+ regs->round_robin = true;
+}
+
static void
ra_add_conflict_list(struct ra_regs *regs, unsigned int r1, unsigned int r2)
{
@@ -436,16 +455,19 @@ GLboolean
ra_select(struct ra_graph *g)
{
int i;
+ int start_search_reg = 0;
while (g->stack_count != 0) {
- unsigned int r;
+ unsigned int ri;
+ unsigned int r = -1;
int n = g->stack[g->stack_count - 1];
struct ra_class *c = g->regs->classes[g->nodes[n].class];
/* Find the lowest-numbered reg which is not used by a member
* of the graph adjacent to us.
*/
- for (r = 0; r < g->regs->count; r++) {
+ for (ri = 0; ri < g->regs->count; ri++) {
+ r = (start_search_reg + ri) % g->regs->count;
if (!c->regs[r])
continue;
@@ -461,12 +483,15 @@ ra_select(struct ra_graph *g)
if (i == g->nodes[n].adjacency_count)
break;
}
- if (r == g->regs->count)
+ if (ri == g->regs->count)
return GL_FALSE;
g->nodes[n].reg = r;
g->nodes[n].in_stack = GL_FALSE;
g->stack_count--;
+
+ if (g->regs->round_robin)
+ start_search_reg = r + 1;
}
return GL_TRUE;