i965/vs: Add support for splitting virtual GRFs.

This should improve our ability to register allocate without spilling.
Unfortuantely, due to the live variable analysis being ignorant of loops, we
still have register allocation failures on some programs.

v2: Add more context to the comment explaining the function.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> (v1)

diff --git a/src/mesa/drivers/dri/i965/brw_vec4.cpp b/src/mesa/drivers/dri/i965/brw_vec4.cpp
index 2941729c4a0976a474b922891d09d36dc6d391c6..e0b643242a364bfd2bbb0f903ef931bbf56c4d04 100644 (file)
--- a/src/mesa/drivers/dri/i965/brw_vec4.cpp
+++ b/src/mesa/drivers/dri/i965/brw_vec4.cpp
@@ -865,4 +865,64 @@ vec4_visitor::opt_compute_to_mrf()
    return progress;
 }
 
+/**
+ * Splits virtual GRFs requesting more than one contiguous physical register.
+ *
+ * We initially create large virtual GRFs for temporary structures, arrays,
+ * and matrices, so that the dereference visitor functions can add reg_offsets
+ * to work their way down to the actual member being accessed.
+ *
+ * Unlike in the FS visitor, though, we have no SEND messages that return more
+ * than 1 register.  We also don't do any array access in register space,
+ * which would have required contiguous physical registers.  Thus, all those
+ * large virtual GRFs can be split up into independent single-register virtual
+ * GRFs, making allocation and optimization easier.
+ */
+void
+vec4_visitor::split_virtual_grfs()
+{
+   int num_vars = this->virtual_grf_count;
+   int new_virtual_grf[num_vars];
+
+   memset(new_virtual_grf, 0, sizeof(new_virtual_grf));
+
+   /* Allocate new space for split regs.  Note that the virtual
+    * numbers will be contiguous.
+    */
+   for (int i = 0; i < num_vars; i++) {
+      if (this->virtual_grf_sizes[i] == 1)
+         continue;
+
+      new_virtual_grf[i] = virtual_grf_alloc(1);
+      for (int j = 2; j < this->virtual_grf_sizes[i]; j++) {
+         int reg = virtual_grf_alloc(1);
+         assert(reg == new_virtual_grf[i] + j - 1);
+         (void) reg;
+      }
+      this->virtual_grf_sizes[i] = 1;
+   }
+
+   foreach_list(node, &this->instructions) {
+      vec4_instruction *inst = (vec4_instruction *)node;
+
+      if (inst->dst.file == GRF &&
+         new_virtual_grf[inst->dst.reg] &&
+         inst->dst.reg_offset != 0) {
+        inst->dst.reg = (new_virtual_grf[inst->dst.reg] +
+                         inst->dst.reg_offset - 1);
+        inst->dst.reg_offset = 0;
+      }
+      for (int i = 0; i < 3; i++) {
+        if (inst->src[i].file == GRF &&
+            new_virtual_grf[inst->src[i].reg] &&
+            inst->src[i].reg_offset != 0) {
+           inst->src[i].reg = (new_virtual_grf[inst->src[i].reg] +
+                               inst->src[i].reg_offset - 1);
+           inst->src[i].reg_offset = 0;
+        }
+      }
+   }
+   this->live_intervals_valid = false;
+}
+
 } /* namespace brw */

diff --git a/src/mesa/drivers/dri/i965/brw_vec4.h b/src/mesa/drivers/dri/i965/brw_vec4.h
index dce3c89575d2aef4728277a3ab225d09d6a643de..4fdede3ab26bb8b9f7a3dd9661b274d19cd62aa4 100644 (file)
--- a/src/mesa/drivers/dri/i965/brw_vec4.h
+++ b/src/mesa/drivers/dri/i965/brw_vec4.h
@@ -323,6 +323,7 @@ public:
    void split_uniform_registers();
    void pack_uniform_registers();
    void calculate_live_intervals();
+   void split_virtual_grfs();
    bool dead_code_eliminate();
    bool virtual_grf_interferes(int a, int b);
    bool opt_copy_propagation();

diff --git a/src/mesa/drivers/dri/i965/brw_vec4_emit.cpp b/src/mesa/drivers/dri/i965/brw_vec4_emit.cpp
index 5a941d90c309e26685e85b124c3e4fc75aebf4a1..479b0a607aae1ab51ee3b29b33cb69de57a9711f 100644 (file)
--- a/src/mesa/drivers/dri/i965/brw_vec4_emit.cpp
+++ b/src/mesa/drivers/dri/i965/brw_vec4_emit.cpp
@@ -794,6 +794,7 @@ vec4_visitor::run()
    move_uniform_array_access_to_pull_constants();
    pack_uniform_registers();
    move_push_constants_to_pull_constants();
+   split_virtual_grfs();
 
    bool progress;
    do {