i965/fs: Improve register coalescing interference check.

I always thought that the is_control_flow() -> return false check was a
bad hack, and some previous attempts to remove it have failed and have
been reverted.

The previous two patches fix some problems that caused register
coalescing to not notice some interference between registers, which the
is_control_flow() check apparently works around.

With that fixed, we can calculate interference more accurately.

total instructions in shared programs: 6261319 -> 6257917 (-0.05%)
instructions in affected programs:     346282 -> 342880 (-0.98%)
helped:                                1552

Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

diff --git a/src/mesa/drivers/dri/i965/brw_fs_register_coalesce.cpp b/src/mesa/drivers/dri/i965/brw_fs_register_coalesce.cpp
index 218cc614e6d88e86684b9e6c88e8cfac3883c3f4..452aee5a7698871a8dddcc4ced02b2bd7502f2ee 100644 (file)
--- a/src/mesa/drivers/dri/i965/brw_fs_register_coalesce.cpp
+++ b/src/mesa/drivers/dri/i965/brw_fs_register_coalesce.cpp
@@ -110,27 +110,30 @@ can_coalesce_vars(brw::fs_live_variables *live_intervals,
        (end_from > end_to && start_to < start_from))
       return false;
 
-   int start_ip = MIN2(start_to, start_from);
+   /* Check for a write to either register in the intersection of their live
+    * ranges.
+    */
+   int start_ip = MAX2(start_to, start_from);
+   int end_ip = MIN2(end_to, end_from);
    int scan_ip = -1;
 
    foreach_block_and_inst(block, fs_inst, scan_inst, cfg) {
       scan_ip++;
 
+      /* Ignore anything before the intersection of the live ranges */
       if (scan_ip < start_ip)
          continue;
 
-      if (scan_inst->is_control_flow())
-         return false;
-
-      if (scan_ip <= live_intervals->start[var_to])
+      /* Ignore the copying instruction itself */
+      if (scan_inst == inst)
          continue;
 
-      if (scan_ip > live_intervals->end[var_to])
-         return true;
+      if (scan_ip > end_ip)
+         return true; /* registers do not interfere */
 
       if (scan_inst->overwrites_reg(inst->dst) ||
           scan_inst->overwrites_reg(inst->src[0]))
-         return false;
+         return false; /* registers interfere */
    }
 
    return true;