i965: Disable try_emit_b2f_of_compare on Gen4-6.

The optimization relies on CMP setting the destination to 0, which is
equivalent to 0.0f.  However, early platforms only set the least
significant byte, leaving the other bits undefined.  So, we must disable
the optimization on those platforms.

Oddly, Sandybridge wasn't reported as broken.  The PRM states that it
only sets the LSB, but the internal documentation says that it follows
the IVB behavior.  Since it wasn't reported as broken, we believe it
really does follow the IVB behavior.

v2: Allow the optimization on Sandybridge (requested by Matt).

+32 piglits on Ironlake.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?=79963
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Matt Turner <mattst88@gmail.com>

diff --git a/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp b/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp
index eca5d0cb01ceabbf72c5f9ef18c7ac7cac0bc498..6e48be7485251776705bf4d4ac6f3c09715fb853 100644 (file)
--- a/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp
+++ b/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp
@@ -1135,6 +1135,13 @@ vec4_visitor::try_emit_mad(ir_expression *ir)
 bool
 vec4_visitor::try_emit_b2f_of_compare(ir_expression *ir)
 {
+   /* This optimization relies on CMP setting the destination to 0 when
+    * false.  Early hardware only sets the least significant bit, and
+    * leaves the other bits undefined.  So we can't use it.
+    */
+   if (brw->gen < 6)
+      return false;
+
    ir_expression *const cmp = ir->operands[0]->as_expression();
 
    if (cmp == NULL)