prog_optimize: Add support for saturates to _mesa_merge_mov_into_inst.

This fixes the remaining regression from ff_fragment_shader in Mesa IR
instruction count, to now being a 1.9% win overall.

diff --git a/src/mesa/program/prog_optimize.c b/src/mesa/program/prog_optimize.c
index f4a7a638d5f71c71fb463669ccd7367272f643a4..3340ce0498b7967ae00087c862e25632be06c23f 100644 (file)
--- a/src/mesa/program/prog_optimize.c
+++ b/src/mesa/program/prog_optimize.c
@@ -472,8 +472,7 @@ can_downward_mov_be_modifed(const struct prog_instruction *mov)
       mov->SrcReg[0].HasIndex2 == 0 &&
       mov->SrcReg[0].RelAddr2 == 0 &&
       mov->DstReg.RelAddr == 0 &&
-      mov->DstReg.CondMask == COND_TR &&
-      mov->SaturateMode == SATURATE_OFF;
+      mov->DstReg.CondMask == COND_TR;
 }
 
 
@@ -482,7 +481,8 @@ can_upward_mov_be_modifed(const struct prog_instruction *mov)
 {
    return
       can_downward_mov_be_modifed(mov) &&
-      mov->DstReg.File == PROGRAM_TEMPORARY;
+      mov->DstReg.File == PROGRAM_TEMPORARY &&
+      mov->SaturateMode == SATURATE_OFF;
 }
 
 
@@ -657,6 +657,8 @@ _mesa_merge_mov_into_inst(struct prog_instruction *inst,
    if (mask != (inst->DstReg.WriteMask & mask))
       return GL_FALSE;
 
+   inst->SaturateMode |= mov->SaturateMode;
+
    /* Depending on the instruction, we may need to recompute the swizzles.
     * Also, some other instructions (like TEX) are not linear. We will only
     * consider completely active sources and destinations