i965: Use Compr4 instruction compression mode on G4X and newer.

No statistically significant performance difference at n=3 with either
openarena or my GL demo, but cutting program size seems like a good
thing to be doing for the hypothetical app that has a working set near
icache size.

3 files changed, 29 insertions, 17 deletions

diff --git a/src/mesa/drivers/dri/i965/brw_eu.h b/src/mesa/drivers/dri/i965/brw_eu.h
index 30603bdd0e6..39eb88d7c2b 100644
--- a/src/mesa/drivers/dri/i965/brw_eu.h
+++ b/src/mesa/drivers/dri/i965/brw_eu.h
@@ -170,11 +170,11 @@ static INLINE struct brw_reg brw_reg( GLuint file,
                                       GLuint writemask )
 {
    struct brw_reg reg;
-   if (type == BRW_GENERAL_REGISTER_FILE)
+   if (file == BRW_GENERAL_REGISTER_FILE)
       assert(nr < BRW_MAX_GRF);
-   else if (type == BRW_MESSAGE_REGISTER_FILE)
-      assert(nr < BRW_MAX_MRF);
-   else if (type == BRW_ARCHITECTURE_REGISTER_FILE)
+   else if (file == BRW_MESSAGE_REGISTER_FILE)
+      assert((nr & ~(1 << 7)) < BRW_MAX_MRF);
+   else if (file == BRW_ARCHITECTURE_REGISTER_FILE)
       assert(nr <= BRW_ARF_IP);
 
    reg.type = type;
@@ -538,7 +538,7 @@ static INLINE struct brw_reg brw_mask_reg( GLuint subnr )
 
 static INLINE struct brw_reg brw_message_reg( GLuint nr )
 {
-   assert(nr < BRW_MAX_MRF);
+   assert((nr & ~(1 << 7)) < BRW_MAX_MRF);
    return brw_vec8_reg(BRW_MESSAGE_REGISTER_FILE,
 		       nr,
 		       0);
diff --git a/src/mesa/drivers/dri/i965/brw_eu_emit.c b/src/mesa/drivers/dri/i965/brw_eu_emit.c
index 241cdc33f86..7ceabba288a 100644
--- a/src/mesa/drivers/dri/i965/brw_eu_emit.c
+++ b/src/mesa/drivers/dri/i965/brw_eu_emit.c
@@ -55,7 +55,8 @@ static void guess_execution_size( struct brw_instruction *insn,
 static void brw_set_dest( struct brw_instruction *insn,
 			  struct brw_reg dest )
 {
-   if (dest.type != BRW_ARCHITECTURE_REGISTER_FILE)
+   if (dest.file != BRW_ARCHITECTURE_REGISTER_FILE &&
+       dest.file != BRW_MESSAGE_REGISTER_FILE)
       assert(dest.nr < 128);
 
    insn->bits1.da1.dest_reg_file = dest.file;
diff --git a/src/mesa/drivers/dri/i965/brw_wm_emit.c b/src/mesa/drivers/dri/i965/brw_wm_emit.c
index 9cd1fedacb7..eb37ea18640 100644
--- a/src/mesa/drivers/dri/i965/brw_wm_emit.c
+++ b/src/mesa/drivers/dri/i965/brw_wm_emit.c
@@ -1105,6 +1105,7 @@ void emit_fb_write(struct brw_wm_compile *c,
 		   GLuint eot)
 {
    struct brw_compile *p = &c->func;
+   struct brw_context *brw = p->brw;
    GLuint nr = 2;
    GLuint channel;
 
@@ -1119,18 +1120,28 @@ void emit_fb_write(struct brw_wm_compile *c,
    brw_push_insn_state(p);
 
    for (channel = 0; channel < 4; channel++) {
-      /*  mov (8) m2.0<1>:ud   r28.0<8;8,1>:ud  { Align1 } */
-      /*  mov (8) m6.0<1>:ud   r29.0<8;8,1>:ud  { Align1 SecHalf } */
-      brw_set_compression_control(p, BRW_COMPRESSION_NONE);
-      brw_MOV(p,
-	      brw_message_reg(nr + channel),
-	      arg0[channel]);
-
-      if (c->dispatch_width == 16) {
-	 brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF);
+      if (c->dispatch_width == 16 && (BRW_IS_G4X(brw) || BRW_IS_IGDNG(brw))) {
+	 /* By setting the high bit of the MRF register number, we indicate
+	  * that we want COMPR4 mode - instead of doing the usual destination
+	  * + 1 for the second half we get destination + 4.
+	  */
 	 brw_MOV(p,
-		 brw_message_reg(nr + channel + 4),
-		 sechalf(arg0[channel]));
+		 brw_message_reg(nr + channel + (1 << 7)),
+		 arg0[channel]);
+      } else {
+	 /*  mov (8) m2.0<1>:ud   r28.0<8;8,1>:ud  { Align1 } */
+	 /*  mov (8) m6.0<1>:ud   r29.0<8;8,1>:ud  { Align1 SecHalf } */
+	 brw_set_compression_control(p, BRW_COMPRESSION_NONE);
+	 brw_MOV(p,
+		 brw_message_reg(nr + channel),
+		 arg0[channel]);
+
+	 if (c->dispatch_width == 16) {
+	    brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF);
+	    brw_MOV(p,
+		    brw_message_reg(nr + channel + 4),
+		    sechalf(arg0[channel]));
+	 }
       }
    }
    /* skip over the regs populated above: