i965: Handle scalar result swizzling in shared GLSL/non-GLSL code.

This is preparation for merging of brw_wm_glsl.c and
brw_wm_emit.c, and glsl.c doesn't swizzle channel results around.

5 files changed, 103 insertions, 69 deletions

diff --git a/src/mesa/drivers/dri/i965/brw_wm.h b/src/mesa/drivers/dri/i965/brw_wm.h
index 9eda2cb7ca8..7a46b1d5437 100644
--- a/src/mesa/drivers/dri/i965/brw_wm.h
+++ b/src/mesa/drivers/dri/i965/brw_wm.h
@@ -297,6 +297,7 @@ void brw_wm_lookup_iz( GLuint line_aa,
 
 GLboolean brw_wm_is_glsl(const struct gl_fragment_program *fp);
 void brw_wm_glsl_emit(struct brw_context *brw, struct brw_wm_compile *c);
+int brw_num_wm_src_regs(gl_inst_opcode op);
 
 
 #endif
diff --git a/src/mesa/drivers/dri/i965/brw_wm_emit.c b/src/mesa/drivers/dri/i965/brw_wm_emit.c
index b9da7e69126..b3cf524c63e 100644
--- a/src/mesa/drivers/dri/i965/brw_wm_emit.c
+++ b/src/mesa/drivers/dri/i965/brw_wm_emit.c
@@ -543,16 +543,18 @@ static void emit_dp3( struct brw_compile *p,
 		      const struct brw_reg *arg0,
 		      const struct brw_reg *arg1 )
 {
+   int dst_chan = _mesa_ffs(mask & WRITEMASK_XYZW) - 1;
+
    if (!(mask & WRITEMASK_XYZW))
       return; /* Do not emit dead code */
 
-   assert((mask & WRITEMASK_XYZW) == WRITEMASK_X);
+   assert(is_power_of_two(mask & WRITEMASK_XYZW));
 
    brw_MUL(p, brw_null_reg(), arg0[0], arg1[0]);
    brw_MAC(p, brw_null_reg(), arg0[1], arg1[1]);
 
    brw_set_saturate(p, (mask & SATURATE) ? 1 : 0);
-   brw_MAC(p, dst[0], arg0[2], arg1[2]);
+   brw_MAC(p, dst[dst_chan], arg0[2], arg1[2]);
    brw_set_saturate(p, 0);
 }
 
@@ -563,17 +565,19 @@ static void emit_dp4( struct brw_compile *p,
 		      const struct brw_reg *arg0,
 		      const struct brw_reg *arg1 )
 {
+   int dst_chan = _mesa_ffs(mask & WRITEMASK_XYZW) - 1;
+
    if (!(mask & WRITEMASK_XYZW))
       return; /* Do not emit dead code */
 
-   assert((mask & WRITEMASK_XYZW) == WRITEMASK_X);
+   assert(is_power_of_two(mask & WRITEMASK_XYZW));
 
    brw_MUL(p, brw_null_reg(), arg0[0], arg1[0]);
    brw_MAC(p, brw_null_reg(), arg0[1], arg1[1]);
    brw_MAC(p, brw_null_reg(), arg0[2], arg1[2]);
 
    brw_set_saturate(p, (mask & SATURATE) ? 1 : 0);
-   brw_MAC(p, dst[0], arg0[3], arg1[3]);
+   brw_MAC(p, dst[dst_chan], arg0[3], arg1[3]);
    brw_set_saturate(p, 0);
 }
 
@@ -630,18 +634,19 @@ static void emit_math1( struct brw_compile *p,
 			GLuint mask,
 			const struct brw_reg *arg0 )
 {
+   int dst_chan = _mesa_ffs(mask & WRITEMASK_XYZW) - 1;
+
    if (!(mask & WRITEMASK_XYZW))
       return; /* Do not emit dead code */
 
-   //assert((mask & WRITEMASK_XYZW) == WRITEMASK_X ||
-   //	  function == BRW_MATH_FUNCTION_SINCOS);
-   
+   assert(is_power_of_two(mask & WRITEMASK_XYZW));
+
    brw_MOV(p, brw_message_reg(2), arg0[0]);
 
    /* Send two messages to perform all 16 operations:
     */
    brw_math_16(p, 
-	       dst[0],
+	       dst[dst_chan],
 	       function,
 	       (mask & SATURATE) ? BRW_MATH_SATURATE_SATURATE : BRW_MATH_SATURATE_NONE,
 	       2,
@@ -657,10 +662,12 @@ static void emit_math2( struct brw_compile *p,
 			const struct brw_reg *arg0,
 			const struct brw_reg *arg1)
 {
+   int dst_chan = _mesa_ffs(mask & WRITEMASK_XYZW) - 1;
+
    if (!(mask & WRITEMASK_XYZW))
       return; /* Do not emit dead code */
 
-   assert((mask & WRITEMASK_XYZW) == WRITEMASK_X);
+   assert(is_power_of_two(mask & WRITEMASK_XYZW));
 
    brw_push_insn_state(p);
 
@@ -679,7 +686,7 @@ static void emit_math2( struct brw_compile *p,
     */
    brw_set_compression_control(p, BRW_COMPRESSION_NONE);
    brw_math(p, 
-	    dst[0],
+	    dst[dst_chan],
 	    function,
 	    (mask & SATURATE) ? BRW_MATH_SATURATE_SATURATE : BRW_MATH_SATURATE_NONE,
 	    2,
@@ -689,7 +696,7 @@ static void emit_math2( struct brw_compile *p,
 
    brw_set_compression_control(p, BRW_COMPRESSION_2NDHALF);
    brw_math(p, 
-	    offset(dst[0],1),
+	    offset(dst[dst_chan],1),
 	    function,
 	    (mask & SATURATE) ? BRW_MATH_SATURATE_SATURATE : BRW_MATH_SATURATE_NONE,
 	    4,
diff --git a/src/mesa/drivers/dri/i965/brw_wm_fp.c b/src/mesa/drivers/dri/i965/brw_wm_fp.c
index a8312228496..8e37a01ff16 100644
--- a/src/mesa/drivers/dri/i965/brw_wm_fp.c
+++ b/src/mesa/drivers/dri/i965/brw_wm_fp.c
@@ -226,9 +226,42 @@ static struct prog_instruction * emit_op(struct brw_wm_compile *c,
                       0, 0, 0,  /* tex unit, target, shadow */
                       src0, src1, src2);
 }
-   
 
 
+/* Many Mesa opcodes produce the same value across all the result channels.
+ * We'd rather not have to support that splatting in the opcode implementations,
+ * and brw_wm_pass*.c wants to optimize them out by shuffling references around
+ * anyway.  We can easily get both by emitting the opcode to one channel, and
+ * then MOVing it to the others, which brw_wm_pass*.c already understands.
+ */
+static struct prog_instruction *emit_scalar_insn(struct brw_wm_compile *c,
+						 const struct prog_instruction *inst0)
+{
+   struct prog_instruction *inst;
+   unsigned int dst_chan;
+   unsigned int other_channel_mask;
+
+   if (inst0->DstReg.WriteMask == 0)
+      return NULL;
+
+   dst_chan = _mesa_ffs(inst0->DstReg.WriteMask) - 1;
+   inst = get_fp_inst(c);
+   *inst = *inst0;
+   inst->DstReg.WriteMask = 1 << dst_chan;
+
+   other_channel_mask = inst0->DstReg.WriteMask & ~(1 << dst_chan);
+   if (other_channel_mask != 0) {
+      inst = emit_op(c,
+		     OPCODE_MOV,
+		     dst_mask(inst0->DstReg, other_channel_mask),
+		     0,
+		     src_swizzle1(src_reg_from_dst(inst0->DstReg), dst_chan),
+		     src_undef(),
+		     src_undef());
+   }
+   return inst;
+}
+
 
 /***********************************************************************
  * Special instructions for interpolation and other tasks
@@ -1138,9 +1171,11 @@ void brw_wm_pass_fp( struct brw_wm_compile *c )
 	 break;
       case OPCODE_PRINT:
 	 break;
-	 
       default:
-	 emit_insn(c, inst);
+	 if (brw_wm_is_scalar_result(inst->Opcode))
+	    emit_scalar_insn(c, inst);
+	 else
+	    emit_insn(c, inst);
 	 break;
       }
    }
diff --git a/src/mesa/drivers/dri/i965/brw_wm_glsl.c b/src/mesa/drivers/dri/i965/brw_wm_glsl.c
index 7ff6125dcac..5f49279f0e6 100644
--- a/src/mesa/drivers/dri/i965/brw_wm_glsl.c
+++ b/src/mesa/drivers/dri/i965/brw_wm_glsl.c
@@ -1032,12 +1032,20 @@ static void emit_dp3(struct brw_wm_compile *c,
     struct brw_reg src0[3], src1[3], dst;
     int i;
     struct brw_compile *p = &c->func;
+    GLuint mask = inst->DstReg.WriteMask;
+    int dst_chan = _mesa_ffs(mask & WRITEMASK_XYZW) - 1;
+
+    if (!(mask & WRITEMASK_XYZW))
+	return;
+
+    assert(is_power_of_two(mask & WRITEMASK_XYZW));
+
     for (i = 0; i < 3; i++) {
 	src0[i] = get_src_reg(c, inst, 0, i);
 	src1[i] = get_src_reg_imm(c, inst, 1, i);
     }
 
-    dst = get_dst_reg(c, inst, get_scalar_dst_index(inst));
+    dst = get_dst_reg(c, inst, dst_chan);
     brw_MUL(p, brw_null_reg(), src0[0], src1[0]);
     brw_MAC(p, brw_null_reg(), src0[1], src1[1]);
     brw_set_saturate(p, (inst->SaturateMode != SATURATE_OFF) ? 1 : 0);
@@ -1051,11 +1059,19 @@ static void emit_dp4(struct brw_wm_compile *c,
     struct brw_reg src0[4], src1[4], dst;
     int i;
     struct brw_compile *p = &c->func;
+    GLuint mask = inst->DstReg.WriteMask;
+    int dst_chan = _mesa_ffs(mask & WRITEMASK_XYZW) - 1;
+
+    if (!(mask & WRITEMASK_XYZW))
+	return;
+
+    assert(is_power_of_two(mask & WRITEMASK_XYZW));
+
     for (i = 0; i < 4; i++) {
 	src0[i] = get_src_reg(c, inst, 0, i);
 	src1[i] = get_src_reg_imm(c, inst, 1, i);
     }
-    dst = get_dst_reg(c, inst, get_scalar_dst_index(inst));
+    dst = get_dst_reg(c, inst, dst_chan);
     brw_MUL(p, brw_null_reg(), src0[0], src1[0]);
     brw_MAC(p, brw_null_reg(), src0[1], src1[1]);
     brw_MAC(p, brw_null_reg(), src0[2], src1[2]);
@@ -1070,11 +1086,19 @@ static void emit_dph(struct brw_wm_compile *c,
     struct brw_reg src0[4], src1[4], dst;
     int i;
     struct brw_compile *p = &c->func;
+    GLuint mask = inst->DstReg.WriteMask;
+    int dst_chan = _mesa_ffs(mask & WRITEMASK_XYZW) - 1;
+
+    if (!(mask & WRITEMASK_XYZW))
+	return;
+
+    assert(is_power_of_two(mask & WRITEMASK_XYZW));
+
     for (i = 0; i < 4; i++) {
 	src0[i] = get_src_reg(c, inst, 0, i);
 	src1[i] = get_src_reg_imm(c, inst, 1, i);
     }
-    dst = get_dst_reg(c, inst, get_scalar_dst_index(inst));
+    dst = get_dst_reg(c, inst, dst_chan);
     brw_MUL(p, brw_null_reg(), src0[0], src1[0]);
     brw_MAC(p, brw_null_reg(), src0[1], src1[1]);
     brw_MAC(p, dst, src0[2], src1[2]);
@@ -1092,37 +1116,28 @@ static void emit_math1(struct brw_wm_compile *c,
                        const struct prog_instruction *inst, GLuint func)
 {
     struct brw_compile *p = &c->func;
-    struct brw_reg src0, dst, tmp;
-    const int mark = mark_tmps( c );
-    int i;
+    struct brw_reg src0, dst;
+    GLuint mask = inst->DstReg.WriteMask;
+    int dst_chan = _mesa_ffs(mask & WRITEMASK_XYZW) - 1;
+
+    if (!(mask & WRITEMASK_XYZW))
+	return;
 
-    tmp = alloc_tmp(c);
+    assert(is_power_of_two(mask & WRITEMASK_XYZW));
 
     /* Get first component of source register */
+    dst = get_dst_reg(c, inst, dst_chan);
     src0 = get_src_reg(c, inst, 0, 0);
 
-    /* tmp = func(src0) */
     brw_MOV(p, brw_message_reg(2), src0);
     brw_math(p,
-             tmp,
+             dst,
              func,
              (inst->SaturateMode != SATURATE_OFF) ? BRW_MATH_SATURATE_SATURATE : BRW_MATH_SATURATE_NONE,
              2,
              brw_null_reg(),
              BRW_MATH_DATA_VECTOR,
              BRW_MATH_PRECISION_FULL);
-
-    /*tmp.dw1.bits.swizzle = SWIZZLE_XXXX;*/
-
-    /* replicate tmp value across enabled dest channels */
-    for (i = 0; i < 4; i++) {
-       if (inst->DstReg.WriteMask & (1 << i)) {
-          dst = get_dst_reg(c, inst, i);
-          brw_MOV(p, dst, tmp);
-       }
-    }
-
-    release_tmps(c, mark);
 }
 
 static void emit_rcp(struct brw_wm_compile *c,
@@ -1322,7 +1337,15 @@ static void emit_pow(struct brw_wm_compile *c,
 {
     struct brw_compile *p = &c->func;
     struct brw_reg dst, src0, src1;
-    dst = get_dst_reg(c, inst, get_scalar_dst_index(inst));
+    GLuint mask = inst->DstReg.WriteMask;
+    int dst_chan = _mesa_ffs(mask & WRITEMASK_XYZW) - 1;
+
+    if (!(mask & WRITEMASK_XYZW))
+	return;
+
+    assert(is_power_of_two(mask & WRITEMASK_XYZW));
+
+    dst = get_dst_reg(c, inst, dst_chan);
     src0 = get_src_reg_imm(c, inst, 0, 0);
     src1 = get_src_reg_imm(c, inst, 1, 0);
 
@@ -3042,7 +3065,6 @@ static void brw_wm_emit_glsl(struct brw_context *brw, struct brw_wm_compile *c)
     }
 }
 
-
 /**
  * Do GPU code generation for shaders that use GLSL features such as
  * flow control.  Other shaders will be compiled with the 
diff --git a/src/mesa/drivers/dri/i965/brw_wm_pass0.c b/src/mesa/drivers/dri/i965/brw_wm_pass0.c
index 92142764f5d..62792583396 100644
--- a/src/mesa/drivers/dri/i965/brw_wm_pass0.c
+++ b/src/mesa/drivers/dri/i965/brw_wm_pass0.c
@@ -257,34 +257,6 @@ static void pass0_set_dst( struct brw_wm_compile *c,
 }
 
 
-static void pass0_set_dst_scalar( struct brw_wm_compile *c,
-				  struct brw_wm_instruction *out,
-                                  const struct prog_instruction *inst,
-				  GLuint writemask )
-{
-   if (writemask) {
-      const struct prog_dst_register *dst = &inst->DstReg;
-      GLuint i;
-
-      /* Compute only the first (X) value:
-       */
-      out->writemask = WRITEMASK_X;
-      out->dst[0] = get_value(c);
-
-      /* Update our tracking register file for all the components in
-       * writemask:
-       */
-      for (i = 0; i < 4; i++) {
-	 if (writemask & (1<<i)) {
-	    pass0_set_fpreg_value(c, dst->File, dst->Index, i, out->dst[0]);
-	 }
-      }
-   }
-   else
-      out->writemask = 0;
-}
-
-
 static const struct brw_wm_ref *get_fp_src_reg_ref( struct brw_wm_compile *c,
 						    struct prog_src_register src,
 						    GLuint i )
@@ -363,10 +335,7 @@ translate_insn(struct brw_wm_compile *c,
 
    /* Dst:
     */
-   if (brw_wm_is_scalar_result(out->opcode)) 
-      pass0_set_dst_scalar(c, out, inst, writemask);
-   else 
-      pass0_set_dst(c, out, inst, writemask);
+   pass0_set_dst(c, out, inst, writemask);
 }