Merge branch 'glsl-to-tgsi'
authorBryan Cain <bryancain3@gmail.com>
Thu, 4 Aug 2011 20:43:34 +0000 (15:43 -0500)
committerBryan Cain <bryancain3@gmail.com>
Thu, 4 Aug 2011 20:43:34 +0000 (15:43 -0500)
Conflicts:
src/mesa/state_tracker/st_atom_pixeltransfer.c
src/mesa/state_tracker/st_program.c

43 files changed:
src/gallium/auxiliary/tgsi/tgsi_exec.c
src/gallium/auxiliary/tgsi/tgsi_exec.h
src/gallium/drivers/i915/i915_screen.c
src/gallium/drivers/i965/brw_screen.c
src/gallium/drivers/nv50/nv50_screen.c
src/gallium/drivers/nvc0/nvc0_screen.c
src/gallium/drivers/nvfx/nvfx_screen.c
src/gallium/drivers/r300/r300_screen.c
src/gallium/drivers/r600/r600_pipe.c
src/gallium/drivers/svga/svga_screen.c
src/gallium/include/pipe/p_defines.h
src/mesa/drivers/dri/i965/brw_fs.cpp
src/mesa/drivers/dri/i965/brw_vs_emit.c
src/mesa/drivers/dri/r200/r200_vertprog.c
src/mesa/drivers/dri/r600/evergreen_fragprog.c
src/mesa/drivers/dri/r600/evergreen_vertprog.c
src/mesa/drivers/dri/r600/r700_fragprog.c
src/mesa/drivers/dri/r600/r700_vertprog.c
src/mesa/main/ff_fragment_shader.cpp
src/mesa/main/ffvertex_prog.c
src/mesa/main/uniforms.c
src/mesa/program/ir_to_mesa.cpp
src/mesa/program/nvfragparse.c
src/mesa/program/prog_execute.c
src/mesa/program/prog_parameter.c
src/mesa/program/prog_parameter.h
src/mesa/program/prog_parameter_layout.c
src/mesa/program/prog_print.c
src/mesa/program/program.c
src/mesa/program/sampler.cpp
src/mesa/sources.mak
src/mesa/state_tracker/st_atom_pixeltransfer.c
src/mesa/state_tracker/st_cb_bitmap.c
src/mesa/state_tracker/st_cb_drawpixels.c
src/mesa/state_tracker/st_cb_program.c
src/mesa/state_tracker/st_glsl_to_tgsi.cpp [new file with mode: 0644]
src/mesa/state_tracker/st_glsl_to_tgsi.h [new file with mode: 0644]
src/mesa/state_tracker/st_mesa_to_tgsi.c
src/mesa/state_tracker/st_mesa_to_tgsi.h
src/mesa/state_tracker/st_program.c
src/mesa/state_tracker/st_program.h
src/mesa/state_tracker/st_texture.c
src/mesa/state_tracker/st_texture.h

index 712e8aca794c4252804f8bb76f799bde68b3e60f..587483c049b893f4f49c064287131c8daf4d0e2a 100644 (file)
@@ -2988,6 +2988,17 @@ micro_xor(union tgsi_exec_channel *dst,
    dst->u[3] = src0->u[3] ^ src1->u[3];
 }
 
+static void
+micro_mod(union tgsi_exec_channel *dst,
+          const union tgsi_exec_channel *src0,
+          const union tgsi_exec_channel *src1)
+{
+   dst->i[0] = src0->i[0] % src1->i[0];
+   dst->i[1] = src0->i[1] % src1->i[1];
+   dst->i[2] = src0->i[2] % src1->i[2];
+   dst->i[3] = src0->i[3] % src1->i[3];
+}
+
 static void
 micro_f2i(union tgsi_exec_channel *dst,
           const union tgsi_exec_channel *src)
@@ -3691,7 +3702,7 @@ exec_instruction(
       break;
 
    case TGSI_OPCODE_MOD:
-      assert (0);
+      exec_vector_binary(mach, inst, micro_mod, TGSI_EXEC_DATA_INT, TGSI_EXEC_DATA_INT);
       break;
 
    case TGSI_OPCODE_XOR:
index 33f33aa82c73b5d0851df2569e9bd9dfd6df99a6..6c32ccff323a317b2189a32f45c4fadfe57c803e 100644 (file)
@@ -400,6 +400,8 @@ tgsi_exec_get_shader_param(enum pipe_shader_cap param)
       return 1;
    case PIPE_SHADER_CAP_SUBROUTINES:
       return 1;
+   case PIPE_SHADER_CAP_INTEGERS:
+      return 1;
    default:
       return 0;
    }
index e743f6031eb765609d2b82aea85461c25c58614c..c108c70298358350dd8fe388d33872812c5c597e 100644 (file)
@@ -221,6 +221,8 @@ i915_get_shader_param(struct pipe_screen *screen, unsigned shader, enum pipe_sha
          return 1;
       case PIPE_SHADER_CAP_SUBROUTINES:
          return 0;
+      case PIPE_SHADER_CAP_INTEGERS:
+         return 0;
       default:
          debug_printf("%s: Unknown cap %u.\n", __FUNCTION__, cap);
          return 0;
index 9178dfa8f69c0daca3472dcbf7dc3c580614892b..39e9e2fa6ac92caa5b8d77fdd4aad90c93e645da 100644 (file)
@@ -243,6 +243,8 @@ brw_get_shader_param(struct pipe_screen *screen, unsigned shader, enum pipe_shad
           return 1;
       case PIPE_SHADER_CAP_SUBROUTINES:
           return 1;
+      case PIPE_SHADER_CAP_INTEGERS:
+         return 0;
       default:
          assert(0);
          return 0;
index 07a74cc2b9f4cc88e161659cf89fb2bad3b8fa76..581aad196273a6c9b113e2b93d0420870bd5d357 100644 (file)
@@ -183,6 +183,8 @@ nv50_screen_get_shader_param(struct pipe_screen *pscreen, unsigned shader,
       return 1;
    case PIPE_SHADER_CAP_SUBROUTINES:
       return 0; /* please inline, or provide function declarations */
+   case PIPE_SHADER_CAP_INTEGERS:
+      return 0;
    default:
       NOUVEAU_ERR("unknown PIPE_SHADER_CAP %d\n", param);
       return 0;
index 5d1b324dbff7ea24f4d70885822c65d072899ea5..c79256a6ba2af3ff04738bc31db8bbb3953bfe97 100644 (file)
@@ -170,6 +170,8 @@ nvc0_screen_get_shader_param(struct pipe_screen *pscreen, unsigned shader,
       return 1;
    case PIPE_SHADER_CAP_SUBROUTINES:
       return 0; /* please inline, or provide function declarations */
+   case PIPE_SHADER_CAP_INTEGERS:
+      return 0;
    default:
       NOUVEAU_ERR("unknown PIPE_SHADER_CAP %d\n", param);
       return 0;
index 8cb6cd0938e9890ca7cee0d25685ffebdd1dacaf..0e8f96772c696ad42ce1fc3fef191cb0d6d38ccb 100644 (file)
@@ -179,6 +179,8 @@ nvfx_screen_get_shader_param(struct pipe_screen *pscreen, unsigned shader, enum
                        return 1;
                case PIPE_SHADER_CAP_SUBROUTINES:
                        return 1;
+               case PIPE_SHADER_CAP_INTEGERS:
+                       return 0;
                default:
                        break;
                }
index d9378308ad025204ceb0bf734cb5a627b7ad2871..47de4005c3795bd4a2f160580b2383fdce6bd045 100644 (file)
@@ -258,6 +258,8 @@ static int r300_get_shader_param(struct pipe_screen *pscreen, unsigned shader, e
             return 1;
         case PIPE_SHADER_CAP_SUBROUTINES:
             return 0;
+        case PIPE_SHADER_CAP_INTEGERS:
+            return 0;
         default:
             break;
         }
index 5159ba350e42ddb8d6d51ad6829fe07c54232e6d..461f59439e88dd488f95d0d5bf01f526f429b067 100644 (file)
@@ -468,6 +468,8 @@ static int r600_get_shader_param(struct pipe_screen* pscreen, unsigned shader, e
                return 1;
        case PIPE_SHADER_CAP_SUBROUTINES:
                return 0;
+       case PIPE_SHADER_CAP_INTEGERS:
+               return 0;
        default:
                return 0;
        }
index b847cf331b3a7faf84264c969e0f6ca94587b870..4be10ef5821821bd57543d2e973957618480c527 100644 (file)
@@ -286,6 +286,8 @@ static int svga_get_shader_param(struct pipe_screen *screen, unsigned shader, en
          return 1;
       case PIPE_SHADER_CAP_SUBROUTINES:
          return 0;
+      case PIPE_SHADER_CAP_INTEGERS:
+         return 0;
       default:
          break;
       }
index 7ffdf97fdfbf1e848ca03f9a2fd51d0d3758515c..1ef3ae71e762f8b1801a4dcd07973ed239ce2cc1 100644 (file)
@@ -495,6 +495,7 @@ enum pipe_shader_cap
    PIPE_SHADER_CAP_INDIRECT_TEMP_ADDR = 14,
    PIPE_SHADER_CAP_INDIRECT_CONST_ADDR = 15,
    PIPE_SHADER_CAP_SUBROUTINES = 16, /* BGNSUB, ENDSUB, CAL, RET */
+   PIPE_SHADER_CAP_INTEGERS = 17,
 };
 
 
index e07798cebc1b3510ea5cb77a04b535ca0d98f6e8..02041b3bc0394886ac46557e83e1b078816f4de0 100644 (file)
@@ -617,7 +617,7 @@ fs_visitor::setup_paramvalues_refs()
    /* Set up the pointers to ParamValues now that that array is finalized. */
    for (unsigned int i = 0; i < c->prog_data.nr_params; i++) {
       c->prog_data.param[i] =
-        fp->Base.Parameters->ParameterValues[this->param_index[i]] +
+        (const float *)fp->Base.Parameters->ParameterValues[this->param_index[i]] +
         this->param_offset[i];
    }
 }
index 674a994bace565d741c25f1b86f6ce2f5d3f6226..dbabb44e45cb1098066485111bcd0b539eabb3bc 100644 (file)
@@ -1359,7 +1359,7 @@ get_src_reg( struct brw_vs_compile *c,
 
         if (component >= 0) {
            params = c->vp->program.Base.Parameters;
-           f = params->ParameterValues[src->Index][component];
+           f = params->ParameterValues[src->Index][component].f;
 
            if (src->Abs)
               f = fabs(f);
index 63e03b0e0c725c457718f8813516afe630cb4b63..cf44d7f459c6dbab5a764569b599a59aa26577b6 100644 (file)
@@ -126,10 +126,10 @@ static GLboolean r200VertexProgUpdateParams(struct gl_context *ctx, struct r200_
       case PROGRAM_NAMED_PARAM:
       //fprintf(stderr, "%s", vp->Parameters->Parameters[pi].Name);
       case PROGRAM_CONSTANT:
-        *fcmd++ = paramList->ParameterValues[pi][0];
-        *fcmd++ = paramList->ParameterValues[pi][1];
-        *fcmd++ = paramList->ParameterValues[pi][2];
-        *fcmd++ = paramList->ParameterValues[pi][3];
+        *fcmd++ = paramList->ParameterValues[pi][0].f;
+        *fcmd++ = paramList->ParameterValues[pi][1].f;
+        *fcmd++ = paramList->ParameterValues[pi][2].f;
+        *fcmd++ = paramList->ParameterValues[pi][3].f;
         break;
       default:
         _mesa_problem(NULL, "Bad param type in %s", __FUNCTION__);
index e527c379b6203448818a6e3a33608ae67abd3016..cc584ca2b35fcf839c64b985878e297d32fece0a 100644 (file)
@@ -752,10 +752,10 @@ GLboolean evergreenSetupFPconstants(struct gl_context * ctx)
            unNumParamData = paramList->NumParameters;
 
            for(ui=0; ui<unNumParamData; ui++) {
-                       evergreen->ps.consts[ui][0].f32All = paramList->ParameterValues[ui][0];
-                       evergreen->ps.consts[ui][1].f32All = paramList->ParameterValues[ui][1];
-                       evergreen->ps.consts[ui][2].f32All = paramList->ParameterValues[ui][2];
-                       evergreen->ps.consts[ui][3].f32All = paramList->ParameterValues[ui][3];
+                       evergreen->ps.consts[ui][0].f32All = paramList->ParameterValues[ui][0].f;
+                       evergreen->ps.consts[ui][1].f32All = paramList->ParameterValues[ui][1].f;
+                       evergreen->ps.consts[ui][2].f32All = paramList->ParameterValues[ui][2].f;
+                       evergreen->ps.consts[ui][3].f32All = paramList->ParameterValues[ui][3].f;
            }
 
            /* alloc multiple of 16 constants */
index 018869b9996a9f0fbc524d40be263d49a75260a8..117916ac78f512e6998353dcc642b1c33f161922 100644 (file)
@@ -684,17 +684,17 @@ GLboolean evergreenSetupVPconstants(struct gl_context * ctx)
            for(ui=0; ui<unNumParamData; ui++) {
             if(paramList->Parameters[ui].Type == PROGRAM_UNIFORM) 
             {
-                evergreen->vs.consts[ui][0].f32All = paramListOrginal->ParameterValues[ui][0];
-                       evergreen->vs.consts[ui][1].f32All = paramListOrginal->ParameterValues[ui][1];
-                       evergreen->vs.consts[ui][2].f32All = paramListOrginal->ParameterValues[ui][2];
-                       evergreen->vs.consts[ui][3].f32All = paramListOrginal->ParameterValues[ui][3];
+                evergreen->vs.consts[ui][0].f32All = paramListOrginal->ParameterValues[ui][0].f;
+                       evergreen->vs.consts[ui][1].f32All = paramListOrginal->ParameterValues[ui][1].f;
+                       evergreen->vs.consts[ui][2].f32All = paramListOrginal->ParameterValues[ui][2].f;
+                       evergreen->vs.consts[ui][3].f32All = paramListOrginal->ParameterValues[ui][3].f;
             }
             else
             {
-                       evergreen->vs.consts[ui][0].f32All = paramList->ParameterValues[ui][0];
-                       evergreen->vs.consts[ui][1].f32All = paramList->ParameterValues[ui][1];
-                       evergreen->vs.consts[ui][2].f32All = paramList->ParameterValues[ui][2];
-                       evergreen->vs.consts[ui][3].f32All = paramList->ParameterValues[ui][3];
+                       evergreen->vs.consts[ui][0].f32All = paramList->ParameterValues[ui][0].f;
+                       evergreen->vs.consts[ui][1].f32All = paramList->ParameterValues[ui][1].f;
+                       evergreen->vs.consts[ui][2].f32All = paramList->ParameterValues[ui][2].f;
+                       evergreen->vs.consts[ui][3].f32All = paramList->ParameterValues[ui][3].f;
             }
            }
 
index 40494cd6af087f9297de0c532b7779147e9c0ad0..6f9834e68fec3b3201e3b535ea3a1bfcfc5a96a8 100644 (file)
@@ -778,10 +778,10 @@ GLboolean r700SetupFragmentProgram(struct gl_context * ctx)
            unNumParamData = paramList->NumParameters;
 
            for(ui=0; ui<unNumParamData; ui++) {
-                       r700->ps.consts[ui][0].f32All = paramList->ParameterValues[ui][0];
-                       r700->ps.consts[ui][1].f32All = paramList->ParameterValues[ui][1];
-                       r700->ps.consts[ui][2].f32All = paramList->ParameterValues[ui][2];
-                       r700->ps.consts[ui][3].f32All = paramList->ParameterValues[ui][3];
+                       r700->ps.consts[ui][0].f32All = paramList->ParameterValues[ui][0].f;
+                       r700->ps.consts[ui][1].f32All = paramList->ParameterValues[ui][1].f;
+                       r700->ps.consts[ui][2].f32All = paramList->ParameterValues[ui][2].f;
+                       r700->ps.consts[ui][3].f32All = paramList->ParameterValues[ui][3].f;
            }
 
         /* Load fp constants to gpu */
index 7d4be9180a090fc728ed455450571e81765e497f..b1e2742b27d70d98557edd90fb13f31b79d4a1fc 100644 (file)
@@ -720,17 +720,17 @@ GLboolean r700SetupVertexProgram(struct gl_context * ctx)
            for(ui=0; ui<unNumParamData; ui++) {
             if(paramList->Parameters[ui].Type == PROGRAM_UNIFORM) 
             {
-                r700->vs.consts[ui][0].f32All = paramListOrginal->ParameterValues[ui][0];
-                       r700->vs.consts[ui][1].f32All = paramListOrginal->ParameterValues[ui][1];
-                       r700->vs.consts[ui][2].f32All = paramListOrginal->ParameterValues[ui][2];
-                       r700->vs.consts[ui][3].f32All = paramListOrginal->ParameterValues[ui][3];
+              r700->vs.consts[ui][0].f32All = paramListOrginal->ParameterValues[ui][0].f;
+                       r700->vs.consts[ui][1].f32All = paramListOrginal->ParameterValues[ui][1].f;
+                       r700->vs.consts[ui][2].f32All = paramListOrginal->ParameterValues[ui][2].f;
+                       r700->vs.consts[ui][3].f32All = paramListOrginal->ParameterValues[ui][3].f;
             }
             else
             {
-                       r700->vs.consts[ui][0].f32All = paramList->ParameterValues[ui][0];
-                       r700->vs.consts[ui][1].f32All = paramList->ParameterValues[ui][1];
-                       r700->vs.consts[ui][2].f32All = paramList->ParameterValues[ui][2];
-                       r700->vs.consts[ui][3].f32All = paramList->ParameterValues[ui][3];
+                       r700->vs.consts[ui][0].f32All = paramList->ParameterValues[ui][0].f;
+                       r700->vs.consts[ui][1].f32All = paramList->ParameterValues[ui][1].f;
+                       r700->vs.consts[ui][2].f32All = paramList->ParameterValues[ui][2].f;
+                       r700->vs.consts[ui][3].f32All = paramList->ParameterValues[ui][3].f;
             }
            }
 
index dbfa6b57d4d62c6f2d46ab56459aedefee5ee7f1..7cc1721688441f4e226443966c08142c4241bcd0 100644 (file)
@@ -874,7 +874,8 @@ static struct ureg register_const4f( struct texenv_fragment_program *p,
    values[1] = s1;
    values[2] = s2;
    values[3] = s3;
-   idx = _mesa_add_unnamed_constant( p->program->Base.Parameters, values, 4,
+   idx = _mesa_add_unnamed_constant( p->program->Base.Parameters,
+                                     (gl_constant_value *) values, 4,
                                      &swizzle );
    r = make_ureg(PROGRAM_CONSTANT, idx);
    r.swz = swizzle;
index b8e49a3757f2e0815179d371bbe08fadb8b0efb9..2d2485c9e06bf7c764b18e2856a19db9404c49b1 100644 (file)
@@ -455,13 +455,13 @@ static struct ureg register_const4f( struct tnl_program *p,
                              GLfloat s2,
                              GLfloat s3)
 {
-   GLfloat values[4];
+   gl_constant_value values[4];
    GLint idx;
    GLuint swizzle;
-   values[0] = s0;
-   values[1] = s1;
-   values[2] = s2;
-   values[3] = s3;
+   values[0].f = s0;
+   values[1].f = s1;
+   values[2].f = s2;
+   values[3].f = s3;
    idx = _mesa_add_unnamed_constant( p->program->Base.Parameters, values, 4,
                                      &swizzle );
    ASSERT(swizzle == SWIZZLE_NOOP);
index dd069a3a4d1464415b7ad7977fe71f09e80499f5..1329af4cd7e378505ad0dd48acad28fec7e3364c 100644 (file)
@@ -429,7 +429,7 @@ get_uniform(struct gl_context *ctx, GLuint program, GLint location,
             for (i = 0; i < rows; i++) {
                const int base = paramPos + offset + i;
                for (j = 0; j < cols; j++ ) {
-                  params[k++] = prog->Parameters->ParameterValues[base][j];
+                  params[k++] = prog->Parameters->ParameterValues[base][j].f;
                }
             }
          }
@@ -442,7 +442,7 @@ get_uniform(struct gl_context *ctx, GLuint program, GLint location,
                const int base = paramPos + offset + i;
                for (j = 0; j < cols; j++ ) {
                   params[k++] = (GLdouble)
-                     prog->Parameters->ParameterValues[base][j];
+                     prog->Parameters->ParameterValues[base][j].f;
                }
             }
          }
@@ -454,8 +454,9 @@ get_uniform(struct gl_context *ctx, GLuint program, GLint location,
             for (i = 0; i < rows; i++) {
                const int base = paramPos + offset + i;
                for (j = 0; j < cols; j++ ) {
-                  params[k++] = (GLint)
-                     prog->Parameters->ParameterValues[base][j];
+                  params[k++] = ctx->Const.GLSLVersion <= 120 ? 
+                     (GLint) prog->Parameters->ParameterValues[base][j].f : 
+                     prog->Parameters->ParameterValues[base][j].i;
                }
             }
          }
@@ -467,8 +468,9 @@ get_uniform(struct gl_context *ctx, GLuint program, GLint location,
             for (i = 0; i < rows; i++) {
                const int base = paramPos + offset + i;
                for (j = 0; j < cols; j++ ) {
-                  params[k++] = (GLuint)
-                     prog->Parameters->ParameterValues[base][j];
+                  params[k++] = ctx->Const.GLSLVersion <= 120 ? 
+                     (GLuint) prog->Parameters->ParameterValues[base][j].f : 
+                     prog->Parameters->ParameterValues[base][j].u;
                }
             }
          }
@@ -670,7 +672,7 @@ set_program_uniform(struct gl_context *ctx, struct gl_program *program,
       /* loop over number of samplers to change */
       for (i = 0; i < count; i++) {
          GLuint sampler = (GLuint)
-            program->Parameters->ParameterValues[index + offset + i][0];
+            program->Parameters->ParameterValues[index+offset + i][0].f;
          GLuint texUnit = ((GLuint *) values)[i];
 
          /* check that the sampler (tex unit index) is legal */
@@ -735,42 +737,52 @@ set_program_uniform(struct gl_context *ctx, struct gl_program *program,
 
       /* loop over number of array elements */
       for (k = 0; k < count; k++) {
-         GLfloat *uniformVal;
+         gl_constant_value *uniformVal;
 
          if (offset + k >= slots) {
             /* Extra array data is ignored */
             break;
          }
 
-         /* uniformVal (the destination) is always float[4] */
+         /* uniformVal (the destination) is always gl_constant_value[4] */
          uniformVal = program->Parameters->ParameterValues[index + offset + k];
 
          if (basicType == GL_INT) {
-            /* convert user's ints to floats */
             const GLint *iValues = ((const GLint *) values) + k * elems;
             for (i = 0; i < elems; i++) {
-               uniformVal[i] = (GLfloat) iValues[i];
+               if (ctx->Const.GLSLVersion <= 120)
+                  uniformVal[i].f = (GLfloat) iValues[i];
+               else
+                  uniformVal[i].i = iValues[i];
             }
          }
          else if (basicType == GL_UNSIGNED_INT) {
-            /* convert user's uints to floats */
             const GLuint *iValues = ((const GLuint *) values) + k * elems;
             for (i = 0; i < elems; i++) {
-               uniformVal[i] = (GLfloat) iValues[i];
+               if (ctx->Const.GLSLVersion <= 120)
+                  uniformVal[i].f = (GLfloat)(GLuint) iValues[i];
+               else
+                  uniformVal[i].u = iValues[i];
             }
          }
          else {
             const GLfloat *fValues = ((const GLfloat *) values) + k * elems;
             assert(basicType == GL_FLOAT);
             for (i = 0; i < elems; i++) {
-               uniformVal[i] = fValues[i];
+               uniformVal[i].f = fValues[i];
             }
          }
 
-         /* if the uniform is bool-valued, convert to 1.0 or 0.0 */
+         /* if the uniform is bool-valued, convert to 1 or 0 */
          if (isUniformBool) {
             for (i = 0; i < elems; i++) {
-               uniformVal[i] = uniformVal[i] ? 1.0f : 0.0f;
+               if (basicType == GL_FLOAT)
+                  uniformVal[i].b = uniformVal[i].f != 0.0f ? 1 : 0;
+               else
+                  uniformVal[i].b = uniformVal[i].u ? 1 : 0;
+               
+               if (ctx->Const.GLSLVersion <= 120)
+                  uniformVal[i].f = uniformVal[i].b ? 1.0f : 0.0f;
             }
          }
       }
@@ -936,7 +948,7 @@ set_program_uniform_matrix(struct gl_context *ctx, struct gl_program *program,
             /* Ignore writes beyond the end of (the used part of) an array */
             return;
          }
-         v = program->Parameters->ParameterValues[index + offset];
+         v = (GLfloat *) program->Parameters->ParameterValues[index + offset];
          for (row = 0; row < rows; row++) {
             if (transpose) {
                v[row] = values[src + row * cols + col];
index 382cda0c703e1066f10a3ea9aedf9dcdabd21524..debadb9a398cea81c2009ba173a6afe7e8077a5e 100644 (file)
@@ -585,7 +585,7 @@ ir_to_mesa_visitor::src_reg_for_float(float val)
    src_reg src(PROGRAM_CONSTANT, -1, NULL);
 
    src.index = _mesa_add_unnamed_constant(this->prog->Parameters,
-                                         &val, 1, &src.swizzle);
+                                         (const gl_constant_value *)&val, 1, &src.swizzle);
 
    return src;
 }
@@ -1795,7 +1795,7 @@ ir_to_mesa_visitor::visit(ir_constant *ir)
 
         src = src_reg(PROGRAM_CONSTANT, -1, NULL);
         src.index = _mesa_add_unnamed_constant(this->prog->Parameters,
-                                               values,
+                                               (gl_constant_value *) values,
                                                ir->type->vector_elements,
                                                &src.swizzle);
         emit(ir, OPCODE_MOV, mat_column, src);
@@ -1833,7 +1833,7 @@ ir_to_mesa_visitor::visit(ir_constant *ir)
 
    this->result = src_reg(PROGRAM_CONSTANT, -1, ir->type);
    this->result.index = _mesa_add_unnamed_constant(this->prog->Parameters,
-                                                  values,
+                                                  (gl_constant_value *) values,
                                                   ir->type->vector_elements,
                                                   &this->result.swizzle);
 }
@@ -2533,7 +2533,7 @@ add_uniforms_to_parameters_list(struct gl_shader_program *shader_program,
          */
         if (file == PROGRAM_SAMPLER) {
            for (unsigned int j = 0; j < size / 4; j++)
-              prog->Parameters->ParameterValues[index + j][0] = next_sampler++;
+              prog->Parameters->ParameterValues[index + j][0].f = next_sampler++;
         }
 
         /* The location chosen in the Parameters list here (returned
index 8516b5fc1ff07212042299102b56aa5c52c7436f..ce72c610d891220cefbb95726074163094bf5faa 100644 (file)
@@ -472,8 +472,9 @@ Parse_ScalarConstant(struct parse_state *parseState, GLfloat *number)
       const GLfloat *constant;
       if (!Parse_Identifier(parseState, ident))
          RETURN_ERROR1("Expected an identifier");
-      constant = _mesa_lookup_parameter_value(parseState->parameters,
-                                              -1, (const char *) ident);
+      constant = (GLfloat *)_mesa_lookup_parameter_value(parseState->parameters,
+                                                         -1, 
+                                                         (const char *) ident);
       /* XXX Check that it's a constant and not a parameter */
       if (!constant) {
          RETURN_ERROR1("Undefined symbol");
@@ -1039,7 +1040,8 @@ Parse_VectorSrc(struct parse_state *parseState,
       if (!Parse_ScalarConstant(parseState, values))
          RETURN_ERROR;
       paramIndex = _mesa_add_unnamed_constant(parseState->parameters,
-                                              values, 4, NULL);
+                                              (gl_constant_value *) values,
+                                              4, NULL);
       srcReg->File = PROGRAM_NAMED_PARAM;
       srcReg->Index = paramIndex;
    }
@@ -1051,7 +1053,8 @@ Parse_VectorSrc(struct parse_state *parseState,
       if (!Parse_VectorConstant(parseState, values))
          RETURN_ERROR;
       paramIndex = _mesa_add_unnamed_constant(parseState->parameters,
-                                              values, 4, NULL);
+                                              (gl_constant_value *) values,
+                                              4, NULL);
       srcReg->File = PROGRAM_NAMED_PARAM;
       srcReg->Index = paramIndex;      
    }
@@ -1145,7 +1148,8 @@ Parse_ScalarSrcReg(struct parse_state *parseState,
       if (!Parse_VectorConstant(parseState, values))
          RETURN_ERROR;
       paramIndex = _mesa_add_unnamed_constant(parseState->parameters,
-                                              values, 4, NULL);
+                                              (gl_constant_value *) values,
+                                              4, NULL);
       srcReg->File = PROGRAM_NAMED_PARAM;
       srcReg->Index = paramIndex;      
    }
@@ -1170,7 +1174,8 @@ Parse_ScalarSrcReg(struct parse_state *parseState,
       if (!Parse_ScalarConstant(parseState, values))
          RETURN_ERROR;
       paramIndex = _mesa_add_unnamed_constant(parseState->parameters,
-                                              values, 4, NULL);
+                                              (gl_constant_value *) values,
+                                              4, NULL);
       srcReg->Index = paramIndex;      
       srcReg->File = PROGRAM_NAMED_PARAM;
       needSuffix = GL_FALSE;
@@ -1296,7 +1301,8 @@ Parse_InstructionSequence(struct parse_state *parseState,
             RETURN_ERROR2(id, "already defined");
          }
          _mesa_add_named_parameter(parseState->parameters,
-                                   (const char *) id, value);
+                                   (const char *) id,
+                                   (gl_constant_value *) value);
       }
       else if (Parse_String(parseState, "DECLARE")) {
          GLubyte id[100];
@@ -1315,7 +1321,8 @@ Parse_InstructionSequence(struct parse_state *parseState,
             RETURN_ERROR2(id, "already declared");
          }
          _mesa_add_named_parameter(parseState->parameters,
-                                   (const char *) id, value);
+                                   (const char *) id,
+                                   (gl_constant_value *) value);
       }
       else if (Parse_String(parseState, "END")) {
          inst->Opcode = OPCODE_END;
index e7553c69dbed71622f4e898b7b3ffa4453b0d423..dbfd1b91875d18f811e4a93d4ce9e11287bd7de8 100644 (file)
@@ -157,7 +157,7 @@ get_src_register_pointer(const struct prog_src_register *source,
    case PROGRAM_NAMED_PARAM:
       if (reg >= (GLint) prog->Parameters->NumParameters)
          return ZeroVec;
-      return prog->Parameters->ParameterValues[reg];
+      return (GLfloat *) prog->Parameters->ParameterValues[reg];
 
    case PROGRAM_SYSTEM_VALUE:
       assert(reg < Elements(machine->SystemValues));
index 3570cab118b22cb8d2192cc21f81155292e173e9..49b3ffbdd5c9970ca9dd74ec95af9b3312e8316c 100644 (file)
@@ -56,8 +56,8 @@ _mesa_new_parameter_list_sized(unsigned size)
       p->Parameters = (struct gl_program_parameter *)
         calloc(1, size * sizeof(struct gl_program_parameter));
 
-      p->ParameterValues = (GLfloat (*)[4])
-         _mesa_align_malloc(size * 4 *sizeof(GLfloat), 16);
+      p->ParameterValues = (gl_constant_value (*)[4])
+         _mesa_align_malloc(size * 4 *sizeof(gl_constant_value), 16);
 
 
       if ((p->Parameters == NULL) || (p->ParameterValues == NULL)) {
@@ -101,14 +101,15 @@ _mesa_free_parameter_list(struct gl_program_parameter_list *paramList)
  * \param name  the parameter name, will be duplicated/copied!
  * \param size  number of elements in 'values' vector (1..4, or more)
  * \param datatype  GL_FLOAT, GL_FLOAT_VECx, GL_INT, GL_INT_VECx or GL_NONE.
- * \param values  initial parameter value, up to 4 GLfloats, or NULL
+ * \param values  initial parameter value, up to 4 gl_constant_values, or NULL
  * \param state  state indexes, or NULL
  * \return  index of new parameter in the list, or -1 if error (out of mem)
  */
 GLint
 _mesa_add_parameter(struct gl_program_parameter_list *paramList,
                     gl_register_file type, const char *name,
-                    GLuint size, GLenum datatype, const GLfloat *values,
+                    GLuint size, GLenum datatype,
+                    const gl_constant_value *values,
                     const gl_state_index state[STATE_LENGTH],
                     GLbitfield flags)
 {
@@ -127,10 +128,10 @@ _mesa_add_parameter(struct gl_program_parameter_list *paramList,
                       oldNum * sizeof(struct gl_program_parameter),
                       paramList->Size * sizeof(struct gl_program_parameter));
 
-      paramList->ParameterValues = (GLfloat (*)[4])
+      paramList->ParameterValues = (gl_constant_value (*)[4])
          _mesa_align_realloc(paramList->ParameterValues,         /* old buf */
-                             oldNum * 4 * sizeof(GLfloat),      /* old size */
-                             paramList->Size * 4 *sizeof(GLfloat), /* new sz */
+                             oldNum * 4 * sizeof(gl_constant_value),/* old sz */
+                             paramList->Size*4*sizeof(gl_constant_value),/*new*/
                              16);
    }
 
@@ -142,7 +143,7 @@ _mesa_add_parameter(struct gl_program_parameter_list *paramList,
       return -1;
    }
    else {
-      GLuint i;
+      GLuint i, j;
 
       paramList->NumParameters = oldNum + sz4;
 
@@ -163,7 +164,8 @@ _mesa_add_parameter(struct gl_program_parameter_list *paramList,
          }
          else {
             /* silence valgrind */
-            ASSIGN_4V(paramList->ParameterValues[oldNum + i], 0, 0, 0, 0);
+            for (j = 0; j < 4; j++)
+               paramList->ParameterValues[oldNum + i][j].f = 0;
          }
          size -= 4;
       }
@@ -184,7 +186,7 @@ _mesa_add_parameter(struct gl_program_parameter_list *paramList,
  */
 GLint
 _mesa_add_named_parameter(struct gl_program_parameter_list *paramList,
-                          const char *name, const GLfloat values[4])
+                          const char *name, const gl_constant_value values[4])
 {
    return _mesa_add_parameter(paramList, PROGRAM_NAMED_PARAM, name,
                               4, GL_NONE, values, NULL, 0x0);
@@ -204,17 +206,17 @@ _mesa_add_named_parameter(struct gl_program_parameter_list *paramList,
  */
 GLint
 _mesa_add_named_constant(struct gl_program_parameter_list *paramList,
-                         const char *name, const GLfloat values[4],
+                         const char *name, const gl_constant_value values[4],
                          GLuint size)
 {
    /* first check if this is a duplicate constant */
    GLint pos;
    for (pos = 0; pos < (GLint)paramList->NumParameters; pos++) {
-      const GLfloat *pvals = paramList->ParameterValues[pos];
-      if (pvals[0] == values[0] &&
-          pvals[1] == values[1] &&
-          pvals[2] == values[2] &&
-          pvals[3] == values[3] &&
+      const gl_constant_value *pvals = paramList->ParameterValues[pos];
+      if (pvals[0].u == values[0].u &&
+          pvals[1].u == values[1].u &&
+          pvals[2].u == values[2].u &&
+          pvals[3].u == values[3].u &&
           strcmp(paramList->Parameters[pos].Name, name) == 0) {
          /* Same name and value is already in the param list - reuse it */
          return pos;
@@ -239,9 +241,9 @@ _mesa_add_named_constant(struct gl_program_parameter_list *paramList,
  * \return index/position of the new parameter in the parameter list.
  */
 GLint
-_mesa_add_unnamed_constant(struct gl_program_parameter_list *paramList,
-                           const GLfloat values[4], GLuint size,
-                           GLuint *swizzleOut)
+_mesa_add_typed_unnamed_constant(struct gl_program_parameter_list *paramList,
+                           const gl_constant_value values[4], GLuint size,
+                           GLenum datatype, GLuint *swizzleOut)
 {
    GLint pos;
    ASSERT(size >= 1);
@@ -262,7 +264,7 @@ _mesa_add_unnamed_constant(struct gl_program_parameter_list *paramList,
          struct gl_program_parameter *p = paramList->Parameters + pos;
          if (p->Type == PROGRAM_CONSTANT && p->Size + size <= 4) {
             /* ok, found room */
-            GLfloat *pVal = paramList->ParameterValues[pos];
+            gl_constant_value *pVal = paramList->ParameterValues[pos];
             GLuint swz = p->Size; /* 1, 2 or 3 for Y, Z, W */
             pVal[p->Size] = values[0];
             p->Size++;
@@ -274,7 +276,7 @@ _mesa_add_unnamed_constant(struct gl_program_parameter_list *paramList,
 
    /* add a new parameter to store this constant */
    pos = _mesa_add_parameter(paramList, PROGRAM_CONSTANT, NULL,
-                             size, GL_NONE, values, NULL, 0x0);
+                             size, datatype, values, NULL, 0x0);
    if (pos >= 0 && swizzleOut) {
       if (size == 1)
          *swizzleOut = SWIZZLE_XXXX;
@@ -284,6 +286,28 @@ _mesa_add_unnamed_constant(struct gl_program_parameter_list *paramList,
    return pos;
 }
 
+/**
+ * Add a new unnamed constant to the parameter list.  This will be used
+ * when a fragment/vertex program contains something like this:
+ *    MOV r, { 0, 1, 2, 3 };
+ * If swizzleOut is non-null we'll search the parameter list for an
+ * existing instance of the constant which matches with a swizzle.
+ *
+ * \param paramList  the parameter list
+ * \param values  four float values
+ * \param swizzleOut  returns swizzle mask for accessing the constant
+ * \return index/position of the new parameter in the parameter list.
+ * \sa _mesa_add_typed_unnamed_constant
+ */
+GLint
+_mesa_add_unnamed_constant(struct gl_program_parameter_list *paramList,
+                           const gl_constant_value values[4], GLuint size,
+                           GLuint *swizzleOut)
+{
+   return _mesa_add_typed_unnamed_constant(paramList, values, size, GL_NONE,
+                                           swizzleOut);
+}
+
 /**
  * Add parameter representing a varying variable.
  */
@@ -401,7 +425,7 @@ _mesa_add_state_reference(struct gl_program_parameter_list *paramList,
  * Lookup a parameter value by name in the given parameter list.
  * \return pointer to the float[4] values.
  */
-GLfloat *
+gl_constant_value *
 _mesa_lookup_parameter_value(const struct gl_program_parameter_list *paramList,
                              GLsizei nameLen, const char *name)
 {
@@ -465,7 +489,7 @@ _mesa_lookup_parameter_index(const struct gl_program_parameter_list *paramList,
  */
 GLboolean
 _mesa_lookup_parameter_constant(const struct gl_program_parameter_list *list,
-                                const GLfloat v[], GLuint vSize,
+                                const gl_constant_value v[], GLuint vSize,
                                 GLint *posOut, GLuint *swizzleOut)
 {
    GLuint i;
@@ -484,7 +508,7 @@ _mesa_lookup_parameter_constant(const struct gl_program_parameter_list *list,
             /* swizzle not allowed */
             GLuint j, match = 0;
             for (j = 0; j < vSize; j++) {
-               if (v[j] == list->ParameterValues[i][j])
+               if (v[j].u == list->ParameterValues[i][j].u)
                   match++;
             }
             if (match == vSize) {
@@ -498,7 +522,7 @@ _mesa_lookup_parameter_constant(const struct gl_program_parameter_list *list,
                 /* look for v[0] anywhere within float[4] value */
                 GLuint j;
                 for (j = 0; j < list->Parameters[i].Size; j++) {
-                   if (list->ParameterValues[i][j] == v[0]) {
+                   if (list->ParameterValues[i][j].u == v[0].u) {
                       /* found it */
                       *posOut = i;
                       *swizzleOut = MAKE_SWIZZLE4(j, j, j, j);
@@ -511,13 +535,13 @@ _mesa_lookup_parameter_constant(const struct gl_program_parameter_list *list,
                 GLuint swz[4];
                 GLuint match = 0, j, k;
                 for (j = 0; j < vSize; j++) {
-                   if (v[j] == list->ParameterValues[i][j]) {
+                   if (v[j].u == list->ParameterValues[i][j].u) {
                       swz[j] = j;
                       match++;
                    }
                    else {
                       for (k = 0; k < list->Parameters[i].Size; k++) {
-                         if (v[j] == list->ParameterValues[i][k]) {
+                         if (v[j].u == list->ParameterValues[i][k].u) {
                             swz[j] = k;
                             match++;
                             break;
index 10cbbe57a6cb86ad9bd893e3eb258446b2fa459f..f858cf0fa0dbed715598d8c456b2583ebc069b44 100644 (file)
 #define PROG_PARAM_BIT_CYL_WRAP  0x10  /**< XXX gallium debug */
 /*@}*/
 
-
+/**
+ * Actual data for constant values of parameters.
+ */
+typedef union gl_constant_value {
+       GLfloat f;
+       GLboolean b;
+       GLint i;
+       GLuint u;
+} gl_constant_value;
 
 /**
  * Program parameter.
@@ -81,7 +89,7 @@ struct gl_program_parameter_list
    GLuint Size;           /**< allocated size of Parameters, ParameterValues */
    GLuint NumParameters;  /**< number of parameters in arrays */
    struct gl_program_parameter *Parameters; /**< Array [Size] */
-   GLfloat (*ParameterValues)[4];        /**< Array [Size] of GLfloat[4] */
+   gl_constant_value (*ParameterValues)[4]; /**< Array [Size] of constant[4] */
    GLbitfield StateFlags; /**< _NEW_* flags indicating which state changes
                                might invalidate ParameterValues[] */
 };
@@ -112,22 +120,28 @@ _mesa_num_parameters(const struct gl_program_parameter_list *list)
 extern GLint
 _mesa_add_parameter(struct gl_program_parameter_list *paramList,
                     gl_register_file type, const char *name,
-                    GLuint size, GLenum datatype, const GLfloat *values,
+                    GLuint size, GLenum datatype,
+                    const gl_constant_value *values,
                     const gl_state_index state[STATE_LENGTH],
                     GLbitfield flags);
 
 extern GLint
 _mesa_add_named_parameter(struct gl_program_parameter_list *paramList,
-                          const char *name, const GLfloat values[4]);
+                          const char *name, const gl_constant_value values[4]);
 
 extern GLint
 _mesa_add_named_constant(struct gl_program_parameter_list *paramList,
-                         const char *name, const GLfloat values[4],
+                         const char *name, const gl_constant_value values[4],
                          GLuint size);
 
+extern GLint
+_mesa_add_typed_unnamed_constant(struct gl_program_parameter_list *paramList,
+                           const gl_constant_value values[4], GLuint size,
+                           GLenum datatype, GLuint *swizzleOut);
+
 extern GLint
 _mesa_add_unnamed_constant(struct gl_program_parameter_list *paramList,
-                           const GLfloat values[4], GLuint size,
+                           const gl_constant_value values[4], GLuint size,
                            GLuint *swizzleOut);
 
 extern GLint
@@ -143,7 +157,7 @@ extern GLint
 _mesa_add_state_reference(struct gl_program_parameter_list *paramList,
                           const gl_state_index stateTokens[STATE_LENGTH]);
 
-extern GLfloat *
+extern gl_constant_value *
 _mesa_lookup_parameter_value(const struct gl_program_parameter_list *paramList,
                              GLsizei nameLen, const char *name);
 
@@ -153,7 +167,7 @@ _mesa_lookup_parameter_index(const struct gl_program_parameter_list *paramList,
 
 extern GLboolean
 _mesa_lookup_parameter_constant(const struct gl_program_parameter_list *list,
-                                const GLfloat v[], GLuint vSize,
+                                const gl_constant_value v[], GLuint vSize,
                                 GLint *posOut, GLuint *swizzleOut);
 
 extern GLuint
index 90a9771080c89ed4f33ca4da985e1ecac629203c..28fca3b92d917bbb45d7c6434f547d48d9bf65b7 100644 (file)
@@ -182,7 +182,7 @@ _mesa_layout_parameters(struct asm_parser_state *state)
 
         switch (p->Type) {
         case PROGRAM_CONSTANT: {
-           const float *const v =
+           const gl_constant_value *const v =
               state->prog->Parameters->ParameterValues[idx];
 
            inst->Base.SrcReg[i].Index =
index 7c3b4909e73ec97b0636c632faaf7fef2514e2ac..70412b1fa6ae6157dc2b6c40c935f591c0ab42dd 100644 (file)
@@ -985,7 +985,7 @@ _mesa_fprint_parameter_list(FILE *f,
    fprintf(f, "dirty state flags: 0x%x\n", list->StateFlags);
    for (i = 0; i < list->NumParameters; i++){
       struct gl_program_parameter *param = list->Parameters + i;
-      const GLfloat *v = list->ParameterValues[i];
+      const GLfloat *v = (GLfloat *) list->ParameterValues[i];
       fprintf(f, "param[%d] sz=%d %s %s = {%.3g, %.3g, %.3g, %.3g}",
              i, param->Size,
              _mesa_register_file_name(list->Parameters[i].Type),
index adca094ee89daa192dbd8366e67348061b0414b1..ecff2344a4473d02b8aaf855111150daadba58fb 100644 (file)
@@ -388,8 +388,9 @@ _mesa_delete_program(struct gl_context *ctx, struct gl_program *prog)
    if (prog->String)
       free(prog->String);
 
-   _mesa_free_instructions(prog->Instructions, prog->NumInstructions);
-
+   if (prog->Instructions) {
+      _mesa_free_instructions(prog->Instructions, prog->NumInstructions);
+   }
    if (prog->Parameters) {
       _mesa_free_parameter_list(prog->Parameters);
    }
@@ -1031,7 +1032,8 @@ _mesa_postprocess_program(struct gl_context *ctx, struct gl_program *prog)
    GLuint i;
    GLuint whiteSwizzle;
    GLint whiteIndex = _mesa_add_unnamed_constant(prog->Parameters,
-                                                 white, 4, &whiteSwizzle);
+                                                 (gl_constant_value *) white,
+                                                 4, &whiteSwizzle);
 
    (void) whiteIndex;
 
index 1457d1199fac036d291be67e628c15bd88e458d2..e8d34c670a965e09c67b419d81932006f654bc44 100644 (file)
@@ -132,6 +132,6 @@ _mesa_get_sampler_uniform_value(class ir_dereference *sampler,
 
    index += getname.offset;
 
-   return prog->Parameters->ParameterValues[index][0];
+   return prog->Parameters->ParameterValues[index][0].f;
 }
 }
index 4b2ec08bbb0d87b792172ac9380d1fcb26f5c653..ed008f8813e3f3d50ca32b203917dd1b1bcfacb4 100644 (file)
@@ -336,7 +336,8 @@ MESA_GALLIUM_SOURCES = \
 
 MESA_GALLIUM_CXX_SOURCES = \
        $(MAIN_CXX_SOURCES) \
-       $(SHADER_CXX_SOURCES)
+       $(SHADER_CXX_SOURCES) \
+       state_tracker/st_glsl_to_tgsi.cpp
 
 # All the core C sources, for dependency checking
 ALL_SOURCES = \
index 1f833d2821297a5cacf57a8d5680753bb84a6f72..12b5bc5ba79923d3ae9c6870b8f7bd308c62b8c8 100644 (file)
@@ -84,26 +84,6 @@ make_state_key(struct gl_context *ctx,  struct state_key *key)
 }
 
 
-static struct pipe_resource *
-create_color_map_texture(struct gl_context *ctx)
-{
-   struct st_context *st = st_context(ctx);
-   struct pipe_context *pipe = st->pipe;
-   struct pipe_resource *pt;
-   enum pipe_format format;
-   const uint texSize = 256; /* simple, and usually perfect */
-
-   /* find an RGBA texture format */
-   format = st_choose_format(pipe->screen, GL_RGBA, GL_NONE, GL_NONE,
-                             PIPE_TEXTURE_2D, 0, PIPE_BIND_SAMPLER_VIEW);
-
-   /* create texture for color map/table */
-   pt = st_texture_create(st, PIPE_TEXTURE_2D, format, 0,
-                          texSize, texSize, 1, 1, PIPE_BIND_SAMPLER_VIEW);
-   return pt;
-}
-
-
 /**
  * Update the pixelmap texture with the contents of the R/G/B/A pixel maps.
  */
@@ -219,7 +199,7 @@ get_pixel_transfer_program(struct gl_context *ctx, const struct state_key *key)
 
       /* create the colormap/texture now if not already done */
       if (!st->pixel_xfer.pixelmap_texture) {
-         st->pixel_xfer.pixelmap_texture = create_color_map_texture(ctx);
+         st->pixel_xfer.pixelmap_texture = st_create_color_map_texture(ctx);
          st->pixel_xfer.pixelmap_sampler_view =
             st_create_texture_sampler_view(st->pipe,
                                            st->pixel_xfer.pixelmap_texture);
index 067403f396b4488e34e1305f8f07ab622537ff44..beb5e7cab31aa5d80d38148baa2a1ffcf48f2cec 100644 (file)
@@ -172,6 +172,23 @@ make_bitmap_fragment_program(struct gl_context *ctx, GLuint samplerIndex)
 }
 
 
+static struct gl_program *
+make_bitmap_fragment_program_glsl(struct st_context *st,
+                                  struct st_fragment_program *orig,
+                                  GLuint samplerIndex)
+{
+   struct gl_context *ctx = st->ctx;
+   struct st_fragment_program *fp = (struct st_fragment_program *)
+      ctx->Driver.NewProgram(ctx, GL_FRAGMENT_PROGRAM_ARB, 0);
+
+   if (!fp)
+      return NULL;
+   
+   get_bitmap_visitor(fp, orig->glsl_to_tgsi, samplerIndex);
+   return &fp->Base.Base;
+}
+
+
 static int
 find_free_bit(uint bitfield)
 {
@@ -199,6 +216,7 @@ st_make_bitmap_fragment_program(struct st_context *st,
                                 GLuint *bitmap_sampler)
 {
    struct st_fragment_program *bitmap_prog;
+   struct st_fragment_program *stfpIn = (struct st_fragment_program *) fpIn;
    struct gl_program *newProg;
    uint sampler;
 
@@ -207,13 +225,18 @@ st_make_bitmap_fragment_program(struct st_context *st,
     * with the bitmap sampler/kill instructions.
     */
    sampler = find_free_bit(fpIn->Base.SamplersUsed);
-   bitmap_prog = make_bitmap_fragment_program(st->ctx, sampler);
+   
+   if (stfpIn->glsl_to_tgsi)
+      newProg = make_bitmap_fragment_program_glsl(st, stfpIn, sampler);
+   else {
+      bitmap_prog = make_bitmap_fragment_program(st->ctx, sampler);
 
-   newProg = _mesa_combine_programs(st->ctx,
-                                    &bitmap_prog->Base.Base,
-                                    &fpIn->Base);
-   /* done with this after combining */
-   st_reference_fragprog(st, &bitmap_prog, NULL);
+      newProg = _mesa_combine_programs(st->ctx,
+                                       &bitmap_prog->Base.Base,
+                                       &fpIn->Base);
+      /* done with this after combining */
+      st_reference_fragprog(st, &bitmap_prog, NULL);
+   }
 
 #if 0
    {
index 1d908c0317a49c431d7d9fd9fdec2b996795b3ba..390c518699f62651bdfce6f279397ee9ba160b72 100644 (file)
@@ -94,6 +94,46 @@ is_passthrough_program(const struct gl_fragment_program *prog)
 }
 
 
+/**
+ * Returns a fragment program which implements the current pixel transfer ops.
+ */
+static struct gl_fragment_program *
+get_glsl_pixel_transfer_program(struct st_context *st,
+                                struct st_fragment_program *orig)
+{
+   int pixelMaps = 0, scaleAndBias = 0;
+   struct gl_context *ctx = st->ctx;
+   struct st_fragment_program *fp = (struct st_fragment_program *)
+      ctx->Driver.NewProgram(ctx, GL_FRAGMENT_PROGRAM_ARB, 0);
+
+   if (!fp)
+      return NULL;
+
+   if (ctx->Pixel.RedBias != 0.0 || ctx->Pixel.RedScale != 1.0 ||
+       ctx->Pixel.GreenBias != 0.0 || ctx->Pixel.GreenScale != 1.0 ||
+       ctx->Pixel.BlueBias != 0.0 || ctx->Pixel.BlueScale != 1.0 ||
+       ctx->Pixel.AlphaBias != 0.0 || ctx->Pixel.AlphaScale != 1.0) {
+      scaleAndBias = 1;
+   }
+
+   pixelMaps = ctx->Pixel.MapColorFlag;
+
+   if (pixelMaps) {
+      /* create the colormap/texture now if not already done */
+      if (!st->pixel_xfer.pixelmap_texture) {
+         st->pixel_xfer.pixelmap_texture = st_create_color_map_texture(ctx);
+         st->pixel_xfer.pixelmap_sampler_view =
+            st_create_texture_sampler_view(st->pipe,
+                                           st->pixel_xfer.pixelmap_texture);
+      }
+   }
+
+   get_pixel_transfer_visitor(fp, orig->glsl_to_tgsi,
+                              scaleAndBias, pixelMaps);
+
+   return &fp->Base;
+}
+
 
 /**
  * Make fragment shader for glDraw/CopyPixels.  This shader is made
@@ -107,11 +147,15 @@ st_make_drawpix_fragment_program(struct st_context *st,
                                  struct gl_fragment_program **fpOut)
 {
    struct gl_program *newProg;
+   struct st_fragment_program *stfp = (struct st_fragment_program *) fpIn;
 
    if (is_passthrough_program(fpIn)) {
       newProg = (struct gl_program *) _mesa_clone_fragment_program(st->ctx,
                                              &st->pixel_xfer.program->Base);
    }
+   else if (stfp->glsl_to_tgsi != NULL) {
+      newProg = (struct gl_program *) get_glsl_pixel_transfer_program(st, stfp);
+   }
    else {
 #if 0
       /* debug */
index 32694975d1722204c9d66bf1b5f0fd430ab84e37..2abb4d8f0829035702bf962256fe745301e37d17 100644 (file)
@@ -44,6 +44,7 @@
 #include "st_program.h"
 #include "st_mesa_to_tgsi.h"
 #include "st_cb_program.h"
+#include "st_glsl_to_tgsi.h"
 
 
 
@@ -129,6 +130,9 @@ st_delete_program(struct gl_context *ctx, struct gl_program *prog)
       {
          struct st_vertex_program *stvp = (struct st_vertex_program *) prog;
          st_release_vp_variants( st, stvp );
+         
+         if (stvp->glsl_to_tgsi)
+            free_glsl_to_tgsi_visitor(stvp->glsl_to_tgsi);
       }
       break;
    case MESA_GEOMETRY_PROGRAM:
@@ -137,6 +141,9 @@ st_delete_program(struct gl_context *ctx, struct gl_program *prog)
             (struct st_geometry_program *) prog;
 
          st_release_gp_variants(st, stgp);
+         
+         if (stgp->glsl_to_tgsi)
+            free_glsl_to_tgsi_visitor(stgp->glsl_to_tgsi);
 
          if (stgp->tgsi.tokens) {
             st_free_tokens((void *) stgp->tgsi.tokens);
@@ -151,6 +158,9 @@ st_delete_program(struct gl_context *ctx, struct gl_program *prog)
 
          st_release_fp_variants(st, stfp);
          
+         if (stfp->glsl_to_tgsi)
+            free_glsl_to_tgsi_visitor(stfp->glsl_to_tgsi);
+         
          if (stfp->tgsi.tokens) {
             st_free_tokens(stfp->tgsi.tokens);
             stfp->tgsi.tokens = NULL;
@@ -242,4 +252,8 @@ st_init_program_functions(struct dd_function_table *functions)
    functions->DeleteProgram = st_delete_program;
    functions->IsProgramNative = st_is_program_native;
    functions->ProgramStringNotify = st_program_string_notify;
+   
+   functions->NewShader = st_new_shader;
+   functions->NewShaderProgram = st_new_shader_program;
+   functions->LinkShader = st_link_shader;
 }
diff --git a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
new file mode 100644 (file)
index 0000000..9c6a7ed
--- /dev/null
@@ -0,0 +1,4962 @@
+/*
+ * Copyright (C) 2005-2007  Brian Paul   All Rights Reserved.
+ * Copyright (C) 2008  VMware, Inc.   All Rights Reserved.
+ * Copyright © 2010 Intel Corporation
+ * Copyright © 2011 Bryan Cain
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
+ */
+
+/**
+ * \file glsl_to_tgsi.cpp
+ *
+ * Translate GLSL IR to TGSI.
+ */
+
+#include <stdio.h>
+#include "main/compiler.h"
+#include "ir.h"
+#include "ir_visitor.h"
+#include "ir_print_visitor.h"
+#include "ir_expression_flattening.h"
+#include "glsl_types.h"
+#include "glsl_parser_extras.h"
+#include "../glsl/program.h"
+#include "ir_optimization.h"
+#include "ast.h"
+
+extern "C" {
+#include "main/mtypes.h"
+#include "main/shaderapi.h"
+#include "main/shaderobj.h"
+#include "main/uniforms.h"
+#include "program/hash_table.h"
+#include "program/prog_instruction.h"
+#include "program/prog_optimize.h"
+#include "program/prog_print.h"
+#include "program/program.h"
+#include "program/prog_uniform.h"
+#include "program/prog_parameter.h"
+#include "program/sampler.h"
+
+#include "pipe/p_compiler.h"
+#include "pipe/p_context.h"
+#include "pipe/p_screen.h"
+#include "pipe/p_shader_tokens.h"
+#include "pipe/p_state.h"
+#include "util/u_math.h"
+#include "tgsi/tgsi_ureg.h"
+#include "tgsi/tgsi_info.h"
+#include "st_context.h"
+#include "st_program.h"
+#include "st_glsl_to_tgsi.h"
+#include "st_mesa_to_tgsi.h"
+}
+
+#define PROGRAM_IMMEDIATE PROGRAM_FILE_MAX
+#define PROGRAM_ANY_CONST ((1 << PROGRAM_LOCAL_PARAM) |  \
+                           (1 << PROGRAM_ENV_PARAM) |    \
+                           (1 << PROGRAM_STATE_VAR) |    \
+                           (1 << PROGRAM_NAMED_PARAM) |  \
+                           (1 << PROGRAM_CONSTANT) |     \
+                           (1 << PROGRAM_UNIFORM))
+
+#define MAX_TEMPS         4096
+
+class st_src_reg;
+class st_dst_reg;
+
+static int swizzle_for_size(int size);
+
+/**
+ * This struct is a corresponding struct to TGSI ureg_src.
+ */
+class st_src_reg {
+public:
+   st_src_reg(gl_register_file file, int index, const glsl_type *type)
+   {
+      this->file = file;
+      this->index = index;
+      if (type && (type->is_scalar() || type->is_vector() || type->is_matrix()))
+         this->swizzle = swizzle_for_size(type->vector_elements);
+      else
+         this->swizzle = SWIZZLE_XYZW;
+      this->negate = 0;
+      this->type = type ? type->base_type : GLSL_TYPE_ERROR;
+      this->reladdr = NULL;
+   }
+
+   st_src_reg(gl_register_file file, int index, int type)
+   {
+      this->type = type;
+      this->file = file;
+      this->index = index;
+      this->swizzle = SWIZZLE_XYZW;
+      this->negate = 0;
+      this->reladdr = NULL;
+   }
+
+   st_src_reg()
+   {
+      this->type = GLSL_TYPE_ERROR;
+      this->file = PROGRAM_UNDEFINED;
+      this->index = 0;
+      this->swizzle = 0;
+      this->negate = 0;
+      this->reladdr = NULL;
+   }
+
+   explicit st_src_reg(st_dst_reg reg);
+
+   gl_register_file file; /**< PROGRAM_* from Mesa */
+   int index; /**< temporary index, VERT_ATTRIB_*, FRAG_ATTRIB_*, etc. */
+   GLuint swizzle; /**< SWIZZLE_XYZWONEZERO swizzles from Mesa. */
+   int negate; /**< NEGATE_XYZW mask from mesa */
+   int type; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
+   /** Register index should be offset by the integer in this reg. */
+   st_src_reg *reladdr;
+};
+
+class st_dst_reg {
+public:
+   st_dst_reg(gl_register_file file, int writemask, int type)
+   {
+      this->file = file;
+      this->index = 0;
+      this->writemask = writemask;
+      this->cond_mask = COND_TR;
+      this->reladdr = NULL;
+      this->type = type;
+   }
+
+   st_dst_reg()
+   {
+      this->type = GLSL_TYPE_ERROR;
+      this->file = PROGRAM_UNDEFINED;
+      this->index = 0;
+      this->writemask = 0;
+      this->cond_mask = COND_TR;
+      this->reladdr = NULL;
+   }
+
+   explicit st_dst_reg(st_src_reg reg);
+
+   gl_register_file file; /**< PROGRAM_* from Mesa */
+   int index; /**< temporary index, VERT_ATTRIB_*, FRAG_ATTRIB_*, etc. */
+   int writemask; /**< Bitfield of WRITEMASK_[XYZW] */
+   GLuint cond_mask:4;
+   int type; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
+   /** Register index should be offset by the integer in this reg. */
+   st_src_reg *reladdr;
+};
+
+st_src_reg::st_src_reg(st_dst_reg reg)
+{
+   this->type = reg.type;
+   this->file = reg.file;
+   this->index = reg.index;
+   this->swizzle = SWIZZLE_XYZW;
+   this->negate = 0;
+   this->reladdr = reg.reladdr;
+}
+
+st_dst_reg::st_dst_reg(st_src_reg reg)
+{
+   this->type = reg.type;
+   this->file = reg.file;
+   this->index = reg.index;
+   this->writemask = WRITEMASK_XYZW;
+   this->cond_mask = COND_TR;
+   this->reladdr = reg.reladdr;
+}
+
+class glsl_to_tgsi_instruction : public exec_node {
+public:
+   /* Callers of this ralloc-based new need not call delete. It's
+    * easier to just ralloc_free 'ctx' (or any of its ancestors). */
+   static void* operator new(size_t size, void *ctx)
+   {
+      void *node;
+
+      node = rzalloc_size(ctx, size);
+      assert(node != NULL);
+
+      return node;
+   }
+
+   unsigned op;
+   st_dst_reg dst;
+   st_src_reg src[3];
+   /** Pointer to the ir source this tree came from for debugging */
+   ir_instruction *ir;
+   GLboolean cond_update;
+   bool saturate;
+   int sampler; /**< sampler index */
+   int tex_target; /**< One of TEXTURE_*_INDEX */
+   GLboolean tex_shadow;
+   int dead_mask; /**< Used in dead code elimination */
+
+   class function_entry *function; /* Set on TGSI_OPCODE_CAL or TGSI_OPCODE_BGNSUB */
+};
+
+class variable_storage : public exec_node {
+public:
+   variable_storage(ir_variable *var, gl_register_file file, int index)
+      : file(file), index(index), var(var)
+   {
+      /* empty */
+   }
+
+   gl_register_file file;
+   int index;
+   ir_variable *var; /* variable that maps to this, if any */
+};
+
+class immediate_storage : public exec_node {
+public:
+   immediate_storage(gl_constant_value *values, int size, int type)
+   {
+      memcpy(this->values, values, size * sizeof(gl_constant_value));
+      this->size = size;
+      this->type = type;
+   }
+   
+   gl_constant_value values[4];
+   int size; /**< Number of components (1-4) */
+   int type; /**< GL_FLOAT, GL_INT, GL_BOOL, or GL_UNSIGNED_INT */
+};
+
+class function_entry : public exec_node {
+public:
+   ir_function_signature *sig;
+
+   /**
+    * identifier of this function signature used by the program.
+    *
+    * At the point that TGSI instructions for function calls are
+    * generated, we don't know the address of the first instruction of
+    * the function body.  So we make the BranchTarget that is called a
+    * small integer and rewrite them during set_branchtargets().
+    */
+   int sig_id;
+
+   /**
+    * Pointer to first instruction of the function body.
+    *
+    * Set during function body emits after main() is processed.
+    */
+   glsl_to_tgsi_instruction *bgn_inst;
+
+   /**
+    * Index of the first instruction of the function body in actual TGSI.
+    *
+    * Set after conversion from glsl_to_tgsi_instruction to TGSI.
+    */
+   int inst;
+
+   /** Storage for the return value. */
+   st_src_reg return_reg;
+};
+
+class glsl_to_tgsi_visitor : public ir_visitor {
+public:
+   glsl_to_tgsi_visitor();
+   ~glsl_to_tgsi_visitor();
+
+   function_entry *current_function;
+
+   struct gl_context *ctx;
+   struct gl_program *prog;
+   struct gl_shader_program *shader_program;
+   struct gl_shader_compiler_options *options;
+
+   int next_temp;
+
+   int num_address_regs;
+   int samplers_used;
+   bool indirect_addr_temps;
+   bool indirect_addr_consts;
+   
+   int glsl_version;
+
+   variable_storage *find_variable_storage(ir_variable *var);
+
+   int add_constant(gl_register_file file, gl_constant_value values[4],
+                    int size, int datatype, GLuint *swizzle_out);
+
+   function_entry *get_function_signature(ir_function_signature *sig);
+
+   st_src_reg get_temp(const glsl_type *type);
+   void reladdr_to_temp(ir_instruction *ir, st_src_reg *reg, int *num_reladdr);
+
+   st_src_reg st_src_reg_for_float(float val);
+   st_src_reg st_src_reg_for_int(int val);
+   st_src_reg st_src_reg_for_type(int type, int val);
+
+   /**
+    * \name Visit methods
+    *
+    * As typical for the visitor pattern, there must be one \c visit method for
+    * each concrete subclass of \c ir_instruction.  Virtual base classes within
+    * the hierarchy should not have \c visit methods.
+    */
+   /*@{*/
+   virtual void visit(ir_variable *);
+   virtual void visit(ir_loop *);
+   virtual void visit(ir_loop_jump *);
+   virtual void visit(ir_function_signature *);
+   virtual void visit(ir_function *);
+   virtual void visit(ir_expression *);
+   virtual void visit(ir_swizzle *);
+   virtual void visit(ir_dereference_variable  *);
+   virtual void visit(ir_dereference_array *);
+   virtual void visit(ir_dereference_record *);
+   virtual void visit(ir_assignment *);
+   virtual void visit(ir_constant *);
+   virtual void visit(ir_call *);
+   virtual void visit(ir_return *);
+   virtual void visit(ir_discard *);
+   virtual void visit(ir_texture *);
+   virtual void visit(ir_if *);
+   /*@}*/
+
+   st_src_reg result;
+
+   /** List of variable_storage */
+   exec_list variables;
+
+   /** List of immediate_storage */
+   exec_list immediates;
+   int num_immediates;
+
+   /** List of function_entry */
+   exec_list function_signatures;
+   int next_signature_id;
+
+   /** List of glsl_to_tgsi_instruction */
+   exec_list instructions;
+
+   glsl_to_tgsi_instruction *emit(ir_instruction *ir, unsigned op);
+
+   glsl_to_tgsi_instruction *emit(ir_instruction *ir, unsigned op,
+                               st_dst_reg dst, st_src_reg src0);
+
+   glsl_to_tgsi_instruction *emit(ir_instruction *ir, unsigned op,
+                               st_dst_reg dst, st_src_reg src0, st_src_reg src1);
+
+   glsl_to_tgsi_instruction *emit(ir_instruction *ir, unsigned op,
+                               st_dst_reg dst,
+                               st_src_reg src0, st_src_reg src1, st_src_reg src2);
+   
+   unsigned get_opcode(ir_instruction *ir, unsigned op,
+                    st_dst_reg dst,
+                    st_src_reg src0, st_src_reg src1);
+
+   /**
+    * Emit the correct dot-product instruction for the type of arguments
+    */
+   void emit_dp(ir_instruction *ir,
+                st_dst_reg dst,
+                st_src_reg src0,
+                st_src_reg src1,
+                unsigned elements);
+
+   void emit_scalar(ir_instruction *ir, unsigned op,
+                   st_dst_reg dst, st_src_reg src0);
+
+   void emit_scalar(ir_instruction *ir, unsigned op,
+                   st_dst_reg dst, st_src_reg src0, st_src_reg src1);
+
+   void emit_arl(ir_instruction *ir, st_dst_reg dst, st_src_reg src0);
+
+   void emit_scs(ir_instruction *ir, unsigned op,
+                st_dst_reg dst, const st_src_reg &src);
+
+   GLboolean try_emit_mad(ir_expression *ir,
+                         int mul_operand);
+   GLboolean try_emit_sat(ir_expression *ir);
+
+   void emit_swz(ir_expression *ir);
+
+   bool process_move_condition(ir_rvalue *ir);
+
+   void remove_output_reads(gl_register_file type);
+   void simplify_cmp(void);
+
+   void rename_temp_register(int index, int new_index);
+   int get_first_temp_read(int index);
+   int get_first_temp_write(int index);
+   int get_last_temp_read(int index);
+   int get_last_temp_write(int index);
+
+   void copy_propagate(void);
+   void eliminate_dead_code(void);
+   int eliminate_dead_code_advanced(void);
+   void merge_registers(void);
+   void renumber_registers(void);
+
+   void *mem_ctx;
+};
+
+static st_src_reg undef_src = st_src_reg(PROGRAM_UNDEFINED, 0, GLSL_TYPE_ERROR);
+
+static st_dst_reg undef_dst = st_dst_reg(PROGRAM_UNDEFINED, SWIZZLE_NOOP, GLSL_TYPE_ERROR);
+
+static st_dst_reg address_reg = st_dst_reg(PROGRAM_ADDRESS, WRITEMASK_X, GLSL_TYPE_FLOAT);
+
+static void
+fail_link(struct gl_shader_program *prog, const char *fmt, ...) PRINTFLIKE(2, 3);
+
+static void
+fail_link(struct gl_shader_program *prog, const char *fmt, ...)
+{
+   va_list args;
+   va_start(args, fmt);
+   ralloc_vasprintf_append(&prog->InfoLog, fmt, args);
+   va_end(args);
+
+   prog->LinkStatus = GL_FALSE;
+}
+
+static int
+swizzle_for_size(int size)
+{
+   int size_swizzles[4] = {
+      MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_X, SWIZZLE_X, SWIZZLE_X),
+      MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Y, SWIZZLE_Y),
+      MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_Z),
+      MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_W),
+   };
+
+   assert((size >= 1) && (size <= 4));
+   return size_swizzles[size - 1];
+}
+
+static bool
+is_tex_instruction(unsigned opcode)
+{
+   const tgsi_opcode_info* info = tgsi_get_opcode_info(opcode);
+   return info->is_tex;
+}
+
+static unsigned
+num_inst_dst_regs(unsigned opcode)
+{
+   const tgsi_opcode_info* info = tgsi_get_opcode_info(opcode);
+   return info->num_dst;
+}
+
+static unsigned
+num_inst_src_regs(unsigned opcode)
+{
+   const tgsi_opcode_info* info = tgsi_get_opcode_info(opcode);
+   return info->is_tex ? info->num_src - 1 : info->num_src;
+}
+
+glsl_to_tgsi_instruction *
+glsl_to_tgsi_visitor::emit(ir_instruction *ir, unsigned op,
+                        st_dst_reg dst,
+                        st_src_reg src0, st_src_reg src1, st_src_reg src2)
+{
+   glsl_to_tgsi_instruction *inst = new(mem_ctx) glsl_to_tgsi_instruction();
+   int num_reladdr = 0, i;
+   
+   op = get_opcode(ir, op, dst, src0, src1);
+
+   /* If we have to do relative addressing, we want to load the ARL
+    * reg directly for one of the regs, and preload the other reladdr
+    * sources into temps.
+    */
+   num_reladdr += dst.reladdr != NULL;
+   num_reladdr += src0.reladdr != NULL;
+   num_reladdr += src1.reladdr != NULL;
+   num_reladdr += src2.reladdr != NULL;
+
+   reladdr_to_temp(ir, &src2, &num_reladdr);
+   reladdr_to_temp(ir, &src1, &num_reladdr);
+   reladdr_to_temp(ir, &src0, &num_reladdr);
+
+   if (dst.reladdr) {
+      emit_arl(ir, address_reg, *dst.reladdr);
+      num_reladdr--;
+   }
+   assert(num_reladdr == 0);
+
+   inst->op = op;
+   inst->dst = dst;
+   inst->src[0] = src0;
+   inst->src[1] = src1;
+   inst->src[2] = src2;
+   inst->ir = ir;
+   inst->dead_mask = 0;
+
+   inst->function = NULL;
+   
+   if (op == TGSI_OPCODE_ARL)
+      this->num_address_regs = 1;
+   
+   /* Update indirect addressing status used by TGSI */
+   if (dst.reladdr) {
+      switch(dst.file) {
+      case PROGRAM_TEMPORARY:
+         this->indirect_addr_temps = true;
+         break;
+      case PROGRAM_LOCAL_PARAM:
+      case PROGRAM_ENV_PARAM:
+      case PROGRAM_STATE_VAR:
+      case PROGRAM_NAMED_PARAM:
+      case PROGRAM_CONSTANT:
+      case PROGRAM_UNIFORM:
+         this->indirect_addr_consts = true;
+         break;
+      case PROGRAM_IMMEDIATE:
+         assert(!"immediates should not have indirect addressing");
+         break;
+      default:
+         break;
+      }
+   }
+   else {
+      for (i=0; i<3; i++) {
+         if(inst->src[i].reladdr) {
+            switch(inst->src[i].file) {
+            case PROGRAM_TEMPORARY:
+               this->indirect_addr_temps = true;
+               break;
+            case PROGRAM_LOCAL_PARAM:
+            case PROGRAM_ENV_PARAM:
+            case PROGRAM_STATE_VAR:
+            case PROGRAM_NAMED_PARAM:
+            case PROGRAM_CONSTANT:
+            case PROGRAM_UNIFORM:
+               this->indirect_addr_consts = true;
+               break;
+            case PROGRAM_IMMEDIATE:
+               assert(!"immediates should not have indirect addressing");
+               break;
+            default:
+               break;
+            }
+         }
+      }
+   }
+
+   this->instructions.push_tail(inst);
+   
+   return inst;
+}
+
+
+glsl_to_tgsi_instruction *
+glsl_to_tgsi_visitor::emit(ir_instruction *ir, unsigned op,
+                        st_dst_reg dst, st_src_reg src0, st_src_reg src1)
+{
+   return emit(ir, op, dst, src0, src1, undef_src);
+}
+
+glsl_to_tgsi_instruction *
+glsl_to_tgsi_visitor::emit(ir_instruction *ir, unsigned op,
+                        st_dst_reg dst, st_src_reg src0)
+{
+   assert(dst.writemask != 0);
+   return emit(ir, op, dst, src0, undef_src, undef_src);
+}
+
+glsl_to_tgsi_instruction *
+glsl_to_tgsi_visitor::emit(ir_instruction *ir, unsigned op)
+{
+   return emit(ir, op, undef_dst, undef_src, undef_src, undef_src);
+}
+
+/**
+ * Determines whether to use an integer, unsigned integer, or float opcode 
+ * based on the operands and input opcode, then emits the result.
+ * 
+ * TODO: type checking for remaining TGSI opcodes
+ */
+unsigned
+glsl_to_tgsi_visitor::get_opcode(ir_instruction *ir, unsigned op,
+                        st_dst_reg dst,
+                        st_src_reg src0, st_src_reg src1)
+{
+   int type = GLSL_TYPE_FLOAT;
+   
+   if (src0.type == GLSL_TYPE_FLOAT || src1.type == GLSL_TYPE_FLOAT)
+      type = GLSL_TYPE_FLOAT;
+   else if (glsl_version >= 130)
+      type = src0.type;
+
+#define case4(c, f, i, u) \
+   case TGSI_OPCODE_##c: \
+      if (type == GLSL_TYPE_INT) op = TGSI_OPCODE_##i; \
+      else if (type == GLSL_TYPE_UINT) op = TGSI_OPCODE_##u; \
+      else op = TGSI_OPCODE_##f; \
+      break;
+#define case3(f, i, u)  case4(f, f, i, u)
+#define case2fi(f, i)   case4(f, f, i, i)
+#define case2iu(i, u)   case4(i, LAST, i, u)
+   
+   switch(op) {
+      case2fi(ADD, UADD);
+      case2fi(MUL, UMUL);
+      case2fi(MAD, UMAD);
+      case3(DIV, IDIV, UDIV);
+      case3(MAX, IMAX, UMAX);
+      case3(MIN, IMIN, UMIN);
+      case2iu(MOD, UMOD);
+      
+      case2fi(SEQ, USEQ);
+      case2fi(SNE, USNE);
+      case3(SGE, ISGE, USGE);
+      case3(SLT, ISLT, USLT);
+      
+      case2iu(SHL, SHL);
+      case2iu(ISHR, USHR);
+      case2iu(NOT, NOT);
+      case2iu(AND, AND);
+      case2iu(OR, OR);
+      case2iu(XOR, XOR);
+      
+      default: break;
+   }
+   
+   assert(op != TGSI_OPCODE_LAST);
+   return op;
+}
+
+void
+glsl_to_tgsi_visitor::emit_dp(ir_instruction *ir,
+                           st_dst_reg dst, st_src_reg src0, st_src_reg src1,
+                           unsigned elements)
+{
+   static const unsigned dot_opcodes[] = {
+      TGSI_OPCODE_DP2, TGSI_OPCODE_DP3, TGSI_OPCODE_DP4
+   };
+
+   emit(ir, dot_opcodes[elements - 2], dst, src0, src1);
+}
+
+/**
+ * Emits TGSI scalar opcodes to produce unique answers across channels.
+ *
+ * Some TGSI opcodes are scalar-only, like ARB_fp/vp.  The src X
+ * channel determines the result across all channels.  So to do a vec4
+ * of this operation, we want to emit a scalar per source channel used
+ * to produce dest channels.
+ */
+void
+glsl_to_tgsi_visitor::emit_scalar(ir_instruction *ir, unsigned op,
+                               st_dst_reg dst,
+                               st_src_reg orig_src0, st_src_reg orig_src1)
+{
+   int i, j;
+   int done_mask = ~dst.writemask;
+
+   /* TGSI RCP is a scalar operation splatting results to all channels,
+    * like ARB_fp/vp.  So emit as many RCPs as necessary to cover our
+    * dst channels.
+    */
+   for (i = 0; i < 4; i++) {
+      GLuint this_mask = (1 << i);
+      glsl_to_tgsi_instruction *inst;
+      st_src_reg src0 = orig_src0;
+      st_src_reg src1 = orig_src1;
+
+      if (done_mask & this_mask)
+         continue;
+
+      GLuint src0_swiz = GET_SWZ(src0.swizzle, i);
+      GLuint src1_swiz = GET_SWZ(src1.swizzle, i);
+      for (j = i + 1; j < 4; j++) {
+         /* If there is another enabled component in the destination that is
+          * derived from the same inputs, generate its value on this pass as
+          * well.
+          */
+         if (!(done_mask & (1 << j)) &&
+             GET_SWZ(src0.swizzle, j) == src0_swiz &&
+             GET_SWZ(src1.swizzle, j) == src1_swiz) {
+            this_mask |= (1 << j);
+         }
+      }
+      src0.swizzle = MAKE_SWIZZLE4(src0_swiz, src0_swiz,
+                                  src0_swiz, src0_swiz);
+      src1.swizzle = MAKE_SWIZZLE4(src1_swiz, src1_swiz,
+                                 src1_swiz, src1_swiz);
+
+      inst = emit(ir, op, dst, src0, src1);
+      inst->dst.writemask = this_mask;
+      done_mask |= this_mask;
+   }
+}
+
+void
+glsl_to_tgsi_visitor::emit_scalar(ir_instruction *ir, unsigned op,
+                               st_dst_reg dst, st_src_reg src0)
+{
+   st_src_reg undef = undef_src;
+
+   undef.swizzle = SWIZZLE_XXXX;
+
+   emit_scalar(ir, op, dst, src0, undef);
+}
+
+void
+glsl_to_tgsi_visitor::emit_arl(ir_instruction *ir,
+                               st_dst_reg dst, st_src_reg src0)
+{
+   st_src_reg tmp = get_temp(glsl_type::float_type);
+
+   if (src0.type == GLSL_TYPE_INT)
+      emit(NULL, TGSI_OPCODE_I2F, st_dst_reg(tmp), src0);
+   else if (src0.type == GLSL_TYPE_UINT)
+      emit(NULL, TGSI_OPCODE_U2F, st_dst_reg(tmp), src0);
+   else
+      tmp = src0;
+   
+   emit(NULL, TGSI_OPCODE_ARL, dst, tmp);
+}
+
+/**
+ * Emit an TGSI_OPCODE_SCS instruction
+ *
+ * The \c SCS opcode functions a bit differently than the other TGSI opcodes.
+ * Instead of splatting its result across all four components of the 
+ * destination, it writes one value to the \c x component and another value to 
+ * the \c y component.
+ *
+ * \param ir        IR instruction being processed
+ * \param op        Either \c TGSI_OPCODE_SIN or \c TGSI_OPCODE_COS depending 
+ *                  on which value is desired.
+ * \param dst       Destination register
+ * \param src       Source register
+ */
+void
+glsl_to_tgsi_visitor::emit_scs(ir_instruction *ir, unsigned op,
+                            st_dst_reg dst,
+                            const st_src_reg &src)
+{
+   /* Vertex programs cannot use the SCS opcode.
+    */
+   if (this->prog->Target == GL_VERTEX_PROGRAM_ARB) {
+      emit_scalar(ir, op, dst, src);
+      return;
+   }
+
+   const unsigned component = (op == TGSI_OPCODE_SIN) ? 0 : 1;
+   const unsigned scs_mask = (1U << component);
+   int done_mask = ~dst.writemask;
+   st_src_reg tmp;
+
+   assert(op == TGSI_OPCODE_SIN || op == TGSI_OPCODE_COS);
+
+   /* If there are compnents in the destination that differ from the component
+    * that will be written by the SCS instrution, we'll need a temporary.
+    */
+   if (scs_mask != unsigned(dst.writemask)) {
+      tmp = get_temp(glsl_type::vec4_type);
+   }
+
+   for (unsigned i = 0; i < 4; i++) {
+      unsigned this_mask = (1U << i);
+      st_src_reg src0 = src;
+
+      if ((done_mask & this_mask) != 0)
+         continue;
+
+      /* The source swizzle specified which component of the source generates
+       * sine / cosine for the current component in the destination.  The SCS
+       * instruction requires that this value be swizzle to the X component.
+       * Replace the current swizzle with a swizzle that puts the source in
+       * the X component.
+       */
+      unsigned src0_swiz = GET_SWZ(src.swizzle, i);
+
+      src0.swizzle = MAKE_SWIZZLE4(src0_swiz, src0_swiz,
+                                  src0_swiz, src0_swiz);
+      for (unsigned j = i + 1; j < 4; j++) {
+         /* If there is another enabled component in the destination that is
+          * derived from the same inputs, generate its value on this pass as
+          * well.
+          */
+         if (!(done_mask & (1 << j)) &&
+             GET_SWZ(src0.swizzle, j) == src0_swiz) {
+            this_mask |= (1 << j);
+         }
+      }
+
+      if (this_mask != scs_mask) {
+         glsl_to_tgsi_instruction *inst;
+         st_dst_reg tmp_dst = st_dst_reg(tmp);
+
+         /* Emit the SCS instruction.
+          */
+         inst = emit(ir, TGSI_OPCODE_SCS, tmp_dst, src0);
+         inst->dst.writemask = scs_mask;
+
+         /* Move the result of the SCS instruction to the desired location in
+          * the destination.
+          */
+         tmp.swizzle = MAKE_SWIZZLE4(component, component,
+                                    component, component);
+         inst = emit(ir, TGSI_OPCODE_SCS, dst, tmp);
+         inst->dst.writemask = this_mask;
+      } else {
+         /* Emit the SCS instruction to write directly to the destination.
+          */
+         glsl_to_tgsi_instruction *inst = emit(ir, TGSI_OPCODE_SCS, dst, src0);
+         inst->dst.writemask = scs_mask;
+      }
+
+      done_mask |= this_mask;
+   }
+}
+
+int
+glsl_to_tgsi_visitor::add_constant(gl_register_file file,
+                            gl_constant_value values[4], int size, int datatype,
+                            GLuint *swizzle_out)
+{
+   if (file == PROGRAM_CONSTANT) {
+      return _mesa_add_typed_unnamed_constant(this->prog->Parameters, values,
+                                              size, datatype, swizzle_out);
+   } else {
+      int index = 0;
+      immediate_storage *entry;
+      assert(file == PROGRAM_IMMEDIATE);
+
+      /* Search immediate storage to see if we already have an identical
+       * immediate that we can use instead of adding a duplicate entry.
+       */
+      foreach_iter(exec_list_iterator, iter, this->immediates) {
+         entry = (immediate_storage *)iter.get();
+         
+         if (entry->size == size &&
+             entry->type == datatype &&
+             !memcmp(entry->values, values, size * sizeof(gl_constant_value))) {
+             return index;
+         }
+         index++;
+      }
+      
+      /* Add this immediate to the list. */
+      entry = new(mem_ctx) immediate_storage(values, size, datatype);
+      this->immediates.push_tail(entry);
+      this->num_immediates++;
+      return index;
+   }
+}
+
+struct st_src_reg
+glsl_to_tgsi_visitor::st_src_reg_for_float(float val)
+{
+   st_src_reg src(PROGRAM_IMMEDIATE, -1, GLSL_TYPE_FLOAT);
+   union gl_constant_value uval;
+
+   uval.f = val;
+   src.index = add_constant(src.file, &uval, 1, GL_FLOAT, &src.swizzle);
+
+   return src;
+}
+
+struct st_src_reg
+glsl_to_tgsi_visitor::st_src_reg_for_int(int val)
+{
+   st_src_reg src(PROGRAM_IMMEDIATE, -1, GLSL_TYPE_INT);
+   union gl_constant_value uval;
+   
+   assert(glsl_version >= 130);
+
+   uval.i = val;
+   src.index = add_constant(src.file, &uval, 1, GL_INT, &src.swizzle);
+
+   return src;
+}
+
+struct st_src_reg
+glsl_to_tgsi_visitor::st_src_reg_for_type(int type, int val)
+{
+   if (glsl_version >= 130)
+      return type == GLSL_TYPE_FLOAT ? st_src_reg_for_float(val) : 
+                                       st_src_reg_for_int(val);
+   else
+      return st_src_reg_for_float(val);
+}
+
+static int
+type_size(const struct glsl_type *type)
+{
+   unsigned int i;
+   int size;
+
+   switch (type->base_type) {
+   case GLSL_TYPE_UINT:
+   case GLSL_TYPE_INT:
+   case GLSL_TYPE_FLOAT:
+   case GLSL_TYPE_BOOL:
+      if (type->is_matrix()) {
+         return type->matrix_columns;
+      } else {
+         /* Regardless of size of vector, it gets a vec4. This is bad
+          * packing for things like floats, but otherwise arrays become a
+          * mess.  Hopefully a later pass over the code can pack scalars
+          * down if appropriate.
+          */
+         return 1;
+      }
+   case GLSL_TYPE_ARRAY:
+      assert(type->length > 0);
+      return type_size(type->fields.array) * type->length;
+   case GLSL_TYPE_STRUCT:
+      size = 0;
+      for (i = 0; i < type->length; i++) {
+         size += type_size(type->fields.structure[i].type);
+      }
+      return size;
+   case GLSL_TYPE_SAMPLER:
+      /* Samplers take up one slot in UNIFORMS[], but they're baked in
+       * at link time.
+       */
+      return 1;
+   default:
+      assert(0);
+      return 0;
+   }
+}
+
+/**
+ * In the initial pass of codegen, we assign temporary numbers to
+ * intermediate results.  (not SSA -- variable assignments will reuse
+ * storage).
+ */
+st_src_reg
+glsl_to_tgsi_visitor::get_temp(const glsl_type *type)
+{
+   st_src_reg src;
+   int swizzle[4];
+   int i;
+
+   src.type = glsl_version >= 130 ? type->base_type : GLSL_TYPE_FLOAT;
+   src.file = PROGRAM_TEMPORARY;
+   src.index = next_temp;
+   src.reladdr = NULL;
+   next_temp += type_size(type);
+
+   if (type->is_array() || type->is_record()) {
+      src.swizzle = SWIZZLE_NOOP;
+   } else {
+      for (i = 0; i < type->vector_elements; i++)
+         swizzle[i] = i;
+      for (; i < 4; i++)
+         swizzle[i] = type->vector_elements - 1;
+      src.swizzle = MAKE_SWIZZLE4(swizzle[0], swizzle[1],
+                                 swizzle[2], swizzle[3]);
+   }
+   src.negate = 0;
+
+   return src;
+}
+
+variable_storage *
+glsl_to_tgsi_visitor::find_variable_storage(ir_variable *var)
+{
+   
+   variable_storage *entry;
+
+   foreach_iter(exec_list_iterator, iter, this->variables) {
+      entry = (variable_storage *)iter.get();
+
+      if (entry->var == var)
+         return entry;
+   }
+
+   return NULL;
+}
+
+void
+glsl_to_tgsi_visitor::visit(ir_variable *ir)
+{
+   if (strcmp(ir->name, "gl_FragCoord") == 0) {
+      struct gl_fragment_program *fp = (struct gl_fragment_program *)this->prog;
+
+      fp->OriginUpperLeft = ir->origin_upper_left;
+      fp->PixelCenterInteger = ir->pixel_center_integer;
+
+   } else if (strcmp(ir->name, "gl_FragDepth") == 0) {
+      struct gl_fragment_program *fp = (struct gl_fragment_program *)this->prog;
+      switch (ir->depth_layout) {
+      case ir_depth_layout_none:
+         fp->FragDepthLayout = FRAG_DEPTH_LAYOUT_NONE;
+         break;
+      case ir_depth_layout_any:
+         fp->FragDepthLayout = FRAG_DEPTH_LAYOUT_ANY;
+         break;
+      case ir_depth_layout_greater:
+         fp->FragDepthLayout = FRAG_DEPTH_LAYOUT_GREATER;
+         break;
+      case ir_depth_layout_less:
+         fp->FragDepthLayout = FRAG_DEPTH_LAYOUT_LESS;
+         break;
+      case ir_depth_layout_unchanged:
+         fp->FragDepthLayout = FRAG_DEPTH_LAYOUT_UNCHANGED;
+         break;
+      default:
+         assert(0);
+         break;
+      }
+   }
+
+   if (ir->mode == ir_var_uniform && strncmp(ir->name, "gl_", 3) == 0) {
+      unsigned int i;
+      const ir_state_slot *const slots = ir->state_slots;
+      assert(ir->state_slots != NULL);
+
+      /* Check if this statevar's setup in the STATE file exactly
+       * matches how we'll want to reference it as a
+       * struct/array/whatever.  If not, then we need to move it into
+       * temporary storage and hope that it'll get copy-propagated
+       * out.
+       */
+      for (i = 0; i < ir->num_state_slots; i++) {
+         if (slots[i].swizzle != SWIZZLE_XYZW) {
+            break;
+         }
+      }
+
+      struct variable_storage *storage;
+      st_dst_reg dst;
+      if (i == ir->num_state_slots) {
+         /* We'll set the index later. */
+         storage = new(mem_ctx) variable_storage(ir, PROGRAM_STATE_VAR, -1);
+         this->variables.push_tail(storage);
+
+         dst = undef_dst;
+      } else {
+         /* The variable_storage constructor allocates slots based on the size
+          * of the type.  However, this had better match the number of state
+          * elements that we're going to copy into the new temporary.
+          */
+         assert((int) ir->num_state_slots == type_size(ir->type));
+
+         storage = new(mem_ctx) variable_storage(ir, PROGRAM_TEMPORARY,
+                                                this->next_temp);
+         this->variables.push_tail(storage);
+         this->next_temp += type_size(ir->type);
+
+         dst = st_dst_reg(st_src_reg(PROGRAM_TEMPORARY, storage->index,
+               glsl_version >= 130 ? ir->type->base_type : GLSL_TYPE_FLOAT));
+      }
+
+
+      for (unsigned int i = 0; i < ir->num_state_slots; i++) {
+         int index = _mesa_add_state_reference(this->prog->Parameters,
+                                              (gl_state_index *)slots[i].tokens);
+
+         if (storage->file == PROGRAM_STATE_VAR) {
+            if (storage->index == -1) {
+               storage->index = index;
+            } else {
+               assert(index == storage->index + (int)i);
+            }
+         } else {
+            st_src_reg src(PROGRAM_STATE_VAR, index,
+                  glsl_version >= 130 ? ir->type->base_type : GLSL_TYPE_FLOAT);
+            src.swizzle = slots[i].swizzle;
+            emit(ir, TGSI_OPCODE_MOV, dst, src);
+            /* even a float takes up a whole vec4 reg in a struct/array. */
+            dst.index++;
+         }
+      }
+
+      if (storage->file == PROGRAM_TEMPORARY &&
+          dst.index != storage->index + (int) ir->num_state_slots) {
+         fail_link(this->shader_program,
+                  "failed to load builtin uniform `%s'  (%d/%d regs loaded)\n",
+                  ir->name, dst.index - storage->index,
+                  type_size(ir->type));
+      }
+   }
+}
+
+void
+glsl_to_tgsi_visitor::visit(ir_loop *ir)
+{
+   ir_dereference_variable *counter = NULL;
+
+   if (ir->counter != NULL)
+      counter = new(ir) ir_dereference_variable(ir->counter);
+
+   if (ir->from != NULL) {
+      assert(ir->counter != NULL);
+
+      ir_assignment *a = new(ir) ir_assignment(counter, ir->from, NULL);
+
+      a->accept(this);
+      delete a;
+   }
+
+   emit(NULL, TGSI_OPCODE_BGNLOOP);
+
+   if (ir->to) {
+      ir_expression *e =
+         new(ir) ir_expression(ir->cmp, glsl_type::bool_type,
+                              counter, ir->to);
+      ir_if *if_stmt =  new(ir) ir_if(e);
+
+      ir_loop_jump *brk = new(ir) ir_loop_jump(ir_loop_jump::jump_break);
+
+      if_stmt->then_instructions.push_tail(brk);
+
+      if_stmt->accept(this);
+
+      delete if_stmt;
+      delete e;
+      delete brk;
+   }
+
+   visit_exec_list(&ir->body_instructions, this);
+
+   if (ir->increment) {
+      ir_expression *e =
+         new(ir) ir_expression(ir_binop_add, counter->type,
+                              counter, ir->increment);
+
+      ir_assignment *a = new(ir) ir_assignment(counter, e, NULL);
+
+      a->accept(this);
+      delete a;
+      delete e;
+   }
+
+   emit(NULL, TGSI_OPCODE_ENDLOOP);
+}
+
+void
+glsl_to_tgsi_visitor::visit(ir_loop_jump *ir)
+{
+   switch (ir->mode) {
+   case ir_loop_jump::jump_break:
+      emit(NULL, TGSI_OPCODE_BRK);
+      break;
+   case ir_loop_jump::jump_continue:
+      emit(NULL, TGSI_OPCODE_CONT);
+      break;
+   }
+}
+
+
+void
+glsl_to_tgsi_visitor::visit(ir_function_signature *ir)
+{
+   assert(0);
+   (void)ir;
+}
+
+void
+glsl_to_tgsi_visitor::visit(ir_function *ir)
+{
+   /* Ignore function bodies other than main() -- we shouldn't see calls to
+    * them since they should all be inlined before we get to glsl_to_tgsi.
+    */
+   if (strcmp(ir->name, "main") == 0) {
+      const ir_function_signature *sig;
+      exec_list empty;
+
+      sig = ir->matching_signature(&empty);
+
+      assert(sig);
+
+      foreach_iter(exec_list_iterator, iter, sig->body) {
+         ir_instruction *ir = (ir_instruction *)iter.get();
+
+         ir->accept(this);
+      }
+   }
+}
+
+GLboolean
+glsl_to_tgsi_visitor::try_emit_mad(ir_expression *ir, int mul_operand)
+{
+   int nonmul_operand = 1 - mul_operand;
+   st_src_reg a, b, c;
+   st_dst_reg result_dst;
+
+   ir_expression *expr = ir->operands[mul_operand]->as_expression();
+   if (!expr || expr->operation != ir_binop_mul)
+      return false;
+
+   expr->operands[0]->accept(this);
+   a = this->result;
+   expr->operands[1]->accept(this);
+   b = this->result;
+   ir->operands[nonmul_operand]->accept(this);
+   c = this->result;
+
+   this->result = get_temp(ir->type);
+   result_dst = st_dst_reg(this->result);
+   result_dst.writemask = (1 << ir->type->vector_elements) - 1;
+   emit(ir, TGSI_OPCODE_MAD, result_dst, a, b, c);
+
+   return true;
+}
+
+GLboolean
+glsl_to_tgsi_visitor::try_emit_sat(ir_expression *ir)
+{
+   /* Saturates were only introduced to vertex programs in
+    * NV_vertex_program3, so don't give them to drivers in the VP.
+    */
+   if (this->prog->Target == GL_VERTEX_PROGRAM_ARB)
+      return false;
+
+   ir_rvalue *sat_src = ir->as_rvalue_to_saturate();
+   if (!sat_src)
+      return false;
+
+   sat_src->accept(this);
+   st_src_reg src = this->result;
+
+   this->result = get_temp(ir->type);
+   st_dst_reg result_dst = st_dst_reg(this->result);
+   result_dst.writemask = (1 << ir->type->vector_elements) - 1;
+   glsl_to_tgsi_instruction *inst;
+   inst = emit(ir, TGSI_OPCODE_MOV, result_dst, src);
+   inst->saturate = true;
+
+   return true;
+}
+
+void
+glsl_to_tgsi_visitor::reladdr_to_temp(ir_instruction *ir,
+                                   st_src_reg *reg, int *num_reladdr)
+{
+   if (!reg->reladdr)
+      return;
+
+   emit_arl(ir, address_reg, *reg->reladdr);
+
+   if (*num_reladdr != 1) {
+      st_src_reg temp = get_temp(glsl_type::vec4_type);
+
+      emit(ir, TGSI_OPCODE_MOV, st_dst_reg(temp), *reg);
+      *reg = temp;
+   }
+
+   (*num_reladdr)--;
+}
+
+void
+glsl_to_tgsi_visitor::visit(ir_expression *ir)
+{
+   unsigned int operand;
+   st_src_reg op[Elements(ir->operands)];
+   st_src_reg result_src;
+   st_dst_reg result_dst;
+
+   /* Quick peephole: Emit MAD(a, b, c) instead of ADD(MUL(a, b), c)
+    */
+   if (ir->operation == ir_binop_add) {
+      if (try_emit_mad(ir, 1))
+         return;
+      if (try_emit_mad(ir, 0))
+         return;
+   }
+   if (try_emit_sat(ir))
+      return;
+
+   if (ir->operation == ir_quadop_vector)
+      assert(!"ir_quadop_vector should have been lowered");
+
+   for (operand = 0; operand < ir->get_num_operands(); operand++) {
+      this->result.file = PROGRAM_UNDEFINED;
+      ir->operands[operand]->accept(this);
+      if (this->result.file == PROGRAM_UNDEFINED) {
+         ir_print_visitor v;
+         printf("Failed to get tree for expression operand:\n");
+         ir->operands[operand]->accept(&v);
+         exit(1);
+      }
+      op[operand] = this->result;
+
+      /* Matrix expression operands should have been broken down to vector
+       * operations already.
+       */
+      assert(!ir->operands[operand]->type->is_matrix());
+   }
+
+   int vector_elements = ir->operands[0]->type->vector_elements;
+   if (ir->operands[1]) {
+      vector_elements = MAX2(vector_elements,
+                            ir->operands[1]->type->vector_elements);
+   }
+
+   this->result.file = PROGRAM_UNDEFINED;
+
+   /* Storage for our result.  Ideally for an assignment we'd be using
+    * the actual storage for the result here, instead.
+    */
+   result_src = get_temp(ir->type);
+   /* convenience for the emit functions below. */
+   result_dst = st_dst_reg(result_src);
+   /* Limit writes to the channels that will be used by result_src later.
+    * This does limit this temp's use as a temporary for multi-instruction
+    * sequences.
+    */
+   result_dst.writemask = (1 << ir->type->vector_elements) - 1;
+
+   switch (ir->operation) {
+   case ir_unop_logic_not:
+      emit(ir, TGSI_OPCODE_SEQ, result_dst, op[0], st_src_reg_for_type(result_dst.type, 0));
+      break;
+   case ir_unop_neg:
+      assert(result_dst.type == GLSL_TYPE_FLOAT || result_dst.type == GLSL_TYPE_INT);
+      if (result_dst.type == GLSL_TYPE_INT)
+         emit(ir, TGSI_OPCODE_INEG, result_dst, op[0]);
+      else {
+         op[0].negate = ~op[0].negate;
+         result_src = op[0];
+      }
+      break;
+   case ir_unop_abs:
+      assert(result_dst.type == GLSL_TYPE_FLOAT);
+      emit(ir, TGSI_OPCODE_ABS, result_dst, op[0]);
+      break;
+   case ir_unop_sign:
+      emit(ir, TGSI_OPCODE_SSG, result_dst, op[0]);
+      break;
+   case ir_unop_rcp:
+      emit_scalar(ir, TGSI_OPCODE_RCP, result_dst, op[0]);
+      break;
+
+   case ir_unop_exp2:
+      emit_scalar(ir, TGSI_OPCODE_EX2, result_dst, op[0]);
+      break;
+   case ir_unop_exp:
+   case ir_unop_log:
+      assert(!"not reached: should be handled by ir_explog_to_explog2");
+      break;
+   case ir_unop_log2:
+      emit_scalar(ir, TGSI_OPCODE_LG2, result_dst, op[0]);
+      break;
+   case ir_unop_sin:
+      emit_scalar(ir, TGSI_OPCODE_SIN, result_dst, op[0]);
+      break;
+   case ir_unop_cos:
+      emit_scalar(ir, TGSI_OPCODE_COS, result_dst, op[0]);
+      break;
+   case ir_unop_sin_reduced:
+      emit_scs(ir, TGSI_OPCODE_SIN, result_dst, op[0]);
+      break;
+   case ir_unop_cos_reduced:
+      emit_scs(ir, TGSI_OPCODE_COS, result_dst, op[0]);
+      break;
+
+   case ir_unop_dFdx:
+      emit(ir, TGSI_OPCODE_DDX, result_dst, op[0]);
+      break;
+   case ir_unop_dFdy:
+      op[0].negate = ~op[0].negate;
+      emit(ir, TGSI_OPCODE_DDY, result_dst, op[0]);
+      break;
+
+   case ir_unop_noise: {
+      /* At some point, a motivated person could add a better
+       * implementation of noise.  Currently not even the nvidia
+       * binary drivers do anything more than this.  In any case, the
+       * place to do this is in the GL state tracker, not the poor
+       * driver.
+       */
+      emit(ir, TGSI_OPCODE_MOV, result_dst, st_src_reg_for_float(0.5));
+      break;
+   }
+
+   case ir_binop_add:
+      emit(ir, TGSI_OPCODE_ADD, result_dst, op[0], op[1]);
+      break;
+   case ir_binop_sub:
+      emit(ir, TGSI_OPCODE_SUB, result_dst, op[0], op[1]);
+      break;
+
+   case ir_binop_mul:
+      emit(ir, TGSI_OPCODE_MUL, result_dst, op[0], op[1]);
+      break;
+   case ir_binop_div:
+      if (result_dst.type == GLSL_TYPE_FLOAT)
+         assert(!"not reached: should be handled by ir_div_to_mul_rcp");
+      else
+         emit(ir, TGSI_OPCODE_DIV, result_dst, op[0], op[1]);
+      break;
+   case ir_binop_mod:
+      if (result_dst.type == GLSL_TYPE_FLOAT)
+         assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
+      else
+         emit(ir, TGSI_OPCODE_MOD, result_dst, op[0], op[1]);
+      break;
+
+   case ir_binop_less:
+      emit(ir, TGSI_OPCODE_SLT, result_dst, op[0], op[1]);
+      break;
+   case ir_binop_greater:
+      emit(ir, TGSI_OPCODE_SGT, result_dst, op[0], op[1]);
+      break;
+   case ir_binop_lequal:
+      emit(ir, TGSI_OPCODE_SLE, result_dst, op[0], op[1]);
+      break;
+   case ir_binop_gequal:
+      emit(ir, TGSI_OPCODE_SGE, result_dst, op[0], op[1]);
+      break;
+   case ir_binop_equal:
+      emit(ir, TGSI_OPCODE_SEQ, result_dst, op[0], op[1]);
+      break;
+   case ir_binop_nequal:
+      emit(ir, TGSI_OPCODE_SNE, result_dst, op[0], op[1]);
+      break;
+   case ir_binop_all_equal:
+      /* "==" operator producing a scalar boolean. */
+      if (ir->operands[0]->type->is_vector() ||
+          ir->operands[1]->type->is_vector()) {
+         st_src_reg temp = get_temp(glsl_version >= 130 ? 
+               glsl_type::get_instance(ir->operands[0]->type->base_type, 4, 1) :
+               glsl_type::vec4_type);
+         assert(ir->operands[0]->type->base_type == GLSL_TYPE_FLOAT);
+         emit(ir, TGSI_OPCODE_SNE, st_dst_reg(temp), op[0], op[1]);
+         emit_dp(ir, result_dst, temp, temp, vector_elements);
+         emit(ir, TGSI_OPCODE_SEQ, result_dst, result_src, st_src_reg_for_float(0.0));
+      } else {
+         emit(ir, TGSI_OPCODE_SEQ, result_dst, op[0], op[1]);
+      }
+      break;
+   case ir_binop_any_nequal:
+      /* "!=" operator producing a scalar boolean. */
+      if (ir->operands[0]->type->is_vector() ||
+          ir->operands[1]->type->is_vector()) {
+         st_src_reg temp = get_temp(glsl_version >= 130 ? 
+               glsl_type::get_instance(ir->operands[0]->type->base_type, 4, 1) :
+               glsl_type::vec4_type);
+         assert(ir->operands[0]->type->base_type == GLSL_TYPE_FLOAT);
+         emit(ir, TGSI_OPCODE_SNE, st_dst_reg(temp), op[0], op[1]);
+         emit_dp(ir, result_dst, temp, temp, vector_elements);
+         emit(ir, TGSI_OPCODE_SNE, result_dst, result_src, st_src_reg_for_float(0.0));
+      } else {
+         emit(ir, TGSI_OPCODE_SNE, result_dst, op[0], op[1]);
+      }
+      break;
+
+   case ir_unop_any:
+      assert(ir->operands[0]->type->is_vector());
+      emit_dp(ir, result_dst, op[0], op[0],
+              ir->operands[0]->type->vector_elements);
+      emit(ir, TGSI_OPCODE_SNE, result_dst, result_src, st_src_reg_for_float(0.0));
+      break;
+
+   case ir_binop_logic_xor:
+      emit(ir, TGSI_OPCODE_SNE, result_dst, op[0], op[1]);
+      break;
+
+   case ir_binop_logic_or:
+      /* This could be a saturated add and skip the SNE. */
+      emit(ir, TGSI_OPCODE_ADD, result_dst, op[0], op[1]);
+      emit(ir, TGSI_OPCODE_SNE, result_dst, result_src, st_src_reg_for_float(0.0));
+      break;
+
+   case ir_binop_logic_and:
+      /* the bool args are stored as float 0.0 or 1.0, so "mul" gives us "and". */
+      emit(ir, TGSI_OPCODE_MUL, result_dst, op[0], op[1]);
+      break;
+
+   case ir_binop_dot:
+      assert(ir->operands[0]->type->is_vector());
+      assert(ir->operands[0]->type == ir->operands[1]->type);
+      emit_dp(ir, result_dst, op[0], op[1],
+              ir->operands[0]->type->vector_elements);
+      break;
+
+   case ir_unop_sqrt:
+      /* sqrt(x) = x * rsq(x). */
+      emit_scalar(ir, TGSI_OPCODE_RSQ, result_dst, op[0]);
+      emit(ir, TGSI_OPCODE_MUL, result_dst, result_src, op[0]);
+      /* For incoming channels <= 0, set the result to 0. */
+      op[0].negate = ~op[0].negate;
+      emit(ir, TGSI_OPCODE_CMP, result_dst,
+                         op[0], result_src, st_src_reg_for_float(0.0));
+      break;
+   case ir_unop_rsq:
+      emit_scalar(ir, TGSI_OPCODE_RSQ, result_dst, op[0]);
+      break;
+   case ir_unop_i2f:
+   case ir_unop_b2f:
+      if (glsl_version >= 130) {
+         emit(ir, TGSI_OPCODE_I2F, result_dst, op[0]);
+         break;
+      }
+   case ir_unop_i2u:
+   case ir_unop_u2i:
+      /* Converting between signed and unsigned integers is a no-op. */
+   case ir_unop_b2i:
+      /* Booleans are stored as integers (or floats in GLSL 1.20 and lower). */
+      result_src = op[0];
+      break;
+   case ir_unop_f2i:
+      if (glsl_version >= 130)
+         emit(ir, TGSI_OPCODE_F2I, result_dst, op[0]);
+      else
+         emit(ir, TGSI_OPCODE_TRUNC, result_dst, op[0]);
+      break;
+   case ir_unop_f2b:
+   case ir_unop_i2b:
+      emit(ir, TGSI_OPCODE_SNE, result_dst, op[0], 
+            st_src_reg_for_type(result_dst.type, 0));
+      break;
+   case ir_unop_trunc:
+      emit(ir, TGSI_OPCODE_TRUNC, result_dst, op[0]);
+      break;
+   case ir_unop_ceil:
+      op[0].negate = ~op[0].negate;
+      emit(ir, TGSI_OPCODE_FLR, result_dst, op[0]);
+      result_src.negate = ~result_src.negate;
+      break;
+   case ir_unop_floor:
+      emit(ir, TGSI_OPCODE_FLR, result_dst, op[0]);
+      break;
+   case ir_unop_fract:
+      emit(ir, TGSI_OPCODE_FRC, result_dst, op[0]);
+      break;
+
+   case ir_binop_min:
+      emit(ir, TGSI_OPCODE_MIN, result_dst, op[0], op[1]);
+      break;
+   case ir_binop_max:
+      emit(ir, TGSI_OPCODE_MAX, result_dst, op[0], op[1]);
+      break;
+   case ir_binop_pow:
+      emit_scalar(ir, TGSI_OPCODE_POW, result_dst, op[0], op[1]);
+      break;
+
+   case ir_unop_bit_not:
+      if (glsl_version >= 130) {
+         emit(ir, TGSI_OPCODE_NOT, result_dst, op[0]);
+         break;
+      }
+   case ir_unop_u2f:
+      if (glsl_version >= 130) {
+         emit(ir, TGSI_OPCODE_U2F, result_dst, op[0]);
+         break;
+      }
+   case ir_binop_lshift:
+      if (glsl_version >= 130) {
+         emit(ir, TGSI_OPCODE_SHL, result_dst, op[0]);
+         break;
+      }
+   case ir_binop_rshift:
+      if (glsl_version >= 130) {
+         emit(ir, TGSI_OPCODE_ISHR, result_dst, op[0]);
+         break;
+      }
+   case ir_binop_bit_and:
+      if (glsl_version >= 130) {
+         emit(ir, TGSI_OPCODE_AND, result_dst, op[0]);
+         break;
+      }
+   case ir_binop_bit_xor:
+      if (glsl_version >= 130) {
+         emit(ir, TGSI_OPCODE_XOR, result_dst, op[0]);
+         break;
+      }
+   case ir_binop_bit_or:
+      if (glsl_version >= 130) {
+         emit(ir, TGSI_OPCODE_OR, result_dst, op[0]);
+         break;
+      }
+   case ir_unop_round_even:
+      assert(!"GLSL 1.30 features unsupported");
+      break;
+
+   case ir_quadop_vector:
+      /* This operation should have already been handled.
+       */
+      assert(!"Should not get here.");
+      break;
+   }
+
+   this->result = result_src;
+}
+
+
+void
+glsl_to_tgsi_visitor::visit(ir_swizzle *ir)
+{
+   st_src_reg src;
+   int i;
+   int swizzle[4];
+
+   /* Note that this is only swizzles in expressions, not those on the left
+    * hand side of an assignment, which do write masking.  See ir_assignment
+    * for that.
+    */
+
+   ir->val->accept(this);
+   src = this->result;
+   assert(src.file != PROGRAM_UNDEFINED);
+
+   for (i = 0; i < 4; i++) {
+      if (i < ir->type->vector_elements) {
+         switch (i) {
+         case 0:
+            swizzle[i] = GET_SWZ(src.swizzle, ir->mask.x);
+            break;
+         case 1:
+            swizzle[i] = GET_SWZ(src.swizzle, ir->mask.y);
+            break;
+         case 2:
+            swizzle[i] = GET_SWZ(src.swizzle, ir->mask.z);
+            break;
+         case 3:
+            swizzle[i] = GET_SWZ(src.swizzle, ir->mask.w);
+            break;
+         }
+      } else {
+         /* If the type is smaller than a vec4, replicate the last
+          * channel out.
+          */
+         swizzle[i] = swizzle[ir->type->vector_elements - 1];
+      }
+   }
+
+   src.swizzle = MAKE_SWIZZLE4(swizzle[0], swizzle[1], swizzle[2], swizzle[3]);
+
+   this->result = src;
+}
+
+void
+glsl_to_tgsi_visitor::visit(ir_dereference_variable *ir)
+{
+   variable_storage *entry = find_variable_storage(ir->var);
+   ir_variable *var = ir->var;
+
+   if (!entry) {
+      switch (var->mode) {
+      case ir_var_uniform:
+         entry = new(mem_ctx) variable_storage(var, PROGRAM_UNIFORM,
+                                              var->location);
+         this->variables.push_tail(entry);
+         break;
+      case ir_var_in:
+      case ir_var_inout:
+         /* The linker assigns locations for varyings and attributes,
+          * including deprecated builtins (like gl_Color), user-assign
+          * generic attributes (glBindVertexLocation), and
+          * user-defined varyings.
+          *
+          * FINISHME: We would hit this path for function arguments.  Fix!
+          */
+         assert(var->location != -1);
+         entry = new(mem_ctx) variable_storage(var,
+                                               PROGRAM_INPUT,
+                                               var->location);
+         if (this->prog->Target == GL_VERTEX_PROGRAM_ARB &&
+             var->location >= VERT_ATTRIB_GENERIC0) {
+            _mesa_add_attribute(this->prog->Attributes,
+                                var->name,
+                                _mesa_sizeof_glsl_type(var->type->gl_type),
+                                var->type->gl_type,
+                                var->location - VERT_ATTRIB_GENERIC0);
+         }
+         break;
+      case ir_var_out:
+         assert(var->location != -1);
+         entry = new(mem_ctx) variable_storage(var,
+                                               PROGRAM_OUTPUT,
+                                               var->location);
+         break;
+      case ir_var_system_value:
+         entry = new(mem_ctx) variable_storage(var,
+                                               PROGRAM_SYSTEM_VALUE,
+                                               var->location);
+         break;
+      case ir_var_auto:
+      case ir_var_temporary:
+         entry = new(mem_ctx) variable_storage(var, PROGRAM_TEMPORARY,
+                                              this->next_temp);
+         this->variables.push_tail(entry);
+
+         next_temp += type_size(var->type);
+         break;
+      }
+
+      if (!entry) {
+         printf("Failed to make storage for %s\n", var->name);
+         exit(1);
+      }
+   }
+
+   this->result = st_src_reg(entry->file, entry->index, var->type);
+   if (glsl_version <= 120)
+      this->result.type = GLSL_TYPE_FLOAT;
+}
+
+void
+glsl_to_tgsi_visitor::visit(ir_dereference_array *ir)
+{
+   ir_constant *index;
+   st_src_reg src;
+   int element_size = type_size(ir->type);
+
+   index = ir->array_index->constant_expression_value();
+
+   ir->array->accept(this);
+   src = this->result;
+
+   if (index) {
+      src.index += index->value.i[0] * element_size;
+   } else {
+      st_src_reg array_base = this->result;
+      /* Variable index array dereference.  It eats the "vec4" of the
+       * base of the array and an index that offsets the TGSI register
+       * index.
+       */
+      ir->array_index->accept(this);
+
+      st_src_reg index_reg;
+
+      if (element_size == 1) {
+         index_reg = this->result;
+      } else {
+         index_reg = get_temp(glsl_type::float_type);
+
+         emit(ir, TGSI_OPCODE_MUL, st_dst_reg(index_reg),
+              this->result, st_src_reg_for_float(element_size));
+      }
+
+      /* If there was already a relative address register involved, add the
+       * new and the old together to get the new offset.
+       */
+      if (src.reladdr != NULL) {
+         st_src_reg accum_reg = get_temp(glsl_type::float_type);
+
+         emit(ir, TGSI_OPCODE_ADD, st_dst_reg(accum_reg),
+              index_reg, *src.reladdr);
+
+         index_reg = accum_reg;
+      }
+
+      src.reladdr = ralloc(mem_ctx, st_src_reg);
+      memcpy(src.reladdr, &index_reg, sizeof(index_reg));
+   }
+
+   /* If the type is smaller than a vec4, replicate the last channel out. */
+   if (ir->type->is_scalar() || ir->type->is_vector())
+      src.swizzle = swizzle_for_size(ir->type->vector_elements);
+   else
+      src.swizzle = SWIZZLE_NOOP;
+
+   this->result = src;
+}
+
+void
+glsl_to_tgsi_visitor::visit(ir_dereference_record *ir)
+{
+   unsigned int i;
+   const glsl_type *struct_type = ir->record->type;
+   int offset = 0;
+
+   ir->record->accept(this);
+
+   for (i = 0; i < struct_type->length; i++) {
+      if (strcmp(struct_type->fields.structure[i].name, ir->field) == 0)
+         break;
+      offset += type_size(struct_type->fields.structure[i].type);
+   }
+
+   /* If the type is smaller than a vec4, replicate the last channel out. */
+   if (ir->type->is_scalar() || ir->type->is_vector())
+      this->result.swizzle = swizzle_for_size(ir->type->vector_elements);
+   else
+      this->result.swizzle = SWIZZLE_NOOP;
+
+   this->result.index += offset;
+}
+
+/**
+ * We want to be careful in assignment setup to hit the actual storage
+ * instead of potentially using a temporary like we might with the
+ * ir_dereference handler.
+ */
+static st_dst_reg
+get_assignment_lhs(ir_dereference *ir, glsl_to_tgsi_visitor *v)
+{
+   /* The LHS must be a dereference.  If the LHS is a variable indexed array
+    * access of a vector, it must be separated into a series conditional moves
+    * before reaching this point (see ir_vec_index_to_cond_assign).
+    */
+   assert(ir->as_dereference());
+   ir_dereference_array *deref_array = ir->as_dereference_array();
+   if (deref_array) {
+      assert(!deref_array->array->type->is_vector());
+   }
+
+   /* Use the rvalue deref handler for the most part.  We'll ignore
+    * swizzles in it and write swizzles using writemask, though.
+    */
+   ir->accept(v);
+   return st_dst_reg(v->result);
+}
+
+/**
+ * Process the condition of a conditional assignment
+ *
+ * Examines the condition of a conditional assignment to generate the optimal
+ * first operand of a \c CMP instruction.  If the condition is a relational
+ * operator with 0 (e.g., \c ir_binop_less), the value being compared will be
+ * used as the source for the \c CMP instruction.  Otherwise the comparison
+ * is processed to a boolean result, and the boolean result is used as the
+ * operand to the CMP instruction.
+ */
+bool
+glsl_to_tgsi_visitor::process_move_condition(ir_rvalue *ir)
+{
+   ir_rvalue *src_ir = ir;
+   bool negate = true;
+   bool switch_order = false;
+
+   ir_expression *const expr = ir->as_expression();
+   if ((expr != NULL) && (expr->get_num_operands() == 2)) {
+      bool zero_on_left = false;
+
+      if (expr->operands[0]->is_zero()) {
+         src_ir = expr->operands[1];
+         zero_on_left = true;
+      } else if (expr->operands[1]->is_zero()) {
+         src_ir = expr->operands[0];
+         zero_on_left = false;
+      }
+
+      /*      a is -  0  +            -  0  +
+       * (a <  0)  T  F  F  ( a < 0)  T  F  F
+       * (0 <  a)  F  F  T  (-a < 0)  F  F  T
+       * (a <= 0)  T  T  F  (-a < 0)  F  F  T  (swap order of other operands)
+       * (0 <= a)  F  T  T  ( a < 0)  T  F  F  (swap order of other operands)
+       * (a >  0)  F  F  T  (-a < 0)  F  F  T
+       * (0 >  a)  T  F  F  ( a < 0)  T  F  F
+       * (a >= 0)  F  T  T  ( a < 0)  T  F  F  (swap order of other operands)
+       * (0 >= a)  T  T  F  (-a < 0)  F  F  T  (swap order of other operands)
+       *
+       * Note that exchanging the order of 0 and 'a' in the comparison simply
+       * means that the value of 'a' should be negated.
+       */
+      if (src_ir != ir) {
+         switch (expr->operation) {
+         case ir_binop_less:
+            switch_order = false;
+            negate = zero_on_left;
+            break;
+
+         case ir_binop_greater:
+            switch_order = false;
+            negate = !zero_on_left;
+            break;
+
+         case ir_binop_lequal:
+            switch_order = true;
+            negate = !zero_on_left;
+            break;
+
+         case ir_binop_gequal:
+            switch_order = true;
+            negate = zero_on_left;
+            break;
+
+         default:
+            /* This isn't the right kind of comparison afterall, so make sure
+             * the whole condition is visited.
+             */
+            src_ir = ir;
+            break;
+         }
+      }
+   }
+
+   src_ir->accept(this);
+
+   /* We use the TGSI_OPCODE_CMP (a < 0 ? b : c) for conditional moves, and the
+    * condition we produced is 0.0 or 1.0.  By flipping the sign, we can
+    * choose which value TGSI_OPCODE_CMP produces without an extra instruction
+    * computing the condition.
+    */
+   if (negate)
+      this->result.negate = ~this->result.negate;
+
+   return switch_order;
+}
+
+void
+glsl_to_tgsi_visitor::visit(ir_assignment *ir)
+{
+   st_dst_reg l;
+   st_src_reg r;
+   int i;
+
+   ir->rhs->accept(this);
+   r = this->result;
+
+   l = get_assignment_lhs(ir->lhs, this);
+
+   /* FINISHME: This should really set to the correct maximal writemask for each
+    * FINISHME: component written (in the loops below).  This case can only
+    * FINISHME: occur for matrices, arrays, and structures.
+    */
+   if (ir->write_mask == 0) {
+      assert(!ir->lhs->type->is_scalar() && !ir->lhs->type->is_vector());
+      l.writemask = WRITEMASK_XYZW;
+   } else if (ir->lhs->type->is_scalar() &&
+              ir->lhs->variable_referenced()->mode == ir_var_out) {
+      /* FINISHME: This hack makes writing to gl_FragDepth, which lives in the
+       * FINISHME: W component of fragment shader output zero, work correctly.
+       */
+      l.writemask = WRITEMASK_XYZW;
+   } else {
+      int swizzles[4];
+      int first_enabled_chan = 0;
+      int rhs_chan = 0;
+
+      l.writemask = ir->write_mask;
+
+      for (int i = 0; i < 4; i++) {
+         if (l.writemask & (1 << i)) {
+            first_enabled_chan = GET_SWZ(r.swizzle, i);
+            break;
+         }
+      }
+
+      /* Swizzle a small RHS vector into the channels being written.
+       *
+       * glsl ir treats write_mask as dictating how many channels are
+       * present on the RHS while TGSI treats write_mask as just
+       * showing which channels of the vec4 RHS get written.
+       */
+      for (int i = 0; i < 4; i++) {
+         if (l.writemask & (1 << i))
+            swizzles[i] = GET_SWZ(r.swizzle, rhs_chan++);
+         else
+            swizzles[i] = first_enabled_chan;
+      }
+      r.swizzle = MAKE_SWIZZLE4(swizzles[0], swizzles[1],
+                               swizzles[2], swizzles[3]);
+   }
+
+   assert(l.file != PROGRAM_UNDEFINED);
+   assert(r.file != PROGRAM_UNDEFINED);
+
+   if (ir->condition) {
+      const bool switch_order = this->process_move_condition(ir->condition);
+      st_src_reg condition = this->result;
+
+      for (i = 0; i < type_size(ir->lhs->type); i++) {
+         st_src_reg l_src = st_src_reg(l);
+         l_src.swizzle = swizzle_for_size(ir->lhs->type->vector_elements);
+         
+         if (switch_order) {
+            emit(ir, TGSI_OPCODE_CMP, l, condition, l_src, r);
+         } else {
+            emit(ir, TGSI_OPCODE_CMP, l, condition, r, l_src);
+         }
+
+         l.index++;
+         r.index++;
+      }
+   } else if (ir->rhs->as_expression() &&
+              this->instructions.get_tail() &&
+              ir->rhs == ((glsl_to_tgsi_instruction *)this->instructions.get_tail())->ir &&
+              type_size(ir->lhs->type) == 1) {
+      /* To avoid emitting an extra MOV when assigning an expression to a 
+       * variable, emit the last instruction of the expression again, but
+       * replace the destination register with the target of the assignment.
+       * Dead code elimination will remove the original instruction.
+       */
+      glsl_to_tgsi_instruction *inst;
+      inst = (glsl_to_tgsi_instruction *)this->instructions.get_tail();
+      emit(ir, inst->op, l, inst->src[0], inst->src[1], inst->src[2]);
+   } else {
+      for (i = 0; i < type_size(ir->lhs->type); i++) {
+         emit(ir, TGSI_OPCODE_MOV, l, r);
+         l.index++;
+         r.index++;
+      }
+   }
+}
+
+
+void
+glsl_to_tgsi_visitor::visit(ir_constant *ir)
+{
+   st_src_reg src;
+   GLfloat stack_vals[4] = { 0 };
+   gl_constant_value *values = (gl_constant_value *) stack_vals;
+   GLenum gl_type = GL_NONE;
+   unsigned int i;
+   static int in_array = 0;
+   gl_register_file file = in_array ? PROGRAM_CONSTANT : PROGRAM_IMMEDIATE;
+
+   /* Unfortunately, 4 floats is all we can get into
+    * _mesa_add_typed_unnamed_constant.  So, make a temp to store an
+    * aggregate constant and move each constant value into it.  If we
+    * get lucky, copy propagation will eliminate the extra moves.
+    */
+   if (ir->type->base_type == GLSL_TYPE_STRUCT) {
+      st_src_reg temp_base = get_temp(ir->type);
+      st_dst_reg temp = st_dst_reg(temp_base);
+
+      foreach_iter(exec_list_iterator, iter, ir->components) {
+         ir_constant *field_value = (ir_constant *)iter.get();
+         int size = type_size(field_value->type);
+
+         assert(size > 0);
+
+         field_value->accept(this);
+         src = this->result;
+
+         for (i = 0; i < (unsigned int)size; i++) {
+            emit(ir, TGSI_OPCODE_MOV, temp, src);
+
+            src.index++;
+            temp.index++;
+         }
+      }
+      this->result = temp_base;
+      return;
+   }
+
+   if (ir->type->is_array()) {
+      st_src_reg temp_base = get_temp(ir->type);
+      st_dst_reg temp = st_dst_reg(temp_base);
+      int size = type_size(ir->type->fields.array);
+
+      assert(size > 0);
+      in_array++;
+
+      for (i = 0; i < ir->type->length; i++) {
+         ir->array_elements[i]->accept(this);
+         src = this->result;
+         for (int j = 0; j < size; j++) {
+            emit(ir, TGSI_OPCODE_MOV, temp, src);
+
+            src.index++;
+            temp.index++;
+         }
+      }
+      this->result = temp_base;
+      in_array--;
+      return;
+   }
+
+   if (ir->type->is_matrix()) {
+      st_src_reg mat = get_temp(ir->type);
+      st_dst_reg mat_column = st_dst_reg(mat);
+
+      for (i = 0; i < ir->type->matrix_columns; i++) {
+         assert(ir->type->base_type == GLSL_TYPE_FLOAT);
+         values = (gl_constant_value *) &ir->value.f[i * ir->type->vector_elements];
+
+         src = st_src_reg(file, -1, ir->type->base_type);
+         src.index = add_constant(file,
+                                  values,
+                                  ir->type->vector_elements,
+                                  GL_FLOAT,
+                                  &src.swizzle);
+         emit(ir, TGSI_OPCODE_MOV, mat_column, src);
+
+         mat_column.index++;
+      }
+
+      this->result = mat;
+      return;
+   }
+
+   switch (ir->type->base_type) {
+   case GLSL_TYPE_FLOAT:
+      gl_type = GL_FLOAT;
+      for (i = 0; i < ir->type->vector_elements; i++) {
+         values[i].f = ir->value.f[i];
+      }
+      break;
+   case GLSL_TYPE_UINT:
+      gl_type = glsl_version >= 130 ? GL_UNSIGNED_INT : GL_FLOAT;
+      for (i = 0; i < ir->type->vector_elements; i++) {
+         if (glsl_version >= 130)
+            values[i].u = ir->value.u[i];
+         else
+            values[i].f = ir->value.u[i];
+      }
+      break;
+   case GLSL_TYPE_INT:
+      gl_type = glsl_version >= 130 ? GL_INT : GL_FLOAT;
+      for (i = 0; i < ir->type->vector_elements; i++) {
+         if (glsl_version >= 130)
+            values[i].i = ir->value.i[i];
+         else
+            values[i].f = ir->value.i[i];
+      }
+      break;
+   case GLSL_TYPE_BOOL:
+      gl_type = glsl_version >= 130 ? GL_BOOL : GL_FLOAT;
+      for (i = 0; i < ir->type->vector_elements; i++) {
+         if (glsl_version >= 130)
+            values[i].b = ir->value.b[i];
+         else
+            values[i].f = ir->value.b[i];
+      }
+      break;
+   default:
+      assert(!"Non-float/uint/int/bool constant");
+   }
+
+   this->result = st_src_reg(file, -1, ir->type);
+   this->result.index = add_constant(file,
+                                     values,
+                                     ir->type->vector_elements,
+                                     gl_type,
+                                     &this->result.swizzle);
+}
+
+function_entry *
+glsl_to_tgsi_visitor::get_function_signature(ir_function_signature *sig)
+{
+   function_entry *entry;
+
+   foreach_iter(exec_list_iterator, iter, this->function_signatures) {
+      entry = (function_entry *)iter.get();
+
+      if (entry->sig == sig)
+         return entry;
+   }
+
+   entry = ralloc(mem_ctx, function_entry);
+   entry->sig = sig;
+   entry->sig_id = this->next_signature_id++;
+   entry->bgn_inst = NULL;
+
+   /* Allocate storage for all the parameters. */
+   foreach_iter(exec_list_iterator, iter, sig->parameters) {
+      ir_variable *param = (ir_variable *)iter.get();
+      variable_storage *storage;
+
+      storage = find_variable_storage(param);
+      assert(!storage);
+
+      storage = new(mem_ctx) variable_storage(param, PROGRAM_TEMPORARY,
+                                             this->next_temp);
+      this->variables.push_tail(storage);
+
+      this->next_temp += type_size(param->type);
+   }
+
+   if (!sig->return_type->is_void()) {
+      entry->return_reg = get_temp(sig->return_type);
+   } else {
+      entry->return_reg = undef_src;
+   }
+
+   this->function_signatures.push_tail(entry);
+   return entry;
+}
+
+void
+glsl_to_tgsi_visitor::visit(ir_call *ir)
+{
+   glsl_to_tgsi_instruction *call_inst;
+   ir_function_signature *sig = ir->get_callee();
+   function_entry *entry = get_function_signature(sig);
+   int i;
+
+   /* Process in parameters. */
+   exec_list_iterator sig_iter = sig->parameters.iterator();
+   foreach_iter(exec_list_iterator, iter, *ir) {
+      ir_rvalue *param_rval = (ir_rvalue *)iter.get();
+      ir_variable *param = (ir_variable *)sig_iter.get();
+
+      if (param->mode == ir_var_in ||
+          param->mode == ir_var_inout) {
+         variable_storage *storage = find_variable_storage(param);
+         assert(storage);
+
+         param_rval->accept(this);
+         st_src_reg r = this->result;
+
+         st_dst_reg l;
+         l.file = storage->file;
+         l.index = storage->index;
+         l.reladdr = NULL;
+         l.writemask = WRITEMASK_XYZW;
+         l.cond_mask = COND_TR;
+
+         for (i = 0; i < type_size(param->type); i++) {
+            emit(ir, TGSI_OPCODE_MOV, l, r);
+            l.index++;
+            r.index++;
+         }
+      }
+
+      sig_iter.next();
+   }
+   assert(!sig_iter.has_next());
+
+   /* Emit call instruction */
+   call_inst = emit(ir, TGSI_OPCODE_CAL);
+   call_inst->function = entry;
+
+   /* Process out parameters. */
+   sig_iter = sig->parameters.iterator();
+   foreach_iter(exec_list_iterator, iter, *ir) {
+      ir_rvalue *param_rval = (ir_rvalue *)iter.get();
+      ir_variable *param = (ir_variable *)sig_iter.get();
+
+      if (param->mode == ir_var_out ||
+          param->mode == ir_var_inout) {
+         variable_storage *storage = find_variable_storage(param);
+         assert(storage);
+
+         st_src_reg r;
+         r.file = storage->file;
+         r.index = storage->index;
+         r.reladdr = NULL;
+         r.swizzle = SWIZZLE_NOOP;
+         r.negate = 0;
+
+         param_rval->accept(this);
+         st_dst_reg l = st_dst_reg(this->result);
+
+         for (i = 0; i < type_size(param->type); i++) {
+            emit(ir, TGSI_OPCODE_MOV, l, r);
+            l.index++;
+            r.index++;
+         }
+      }
+
+      sig_iter.next();
+   }
+   assert(!sig_iter.has_next());
+
+   /* Process return value. */
+   this->result = entry->return_reg;
+}
+
+void
+glsl_to_tgsi_visitor::visit(ir_texture *ir)
+{
+   st_src_reg result_src, coord, lod_info, projector, dx, dy;
+   st_dst_reg result_dst, coord_dst;
+   glsl_to_tgsi_instruction *inst = NULL;
+   unsigned opcode = TGSI_OPCODE_NOP;
+
+   ir->coordinate->accept(this);
+
+   /* Put our coords in a temp.  We'll need to modify them for shadow,
+    * projection, or LOD, so the only case we'd use it as is is if
+    * we're doing plain old texturing.  The optimization passes on
+    * glsl_to_tgsi_visitor should handle cleaning up our mess in that case.
+    */
+   coord = get_temp(glsl_type::vec4_type);
+   coord_dst = st_dst_reg(coord);
+   emit(ir, TGSI_OPCODE_MOV, coord_dst, this->result);
+
+   if (ir->projector) {
+      ir->projector->accept(this);
+      projector = this->result;
+   }
+
+   /* Storage for our result.  Ideally for an assignment we'd be using
+    * the actual storage for the result here, instead.
+    */
+   result_src = get_temp(glsl_type::vec4_type);
+   result_dst = st_dst_reg(result_src);
+
+   switch (ir->op) {
+   case ir_tex:
+      opcode = TGSI_OPCODE_TEX;
+      break;
+   case ir_txb:
+      opcode = TGSI_OPCODE_TXB;
+      ir->lod_info.bias->accept(this);
+      lod_info = this->result;
+      break;
+   case ir_txl:
+      opcode = TGSI_OPCODE_TXL;
+      ir->lod_info.lod->accept(this);
+      lod_info = this->result;
+      break;
+   case ir_txd:
+      opcode = TGSI_OPCODE_TXD;
+      ir->lod_info.grad.dPdx->accept(this);
+      dx = this->result;
+      ir->lod_info.grad.dPdy->accept(this);
+      dy = this->result;
+      break;
+   case ir_txf: /* TODO: use TGSI_OPCODE_TXF here */
+      assert(!"GLSL 1.30 features unsupported");
+      break;
+   }
+
+   if (ir->projector) {
+      if (opcode == TGSI_OPCODE_TEX) {
+         /* Slot the projector in as the last component of the coord. */
+         coord_dst.writemask = WRITEMASK_W;
+         emit(ir, TGSI_OPCODE_MOV, coord_dst, projector);
+         coord_dst.writemask = WRITEMASK_XYZW;
+         opcode = TGSI_OPCODE_TXP;
+      } else {
+         st_src_reg coord_w = coord;
+         coord_w.swizzle = SWIZZLE_WWWW;
+
+         /* For the other TEX opcodes there's no projective version
+          * since the last slot is taken up by LOD info.  Do the
+          * projective divide now.
+          */
+         coord_dst.writemask = WRITEMASK_W;
+         emit(ir, TGSI_OPCODE_RCP, coord_dst, projector);
+
+         /* In the case where we have to project the coordinates "by hand,"
+          * the shadow comparator value must also be projected.
+          */
+         st_src_reg tmp_src = coord;
+         if (ir->shadow_comparitor) {
+            /* Slot the shadow value in as the second to last component of the
+             * coord.
+             */
+            ir->shadow_comparitor->accept(this);
+
+            tmp_src = get_temp(glsl_type::vec4_type);
+            st_dst_reg tmp_dst = st_dst_reg(tmp_src);
+
+            tmp_dst.writemask = WRITEMASK_Z;
+            emit(ir, TGSI_OPCODE_MOV, tmp_dst, this->result);
+
+            tmp_dst.writemask = WRITEMASK_XY;
+            emit(ir, TGSI_OPCODE_MOV, tmp_dst, coord);
+         }
+
+         coord_dst.writemask = WRITEMASK_XYZ;
+         emit(ir, TGSI_OPCODE_MUL, coord_dst, tmp_src, coord_w);
+
+         coord_dst.writemask = WRITEMASK_XYZW;
+         coord.swizzle = SWIZZLE_XYZW;
+      }
+   }
+
+   /* If projection is done and the opcode is not TGSI_OPCODE_TXP, then the shadow
+    * comparator was put in the correct place (and projected) by the code,
+    * above, that handles by-hand projection.
+    */
+   if (ir->shadow_comparitor && (!ir->projector || opcode == TGSI_OPCODE_TXP)) {
+      /* Slot the shadow value in as the second to last component of the
+       * coord.
+       */
+      ir->shadow_comparitor->accept(this);
+      coord_dst.writemask = WRITEMASK_Z;
+      emit(ir, TGSI_OPCODE_MOV, coord_dst, this->result);
+      coord_dst.writemask = WRITEMASK_XYZW;
+   }
+
+   if (opcode == TGSI_OPCODE_TXL || opcode == TGSI_OPCODE_TXB) {
+      /* TGSI stores LOD or LOD bias in the last channel of the coords. */
+      coord_dst.writemask = WRITEMASK_W;
+      emit(ir, TGSI_OPCODE_MOV, coord_dst, lod_info);
+      coord_dst.writemask = WRITEMASK_XYZW;
+   }
+
+   if (opcode == TGSI_OPCODE_TXD)
+      inst = emit(ir, opcode, result_dst, coord, dx, dy);
+   else
+      inst = emit(ir, opcode, result_dst, coord);
+
+   if (ir->shadow_comparitor)
+      inst->tex_shadow = GL_TRUE;
+
+   inst->sampler = _mesa_get_sampler_uniform_value(ir->sampler,
+                                                  this->shader_program,
+                                                  this->prog);
+
+   const glsl_type *sampler_type = ir->sampler->type;
+
+   switch (sampler_type->sampler_dimensionality) {
+   case GLSL_SAMPLER_DIM_1D:
+      inst->tex_target = (sampler_type->sampler_array)
+         ? TEXTURE_1D_ARRAY_INDEX : TEXTURE_1D_INDEX;
+      break;
+   case GLSL_SAMPLER_DIM_2D:
+      inst->tex_target = (sampler_type->sampler_array)
+         ? TEXTURE_2D_ARRAY_INDEX : TEXTURE_2D_INDEX;
+      break;
+   case GLSL_SAMPLER_DIM_3D:
+      inst->tex_target = TEXTURE_3D_INDEX;
+      break;
+   case GLSL_SAMPLER_DIM_CUBE:
+      inst->tex_target = TEXTURE_CUBE_INDEX;
+      break;
+   case GLSL_SAMPLER_DIM_RECT:
+      inst->tex_target = TEXTURE_RECT_INDEX;
+      break;
+   case GLSL_SAMPLER_DIM_BUF:
+      assert(!"FINISHME: Implement ARB_texture_buffer_object");
+      break;
+   default:
+      assert(!"Should not get here.");
+   }
+
+   this->result = result_src;
+}
+
+void
+glsl_to_tgsi_visitor::visit(ir_return *ir)
+{
+   if (ir->get_value()) {
+      st_dst_reg l;
+      int i;
+
+      assert(current_function);
+
+      ir->get_value()->accept(this);
+      st_src_reg r = this->result;
+
+      l = st_dst_reg(current_function->return_reg);
+
+      for (i = 0; i < type_size(current_function->sig->return_type); i++) {
+         emit(ir, TGSI_OPCODE_MOV, l, r);
+         l.index++;
+         r.index++;
+      }
+   }
+
+   emit(ir, TGSI_OPCODE_RET);
+}
+
+void
+glsl_to_tgsi_visitor::visit(ir_discard *ir)
+{
+   struct gl_fragment_program *fp = (struct gl_fragment_program *)this->prog;
+
+   if (ir->condition) {
+      ir->condition->accept(this);
+      this->result.negate = ~this->result.negate;
+      emit(ir, TGSI_OPCODE_KIL, undef_dst, this->result);
+   } else {
+      emit(ir, TGSI_OPCODE_KILP);
+   }
+
+   fp->UsesKill = GL_TRUE;
+}
+
+void
+glsl_to_tgsi_visitor::visit(ir_if *ir)
+{
+   glsl_to_tgsi_instruction *cond_inst, *if_inst, *else_inst = NULL;
+   glsl_to_tgsi_instruction *prev_inst;
+
+   prev_inst = (glsl_to_tgsi_instruction *)this->instructions.get_tail();
+
+   ir->condition->accept(this);
+   assert(this->result.file != PROGRAM_UNDEFINED);
+
+   if (this->options->EmitCondCodes) {
+      cond_inst = (glsl_to_tgsi_instruction *)this->instructions.get_tail();
+
+      /* See if we actually generated any instruction for generating
+       * the condition.  If not, then cook up a move to a temp so we
+       * have something to set cond_update on.
+       */
+      if (cond_inst == prev_inst) {
+         st_src_reg temp = get_temp(glsl_type::bool_type);
+         cond_inst = emit(ir->condition, TGSI_OPCODE_MOV, st_dst_reg(temp), result);
+      }
+      cond_inst->cond_update = GL_TRUE;
+
+      if_inst = emit(ir->condition, TGSI_OPCODE_IF);
+      if_inst->dst.cond_mask = COND_NE;
+   } else {
+      if_inst = emit(ir->condition, TGSI_OPCODE_IF, undef_dst, this->result);
+   }
+
+   this->instructions.push_tail(if_inst);
+
+   visit_exec_list(&ir->then_instructions, this);
+
+   if (!ir->else_instructions.is_empty()) {
+      else_inst = emit(ir->condition, TGSI_OPCODE_ELSE);
+      visit_exec_list(&ir->else_instructions, this);
+   }
+
+   if_inst = emit(ir->condition, TGSI_OPCODE_ENDIF);
+}
+
+glsl_to_tgsi_visitor::glsl_to_tgsi_visitor()
+{
+   result.file = PROGRAM_UNDEFINED;
+   next_temp = 1;
+   next_signature_id = 1;
+   num_immediates = 0;
+   current_function = NULL;
+   num_address_regs = 0;
+   indirect_addr_temps = false;
+   indirect_addr_consts = false;
+   mem_ctx = ralloc_context(NULL);
+}
+
+glsl_to_tgsi_visitor::~glsl_to_tgsi_visitor()
+{
+   ralloc_free(mem_ctx);
+}
+
+extern "C" void free_glsl_to_tgsi_visitor(glsl_to_tgsi_visitor *v)
+{
+   delete v;
+}
+
+
+/**
+ * Count resources used by the given gpu program (number of texture
+ * samplers, etc).
+ */
+static void
+count_resources(glsl_to_tgsi_visitor *v, gl_program *prog)
+{
+   v->samplers_used = 0;
+
+   foreach_iter(exec_list_iterator, iter, v->instructions) {
+      glsl_to_tgsi_instruction *inst = (glsl_to_tgsi_instruction *)iter.get();
+
+      if (is_tex_instruction(inst->op)) {
+         v->samplers_used |= 1 << inst->sampler;
+
+         prog->SamplerTargets[inst->sampler] =
+            (gl_texture_index)inst->tex_target;
+         if (inst->tex_shadow) {
+            prog->ShadowSamplers |= 1 << inst->sampler;
+         }
+      }
+   }
+   
+   prog->SamplersUsed = v->samplers_used;
+   _mesa_update_shader_textures_used(prog);
+}
+
+
+/**
+ * Check if the given vertex/fragment/shader program is within the
+ * resource limits of the context (number of texture units, etc).
+ * If any of those checks fail, record a linker error.
+ *
+ * XXX more checks are needed...
+ */
+static void
+check_resources(const struct gl_context *ctx,
+                struct gl_shader_program *shader_program,
+                glsl_to_tgsi_visitor *prog,
+                struct gl_program *proginfo)
+{
+   switch (proginfo->Target) {
+   case GL_VERTEX_PROGRAM_ARB:
+      if (_mesa_bitcount(prog->samplers_used) >
+          ctx->Const.MaxVertexTextureImageUnits) {
+         fail_link(shader_program, "Too many vertex shader texture samplers");
+      }
+      if (proginfo->Parameters->NumParameters > MAX_UNIFORMS) {
+         fail_link(shader_program, "Too many vertex shader constants");
+      }
+      break;
+   case MESA_GEOMETRY_PROGRAM:
+      if (_mesa_bitcount(prog->samplers_used) >
+          ctx->Const.MaxGeometryTextureImageUnits) {
+         fail_link(shader_program, "Too many geometry shader texture samplers");
+      }
+      if (proginfo->Parameters->NumParameters >
+          MAX_GEOMETRY_UNIFORM_COMPONENTS / 4) {
+         fail_link(shader_program, "Too many geometry shader constants");
+      }
+      break;
+   case GL_FRAGMENT_PROGRAM_ARB:
+      if (_mesa_bitcount(prog->samplers_used) >
+          ctx->Const.MaxTextureImageUnits) {
+         fail_link(shader_program, "Too many fragment shader texture samplers");
+      }
+      if (proginfo->Parameters->NumParameters > MAX_UNIFORMS) {
+         fail_link(shader_program, "Too many fragment shader constants");
+      }
+      break;
+   default:
+      _mesa_problem(ctx, "unexpected program type in check_resources()");
+   }
+}
+
+
+
+struct uniform_sort {
+   struct gl_uniform *u;
+   int pos;
+};
+
+/* The shader_program->Uniforms list is almost sorted in increasing
+ * uniform->{Frag,Vert}Pos locations, but not quite when there are
+ * uniforms shared between targets.  We need to add parameters in
+ * increasing order for the targets.
+ */
+static int
+sort_uniforms(const void *a, const void *b)
+{
+   struct uniform_sort *u1 = (struct uniform_sort *)a;
+   struct uniform_sort *u2 = (struct uniform_sort *)b;
+
+   return u1->pos - u2->pos;
+}
+
+/* Add the uniforms to the parameters.  The linker chose locations
+ * in our parameters lists (which weren't created yet), which the
+ * uniforms code will use to poke values into our parameters list
+ * when uniforms are updated.
+ */
+static void
+add_uniforms_to_parameters_list(struct gl_shader_program *shader_program,
+                               struct gl_shader *shader,
+                               struct gl_program *prog)
+{
+   unsigned int i;
+   unsigned int next_sampler = 0, num_uniforms = 0;
+   struct uniform_sort *sorted_uniforms;
+
+   sorted_uniforms = ralloc_array(NULL, struct uniform_sort,
+                                 shader_program->Uniforms->NumUniforms);
+
+   for (i = 0; i < shader_program->Uniforms->NumUniforms; i++) {
+      struct gl_uniform *uniform = shader_program->Uniforms->Uniforms + i;
+      int parameter_index = -1;
+
+      switch (shader->Type) {
+      case GL_VERTEX_SHADER:
+         parameter_index = uniform->VertPos;
+         break;
+      case GL_FRAGMENT_SHADER:
+         parameter_index = uniform->FragPos;
+         break;
+      case GL_GEOMETRY_SHADER:
+         parameter_index = uniform->GeomPos;
+         break;
+      }
+
+      /* Only add uniforms used in our target. */
+      if (parameter_index != -1) {
+         sorted_uniforms[num_uniforms].pos = parameter_index;
+         sorted_uniforms[num_uniforms].u = uniform;
+         num_uniforms++;
+      }
+   }
+
+   qsort(sorted_uniforms, num_uniforms, sizeof(struct uniform_sort),
+         sort_uniforms);
+
+   for (i = 0; i < num_uniforms; i++) {
+      struct gl_uniform *uniform = sorted_uniforms[i].u;
+      int parameter_index = sorted_uniforms[i].pos;
+      const glsl_type *type = uniform->Type;
+      unsigned int size;
+
+      if (type->is_vector() ||
+          type->is_scalar()) {
+         size = type->vector_elements;
+      } else {
+         size = type_size(type) * 4;
+      }
+
+      gl_register_file file;
+      if (type->is_sampler() ||
+          (type->is_array() && type->fields.array->is_sampler())) {
+         file = PROGRAM_SAMPLER;
+      } else {
+         file = PROGRAM_UNIFORM;
+      }
+
+      GLint index = _mesa_lookup_parameter_index(prog->Parameters, -1,
+                                                uniform->Name);
+
+      if (index < 0) {
+         index = _mesa_add_parameter(prog->Parameters, file,
+                                    uniform->Name, size, type->gl_type,
+                                    NULL, NULL, 0x0);
+
+         /* Sampler uniform values are stored in prog->SamplerUnits,
+          * and the entry in that array is selected by this index we
+          * store in ParameterValues[].
+          */
+         if (file == PROGRAM_SAMPLER) {
+            for (unsigned int j = 0; j < size / 4; j++)
+               prog->Parameters->ParameterValues[index + j][0].f = next_sampler++;
+         }
+
+         /* The location chosen in the Parameters list here (returned
+          * from _mesa_add_uniform) has to match what the linker chose.
+          */
+         if (index != parameter_index) {
+            fail_link(shader_program, "Allocation of uniform `%s' to target "
+                     "failed (%d vs %d)\n",
+                     uniform->Name, index, parameter_index);
+         }
+      }
+   }
+
+   ralloc_free(sorted_uniforms);
+}
+
+static void
+set_uniform_initializer(struct gl_context *ctx, void *mem_ctx,
+                       struct gl_shader_program *shader_program,
+                       const char *name, const glsl_type *type,
+                       ir_constant *val)
+{
+   if (type->is_record()) {
+      ir_constant *field_constant;
+
+      field_constant = (ir_constant *)val->components.get_head();
+
+      for (unsigned int i = 0; i < type->length; i++) {
+         const glsl_type *field_type = type->fields.structure[i].type;
+         const char *field_name = ralloc_asprintf(mem_ctx, "%s.%s", name,
+                                           type->fields.structure[i].name);
+         set_uniform_initializer(ctx, mem_ctx, shader_program, field_name,
+                                field_type, field_constant);
+         field_constant = (ir_constant *)field_constant->next;
+      }
+      return;
+   }
+
+   int loc = _mesa_get_uniform_location(ctx, shader_program, name);
+
+   if (loc == -1) {
+      fail_link(shader_program,
+               "Couldn't find uniform for initializer %s\n", name);
+      return;
+   }
+
+   for (unsigned int i = 0; i < (type->is_array() ? type->length : 1); i++) {
+      ir_constant *element;
+      const glsl_type *element_type;
+      if (type->is_array()) {
+         element = val->array_elements[i];
+         element_type = type->fields.array;
+      } else {
+         element = val;
+         element_type = type;
+      }
+
+      void *values;
+
+      if (element_type->base_type == GLSL_TYPE_BOOL) {
+         int *conv = ralloc_array(mem_ctx, int, element_type->components());
+         for (unsigned int j = 0; j < element_type->components(); j++) {
+            conv[j] = element->value.b[j];
+         }
+         values = (void *)conv;
+         element_type = glsl_type::get_instance(GLSL_TYPE_INT,
+                                               element_type->vector_elements,
+                                               1);
+      } else {
+         values = &element->value;
+      }
+
+      if (element_type->is_matrix()) {
+         _mesa_uniform_matrix(ctx, shader_program,
+                             element_type->matrix_columns,
+                             element_type->vector_elements,
+                             loc, 1, GL_FALSE, (GLfloat *)values);
+         loc += element_type->matrix_columns;
+      } else {
+         _mesa_uniform(ctx, shader_program, loc, element_type->matrix_columns,
+                      values, element_type->gl_type);
+         loc += type_size(element_type);
+      }
+   }
+}
+
+static void
+set_uniform_initializers(struct gl_context *ctx,
+                        struct gl_shader_program *shader_program)
+{
+   void *mem_ctx = NULL;
+
+   for (unsigned int i = 0; i < MESA_SHADER_TYPES; i++) {
+      struct gl_shader *shader = shader_program->_LinkedShaders[i];
+
+      if (shader == NULL)
+         continue;
+
+      foreach_iter(exec_list_iterator, iter, *shader->ir) {
+         ir_instruction *ir = (ir_instruction *)iter.get();
+         ir_variable *var = ir->as_variable();
+
+         if (!var || var->mode != ir_var_uniform || !var->constant_value)
+            continue;
+
+         if (!mem_ctx)
+            mem_ctx = ralloc_context(NULL);
+
+         set_uniform_initializer(ctx, mem_ctx, shader_program, var->name,
+                                var->type, var->constant_value);
+      }
+   }
+
+   ralloc_free(mem_ctx);
+}
+
+/*
+ * Scan/rewrite program to remove reads of custom (output) registers.
+ * The passed type has to be either PROGRAM_OUTPUT or PROGRAM_VARYING
+ * (for vertex shaders).
+ * In GLSL shaders, varying vars can be read and written.
+ * On some hardware, trying to read an output register causes trouble.
+ * So, rewrite the program to use a temporary register in this case.
+ * 
+ * Based on _mesa_remove_output_reads from programopt.c.
+ */
+void
+glsl_to_tgsi_visitor::remove_output_reads(gl_register_file type)
+{
+   GLuint i;
+   GLint outputMap[VERT_RESULT_MAX];
+   GLint outputTypes[VERT_RESULT_MAX];
+   GLuint numVaryingReads = 0;
+   GLboolean usedTemps[MAX_TEMPS];
+   GLuint firstTemp = 0;
+
+   _mesa_find_used_registers(prog, PROGRAM_TEMPORARY,
+                             usedTemps, MAX_TEMPS);
+
+   assert(type == PROGRAM_VARYING || type == PROGRAM_OUTPUT);
+   assert(prog->Target == GL_VERTEX_PROGRAM_ARB || type != PROGRAM_VARYING);
+
+   for (i = 0; i < VERT_RESULT_MAX; i++)
+      outputMap[i] = -1;
+
+   /* look for instructions which read from varying vars */
+   foreach_iter(exec_list_iterator, iter, this->instructions) {
+      glsl_to_tgsi_instruction *inst = (glsl_to_tgsi_instruction *)iter.get();
+      const GLuint numSrc = num_inst_src_regs(inst->op);
+      GLuint j;
+      for (j = 0; j < numSrc; j++) {
+         if (inst->src[j].file == type) {
+            /* replace the read with a temp reg */
+            const GLuint var = inst->src[j].index;
+            if (outputMap[var] == -1) {
+               numVaryingReads++;
+               outputMap[var] = _mesa_find_free_register(usedTemps,
+                                                         MAX_TEMPS,
+                                                         firstTemp);
+               outputTypes[var] = inst->src[j].type;
+               firstTemp = outputMap[var] + 1;
+            }
+            inst->src[j].file = PROGRAM_TEMPORARY;
+            inst->src[j].index = outputMap[var];
+         }
+      }
+   }
+
+   if (numVaryingReads == 0)
+      return; /* nothing to be done */
+
+   /* look for instructions which write to the varying vars identified above */
+   foreach_iter(exec_list_iterator, iter, this->instructions) {
+      glsl_to_tgsi_instruction *inst = (glsl_to_tgsi_instruction *)iter.get();
+      if (inst->dst.file == type && outputMap[inst->dst.index] >= 0) {
+         /* change inst to write to the temp reg, instead of the varying */
+         inst->dst.file = PROGRAM_TEMPORARY;
+         inst->dst.index = outputMap[inst->dst.index];
+      }
+   }
+   
+   /* insert new MOV instructions at the end */
+   for (i = 0; i < VERT_RESULT_MAX; i++) {
+      if (outputMap[i] >= 0) {
+         /* MOV VAR[i], TEMP[tmp]; */
+         st_src_reg src = st_src_reg(PROGRAM_TEMPORARY, outputMap[i], outputTypes[i]);
+         st_dst_reg dst = st_dst_reg(type, WRITEMASK_XYZW, outputTypes[i]);
+         dst.index = i;
+         this->emit(NULL, TGSI_OPCODE_MOV, dst, src);
+      }
+   }
+}
+
+/**
+ * Returns the mask of channels (bitmask of WRITEMASK_X,Y,Z,W) which
+ * are read from the given src in this instruction
+ */
+static int
+get_src_arg_mask(st_dst_reg dst, st_src_reg src)
+{
+   int read_mask = 0, comp;
+
+   /* Now, given the src swizzle and the written channels, find which
+    * components are actually read
+    */
+   for (comp = 0; comp < 4; ++comp) {
+      const unsigned coord = GET_SWZ(src.swizzle, comp);
+      ASSERT(coord < 4);
+      if (dst.writemask & (1 << comp) && coord <= SWIZZLE_W)
+         read_mask |= 1 << coord;
+   }
+
+   return read_mask;
+}
+
+/**
+ * This pass replaces CMP T0, T1 T2 T0 with MOV T0, T2 when the CMP
+ * instruction is the first instruction to write to register T0.  There are
+ * several lowering passes done in GLSL IR (e.g. branches and
+ * relative addressing) that create a large number of conditional assignments
+ * that ir_to_mesa converts to CMP instructions like the one mentioned above.
+ *
+ * Here is why this conversion is safe:
+ * CMP T0, T1 T2 T0 can be expanded to:
+ * if (T1 < 0.0)
+ *     MOV T0, T2;
+ * else
+ *     MOV T0, T0;
+ *
+ * If (T1 < 0.0) evaluates to true then our replacement MOV T0, T2 is the same
+ * as the original program.  If (T1 < 0.0) evaluates to false, executing
+ * MOV T0, T0 will store a garbage value in T0 since T0 is uninitialized.
+ * Therefore, it doesn't matter that we are replacing MOV T0, T0 with MOV T0, T2
+ * because any instruction that was going to read from T0 after this was going
+ * to read a garbage value anyway.
+ */
+void
+glsl_to_tgsi_visitor::simplify_cmp(void)
+{
+   unsigned tempWrites[MAX_TEMPS];
+   unsigned outputWrites[MAX_PROGRAM_OUTPUTS];
+
+   memset(tempWrites, 0, sizeof(tempWrites));
+   memset(outputWrites, 0, sizeof(outputWrites));
+
+   foreach_iter(exec_list_iterator, iter, this->instructions) {
+      glsl_to_tgsi_instruction *inst = (glsl_to_tgsi_instruction *)iter.get();
+      unsigned prevWriteMask = 0;
+
+      /* Give up if we encounter relative addressing or flow control. */
+      if (inst->dst.reladdr ||
+          tgsi_get_opcode_info(inst->op)->is_branch ||
+          inst->op == TGSI_OPCODE_BGNSUB ||
+          inst->op == TGSI_OPCODE_CONT ||
+          inst->op == TGSI_OPCODE_END ||
+          inst->op == TGSI_OPCODE_ENDSUB ||
+          inst->op == TGSI_OPCODE_RET) {
+         return;
+      }
+
+      if (inst->dst.file == PROGRAM_OUTPUT) {
+         assert(inst->dst.index < MAX_PROGRAM_OUTPUTS);
+         prevWriteMask = outputWrites[inst->dst.index];
+         outputWrites[inst->dst.index] |= inst->dst.writemask;
+      } else if (inst->dst.file == PROGRAM_TEMPORARY) {
+         assert(inst->dst.index < MAX_TEMPS);
+         prevWriteMask = tempWrites[inst->dst.index];
+         tempWrites[inst->dst.index] |= inst->dst.writemask;
+      }
+
+      /* For a CMP to be considered a conditional write, the destination
+       * register and source register two must be the same. */
+      if (inst->op == TGSI_OPCODE_CMP
+          && !(inst->dst.writemask & prevWriteMask)
+          && inst->src[2].file == inst->dst.file
+          && inst->src[2].index == inst->dst.index
+          && inst->dst.writemask == get_src_arg_mask(inst->dst, inst->src[2])) {
+
+         inst->op = TGSI_OPCODE_MOV;
+         inst->src[0] = inst->src[1];
+      }
+   }
+}
+
+/* Replaces all references to a temporary register index with another index. */
+void
+glsl_to_tgsi_visitor::rename_temp_register(int index, int new_index)
+{
+   foreach_iter(exec_list_iterator, iter, this->instructions) {
+      glsl_to_tgsi_instruction *inst = (glsl_to_tgsi_instruction *)iter.get();
+      unsigned j;
+      
+      for (j=0; j < num_inst_src_regs(inst->op); j++) {
+         if (inst->src[j].file == PROGRAM_TEMPORARY && 
+             inst->src[j].index == index) {
+            inst->src[j].index = new_index;
+         }
+      }
+      
+      if (inst->dst.file == PROGRAM_TEMPORARY && inst->dst.index == index) {
+         inst->dst.index = new_index;
+      }
+   }
+}
+
+int
+glsl_to_tgsi_visitor::get_first_temp_read(int index)
+{
+   int depth = 0; /* loop depth */
+   int loop_start = -1; /* index of the first active BGNLOOP (if any) */
+   unsigned i = 0, j;
+   
+   foreach_iter(exec_list_iterator, iter, this->instructions) {
+      glsl_to_tgsi_instruction *inst = (glsl_to_tgsi_instruction *)iter.get();
+      
+      for (j=0; j < num_inst_src_regs(inst->op); j++) {
+         if (inst->src[j].file == PROGRAM_TEMPORARY && 
+             inst->src[j].index == index) {
+            return (depth == 0) ? i : loop_start;
+         }
+      }
+      
+      if (inst->op == TGSI_OPCODE_BGNLOOP) {
+         if(depth++ == 0)
+            loop_start = i;
+      } else if (inst->op == TGSI_OPCODE_ENDLOOP) {
+         if (--depth == 0)
+            loop_start = -1;
+      }
+      assert(depth >= 0);
+      
+      i++;
+   }
+   
+   return -1;
+}
+
+int
+glsl_to_tgsi_visitor::get_first_temp_write(int index)
+{
+   int depth = 0; /* loop depth */
+   int loop_start = -1; /* index of the first active BGNLOOP (if any) */
+   int i = 0;
+   
+   foreach_iter(exec_list_iterator, iter, this->instructions) {
+      glsl_to_tgsi_instruction *inst = (glsl_to_tgsi_instruction *)iter.get();
+      
+      if (inst->dst.file == PROGRAM_TEMPORARY && inst->dst.index == index) {
+         return (depth == 0) ? i : loop_start;
+      }
+      
+      if (inst->op == TGSI_OPCODE_BGNLOOP) {
+         if(depth++ == 0)
+            loop_start = i;
+      } else if (inst->op == TGSI_OPCODE_ENDLOOP) {
+         if (--depth == 0)
+            loop_start = -1;
+      }
+      assert(depth >= 0);
+      
+      i++;
+   }
+   
+   return -1;
+}
+
+int
+glsl_to_tgsi_visitor::get_last_temp_read(int index)
+{
+   int depth = 0; /* loop depth */
+   int last = -1; /* index of last instruction that reads the temporary */
+   unsigned i = 0, j;
+   
+   foreach_iter(exec_list_iterator, iter, this->instructions) {
+      glsl_to_tgsi_instruction *inst = (glsl_to_tgsi_instruction *)iter.get();
+      
+      for (j=0; j < num_inst_src_regs(inst->op); j++) {
+         if (inst->src[j].file == PROGRAM_TEMPORARY && 
+             inst->src[j].index == index) {
+            last = (depth == 0) ? i : -2;
+         }
+      }
+      
+      if (inst->op == TGSI_OPCODE_BGNLOOP)
+         depth++;
+      else if (inst->op == TGSI_OPCODE_ENDLOOP)
+         if (--depth == 0 && last == -2)
+            last = i;
+      assert(depth >= 0);
+      
+      i++;
+   }
+   
+   assert(last >= -1);
+   return last;
+}
+
+int
+glsl_to_tgsi_visitor::get_last_temp_write(int index)
+{
+   int depth = 0; /* loop depth */
+   int last = -1; /* index of last instruction that writes to the temporary */
+   int i = 0;
+   
+   foreach_iter(exec_list_iterator, iter, this->instructions) {
+      glsl_to_tgsi_instruction *inst = (glsl_to_tgsi_instruction *)iter.get();
+      
+      if (inst->dst.file == PROGRAM_TEMPORARY && inst->dst.index == index)
+         last = (depth == 0) ? i : -2;
+      
+      if (inst->op == TGSI_OPCODE_BGNLOOP)
+         depth++;
+      else if (inst->op == TGSI_OPCODE_ENDLOOP)
+         if (--depth == 0 && last == -2)
+            last = i;
+      assert(depth >= 0);
+      
+      i++;
+   }
+   
+   assert(last >= -1);
+   return last;
+}
+
+/*
+ * On a basic block basis, tracks available PROGRAM_TEMPORARY register
+ * channels for copy propagation and updates following instructions to
+ * use the original versions.
+ *
+ * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
+ * will occur.  As an example, a TXP production before this pass:
+ *
+ * 0: MOV TEMP[1], INPUT[4].xyyy;
+ * 1: MOV TEMP[1].w, INPUT[4].wwww;
+ * 2: TXP TEMP[2], TEMP[1], texture[0], 2D;
+ *
+ * and after:
+ *
+ * 0: MOV TEMP[1], INPUT[4].xyyy;
+ * 1: MOV TEMP[1].w, INPUT[4].wwww;
+ * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
+ *
+ * which allows for dead code elimination on TEMP[1]'s writes.
+ */
+void
+glsl_to_tgsi_visitor::copy_propagate(void)
+{
+   glsl_to_tgsi_instruction **acp = rzalloc_array(mem_ctx,
+                                                   glsl_to_tgsi_instruction *,
+                                                   this->next_temp * 4);
+   int *acp_level = rzalloc_array(mem_ctx, int, this->next_temp * 4);
+   int level = 0;
+
+   foreach_iter(exec_list_iterator, iter, this->instructions) {
+      glsl_to_tgsi_instruction *inst = (glsl_to_tgsi_instruction *)iter.get();
+
+      assert(inst->dst.file != PROGRAM_TEMPORARY
+             || inst->dst.index < this->next_temp);
+
+      /* First, do any copy propagation possible into the src regs. */
+      for (int r = 0; r < 3; r++) {
+         glsl_to_tgsi_instruction *first = NULL;
+         bool good = true;
+         int acp_base = inst->src[r].index * 4;
+
+         if (inst->src[r].file != PROGRAM_TEMPORARY ||
+             inst->src[r].reladdr)
+            continue;
+
+         /* See if we can find entries in the ACP consisting of MOVs
+          * from the same src register for all the swizzled channels
+          * of this src register reference.
+          */
+         for (int i = 0; i < 4; i++) {
+            int src_chan = GET_SWZ(inst->src[r].swizzle, i);
+            glsl_to_tgsi_instruction *copy_chan = acp[acp_base + src_chan];
+
+            if (!copy_chan) {
+               good = false;
+               break;
+            }
+
+            assert(acp_level[acp_base + src_chan] <= level);
+
+            if (!first) {
+               first = copy_chan;
+            } else {
+               if (first->src[0].file != copy_chan->src[0].file ||
+                  first->src[0].index != copy_chan->src[0].index) {
+                 good = false;
+                 break;
+               }
+            }
+         }
+
+         if (good) {
+            /* We've now validated that we can copy-propagate to
+             * replace this src register reference.  Do it.
+             */
+            inst->src[r].file = first->src[0].file;
+            inst->src[r].index = first->src[0].index;
+
+            int swizzle = 0;
+            for (int i = 0; i < 4; i++) {
+               int src_chan = GET_SWZ(inst->src[r].swizzle, i);
+               glsl_to_tgsi_instruction *copy_inst = acp[acp_base + src_chan];
+               swizzle |= (GET_SWZ(copy_inst->src[0].swizzle, src_chan) <<
+                          (3 * i));
+            }
+            inst->src[r].swizzle = swizzle;
+         }
+      }
+
+      switch (inst->op) {
+      case TGSI_OPCODE_BGNLOOP:
+      case TGSI_OPCODE_ENDLOOP:
+         /* End of a basic block, clear the ACP entirely. */
+         memset(acp, 0, sizeof(*acp) * this->next_temp * 4);
+         break;
+
+      case TGSI_OPCODE_IF:
+         ++level;
+         break;
+
+      case TGSI_OPCODE_ENDIF:
+      case TGSI_OPCODE_ELSE:
+         /* Clear all channels written inside the block from the ACP, but
+          * leaving those that were not touched.
+          */
+         for (int r = 0; r < this->next_temp; r++) {
+            for (int c = 0; c < 4; c++) {
+               if (!acp[4 * r + c])
+                 continue;
+
+               if (acp_level[4 * r + c] >= level)
+                 acp[4 * r + c] = NULL;
+            }
+         }
+         if (inst->op == TGSI_OPCODE_ENDIF)
+            --level;
+         break;
+
+      default:
+         /* Continuing the block, clear any written channels from
+          * the ACP.
+          */
+         if (inst->dst.file == PROGRAM_TEMPORARY && inst->dst.reladdr) {
+            /* Any temporary might be written, so no copy propagation
+             * across this instruction.
+             */
+            memset(acp, 0, sizeof(*acp) * this->next_temp * 4);
+         } else if (inst->dst.file == PROGRAM_OUTPUT &&
+                   inst->dst.reladdr) {
+            /* Any output might be written, so no copy propagation
+             * from outputs across this instruction.
+             */
+            for (int r = 0; r < this->next_temp; r++) {
+               for (int c = 0; c < 4; c++) {
+                 if (!acp[4 * r + c])
+                    continue;
+
+                 if (acp[4 * r + c]->src[0].file == PROGRAM_OUTPUT)
+                    acp[4 * r + c] = NULL;
+               }
+            }
+         } else if (inst->dst.file == PROGRAM_TEMPORARY ||
+                   inst->dst.file == PROGRAM_OUTPUT) {
+            /* Clear where it's used as dst. */
+            if (inst->dst.file == PROGRAM_TEMPORARY) {
+               for (int c = 0; c < 4; c++) {
+                 if (inst->dst.writemask & (1 << c)) {
+                    acp[4 * inst->dst.index + c] = NULL;
+                 }
+               }
+            }
+
+            /* Clear where it's used as src. */
+            for (int r = 0; r < this->next_temp; r++) {
+               for (int c = 0; c < 4; c++) {
+                 if (!acp[4 * r + c])
+                    continue;
+
+                 int src_chan = GET_SWZ(acp[4 * r + c]->src[0].swizzle, c);
+
+                 if (acp[4 * r + c]->src[0].file == inst->dst.file &&
+                     acp[4 * r + c]->src[0].index == inst->dst.index &&
+                     inst->dst.writemask & (1 << src_chan))
+                 {
+                    acp[4 * r + c] = NULL;
+                 }
+               }
+            }
+         }
+         break;
+      }
+
+      /* If this is a copy, add it to the ACP. */
+      if (inst->op == TGSI_OPCODE_MOV &&
+          inst->dst.file == PROGRAM_TEMPORARY &&
+          !inst->dst.reladdr &&
+          !inst->saturate &&
+          !inst->src[0].reladdr &&
+          !inst->src[0].negate) {
+         for (int i = 0; i < 4; i++) {
+            if (inst->dst.writemask & (1 << i)) {
+               acp[4 * inst->dst.index + i] = inst;
+               acp_level[4 * inst->dst.index + i] = level;
+            }
+         }
+      }
+   }
+
+   ralloc_free(acp_level);
+   ralloc_free(acp);
+}
+
+/*
+ * Tracks available PROGRAM_TEMPORARY registers for dead code elimination.
+ *
+ * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
+ * will occur.  As an example, a TXP production after copy propagation but 
+ * before this pass:
+ *
+ * 0: MOV TEMP[1], INPUT[4].xyyy;
+ * 1: MOV TEMP[1].w, INPUT[4].wwww;
+ * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
+ *
+ * and after this pass:
+ *
+ * 0: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
+ * 
+ * FIXME: assumes that all functions are inlined (no support for BGNSUB/ENDSUB)
+ * FIXME: doesn't eliminate all dead code inside of loops; it steps around them
+ */
+void
+glsl_to_tgsi_visitor::eliminate_dead_code(void)
+{
+   int i;
+   
+   for (i=0; i < this->next_temp; i++) {
+      int last_read = get_last_temp_read(i);
+      int j = 0;
+      
+      foreach_iter(exec_list_iterator, iter, this->instructions) {
+         glsl_to_tgsi_instruction *inst = (glsl_to_tgsi_instruction *)iter.get();
+
+         if (inst->dst.file == PROGRAM_TEMPORARY && inst->dst.index == i &&
+             j > last_read)
+         {
+            iter.remove();
+            delete inst;
+         }
+         
+         j++;
+      }
+   }
+}
+
+/*
+ * On a basic block basis, tracks available PROGRAM_TEMPORARY registers for dead
+ * code elimination.  This is less primitive than eliminate_dead_code(), as it
+ * is per-channel and can detect consecutive writes without a read between them
+ * as dead code.  However, there is some dead code that can be eliminated by 
+ * eliminate_dead_code() but not this function - for example, this function 
+ * cannot eliminate an instruction writing to a register that is never read and
+ * is the only instruction writing to that register.
+ *
+ * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
+ * will occur.
+ */
+int
+glsl_to_tgsi_visitor::eliminate_dead_code_advanced(void)
+{
+   glsl_to_tgsi_instruction **writes = rzalloc_array(mem_ctx,
+                                                     glsl_to_tgsi_instruction *,
+                                                     this->next_temp * 4);
+   int *write_level = rzalloc_array(mem_ctx, int, this->next_temp * 4);
+   int level = 0;
+   int removed = 0;
+
+   foreach_iter(exec_list_iterator, iter, this->instructions) {
+      glsl_to_tgsi_instruction *inst = (glsl_to_tgsi_instruction *)iter.get();
+
+      assert(inst->dst.file != PROGRAM_TEMPORARY
+             || inst->dst.index < this->next_temp);
+      
+      switch (inst->op) {
+      case TGSI_OPCODE_BGNLOOP:
+      case TGSI_OPCODE_ENDLOOP:
+         /* End of a basic block, clear the write array entirely.
+          * FIXME: This keeps us from killing dead code when the writes are
+          * on either side of a loop, even when the register isn't touched
+          * inside the loop.
+          */
+         memset(writes, 0, sizeof(*writes) * this->next_temp * 4);
+         break;
+
+      case TGSI_OPCODE_ENDIF:
+         --level;
+         break;
+
+      case TGSI_OPCODE_ELSE:
+         /* Clear all channels written inside the preceding if block from the
+          * write array, but leave those that were not touched.
+          *
+          * FIXME: This destroys opportunities to remove dead code inside of
+          * IF blocks that are followed by an ELSE block.
+          */
+         for (int r = 0; r < this->next_temp; r++) {
+            for (int c = 0; c < 4; c++) {
+               if (!writes[4 * r + c])
+                        continue;
+
+               if (write_level[4 * r + c] >= level)
+                        writes[4 * r + c] = NULL;
+            }
+         }
+         break;
+
+      case TGSI_OPCODE_IF:
+         ++level;
+         /* fallthrough to default case to mark the condition as read */
+      
+      default:
+         /* Continuing the block, clear any channels from the write array that
+          * are read by this instruction.
+          */
+         for (int i = 0; i < 4; i++) {
+            if (inst->src[i].file == PROGRAM_TEMPORARY && inst->src[i].reladdr){
+               /* Any temporary might be read, so no dead code elimination 
+                * across this instruction.
+                */
+               memset(writes, 0, sizeof(*writes) * this->next_temp * 4);
+            } else if (inst->src[i].file == PROGRAM_TEMPORARY) {
+               /* Clear where it's used as src. */
+               int src_chans = 1 << GET_SWZ(inst->src[i].swizzle, 0);
+               src_chans |= 1 << GET_SWZ(inst->src[i].swizzle, 1);
+               src_chans |= 1 << GET_SWZ(inst->src[i].swizzle, 2);
+               src_chans |= 1 << GET_SWZ(inst->src[i].swizzle, 3);
+               
+               for (int c = 0; c < 4; c++) {
+                  if (src_chans & (1 << c)) {
+                     writes[4 * inst->src[i].index + c] = NULL;
+                  }
+               }
+            }
+         }
+         break;
+      }
+
+      /* If this instruction writes to a temporary, add it to the write array.
+       * If there is already an instruction in the write array for one or more
+       * of the channels, flag that channel write as dead.
+       */
+      if (inst->dst.file == PROGRAM_TEMPORARY &&
+          !inst->dst.reladdr &&
+          !inst->saturate) {
+         for (int c = 0; c < 4; c++) {
+            if (inst->dst.writemask & (1 << c)) {
+               if (writes[4 * inst->dst.index + c]) {
+                  if (write_level[4 * inst->dst.index + c] < level)
+                     continue;
+                  else
+                     writes[4 * inst->dst.index + c]->dead_mask |= (1 << c);
+               }
+               writes[4 * inst->dst.index + c] = inst;
+               write_level[4 * inst->dst.index + c] = level;
+            }
+         }
+      }
+   }
+
+   /* Anything still in the write array at this point is dead code. */
+   for (int r = 0; r < this->next_temp; r++) {
+      for (int c = 0; c < 4; c++) {
+         glsl_to_tgsi_instruction *inst = writes[4 * r + c];
+         if (inst)
+            inst->dead_mask |= (1 << c);
+      }
+   }
+
+   /* Now actually remove the instructions that are completely dead and update
+    * the writemask of other instructions with dead channels.
+    */
+   foreach_iter(exec_list_iterator, iter, this->instructions) {
+      glsl_to_tgsi_instruction *inst = (glsl_to_tgsi_instruction *)iter.get();
+      
+      if (!inst->dead_mask || !inst->dst.writemask)
+         continue;
+      else if (inst->dead_mask == inst->dst.writemask) {
+         iter.remove();
+         delete inst;
+         removed++;
+      } else
+         inst->dst.writemask &= ~(inst->dead_mask);
+   }
+
+   ralloc_free(write_level);
+   ralloc_free(writes);
+   
+   return removed;
+}
+
+/* Merges temporary registers together where possible to reduce the number of 
+ * registers needed to run a program.
+ * 
+ * Produces optimal code only after copy propagation and dead code elimination 
+ * have been run. */
+void
+glsl_to_tgsi_visitor::merge_registers(void)
+{
+   int *last_reads = rzalloc_array(mem_ctx, int, this->next_temp);
+   int *first_writes = rzalloc_array(mem_ctx, int, this->next_temp);
+   int i, j;
+   
+   /* Read the indices of the last read and first write to each temp register
+    * into an array so that we don't have to traverse the instruction list as 
+    * much. */
+   for (i=0; i < this->next_temp; i++) {
+      last_reads[i] = get_last_temp_read(i);
+      first_writes[i] = get_first_temp_write(i);
+   }
+   
+   /* Start looking for registers with non-overlapping usages that can be 
+    * merged together. */
+   for (i=0; i < this->next_temp; i++) {
+      /* Don't touch unused registers. */
+      if (last_reads[i] < 0 || first_writes[i] < 0) continue;
+      
+      for (j=0; j < this->next_temp; j++) {
+         /* Don't touch unused registers. */
+         if (last_reads[j] < 0 || first_writes[j] < 0) continue;
+         
+         /* We can merge the two registers if the first write to j is after or 
+          * in the same instruction as the last read from i.  Note that the 
+          * register at index i will always be used earlier or at the same time 
+          * as the register at index j. */
+         if (first_writes[i] <= first_writes[j] && 
+             last_reads[i] <= first_writes[j])
+         {
+            rename_temp_register(j, i); /* Replace all references to j with i.*/
+            
+            /* Update the first_writes and last_reads arrays with the new 
+             * values for the merged register index, and mark the newly unused 
+             * register index as such. */
+            last_reads[i] = last_reads[j];
+            first_writes[j] = -1;
+            last_reads[j] = -1;
+         }
+      }
+   }
+   
+   ralloc_free(last_reads);
+   ralloc_free(first_writes);
+}
+
+/* Reassign indices to temporary registers by reusing unused indices created 
+ * by optimization passes. */
+void
+glsl_to_tgsi_visitor::renumber_registers(void)
+{
+   int i = 0;
+   int new_index = 0;
+   
+   for (i=0; i < this->next_temp; i++) {
+      if (get_first_temp_read(i) < 0) continue;
+      if (i != new_index)
+         rename_temp_register(i, new_index);
+      new_index++;
+   }
+   
+   this->next_temp = new_index;
+}
+
+/**
+ * Returns a fragment program which implements the current pixel transfer ops.
+ * Based on get_pixel_transfer_program in st_atom_pixeltransfer.c.
+ */
+extern "C" void
+get_pixel_transfer_visitor(struct st_fragment_program *fp,
+                           glsl_to_tgsi_visitor *original,
+                           int scale_and_bias, int pixel_maps)
+{
+   glsl_to_tgsi_visitor *v = new glsl_to_tgsi_visitor();
+   struct st_context *st = st_context(original->ctx);
+   struct gl_program *prog = &fp->Base.Base;
+   struct gl_program_parameter_list *params = _mesa_new_parameter_list();
+   st_src_reg coord, src0;
+   st_dst_reg dst0;
+   glsl_to_tgsi_instruction *inst;
+
+   /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
+   v->ctx = original->ctx;
+   v->prog = prog;
+   v->glsl_version = original->glsl_version;
+   v->options = original->options;
+   v->next_temp = original->next_temp;
+   v->num_address_regs = original->num_address_regs;
+   v->samplers_used = prog->SamplersUsed = original->samplers_used;
+   v->indirect_addr_temps = original->indirect_addr_temps;
+   v->indirect_addr_consts = original->indirect_addr_consts;
+   memcpy(&v->immediates, &original->immediates, sizeof(v->immediates));
+
+   /*
+    * Get initial pixel color from the texture.
+    * TEX colorTemp, fragment.texcoord[0], texture[0], 2D;
+    */
+   coord = st_src_reg(PROGRAM_INPUT, FRAG_ATTRIB_TEX0, glsl_type::vec2_type);
+   src0 = v->get_temp(glsl_type::vec4_type);
+   dst0 = st_dst_reg(src0);
+   inst = v->emit(NULL, TGSI_OPCODE_TEX, dst0, coord);
+   inst->sampler = 0;
+   inst->tex_target = TEXTURE_2D_INDEX;
+
+   prog->InputsRead |= (1 << FRAG_ATTRIB_TEX0);
+   prog->SamplersUsed |= (1 << 0); /* mark sampler 0 as used */
+   v->samplers_used |= (1 << 0);
+
+   if (scale_and_bias) {
+      static const gl_state_index scale_state[STATE_LENGTH] =
+         { STATE_INTERNAL, STATE_PT_SCALE,
+           (gl_state_index) 0, (gl_state_index) 0, (gl_state_index) 0 };
+      static const gl_state_index bias_state[STATE_LENGTH] =
+         { STATE_INTERNAL, STATE_PT_BIAS,
+           (gl_state_index) 0, (gl_state_index) 0, (gl_state_index) 0 };
+      GLint scale_p, bias_p;
+      st_src_reg scale, bias;
+
+      scale_p = _mesa_add_state_reference(params, scale_state);
+      bias_p = _mesa_add_state_reference(params, bias_state);
+
+      /* MAD colorTemp, colorTemp, scale, bias; */
+      scale = st_src_reg(PROGRAM_STATE_VAR, scale_p, GLSL_TYPE_FLOAT);
+      bias = st_src_reg(PROGRAM_STATE_VAR, bias_p, GLSL_TYPE_FLOAT);
+      inst = v->emit(NULL, TGSI_OPCODE_MAD, dst0, src0, scale, bias);
+   }
+
+   if (pixel_maps) {
+      st_src_reg temp = v->get_temp(glsl_type::vec4_type);
+      st_dst_reg temp_dst = st_dst_reg(temp);
+
+      assert(st->pixel_xfer.pixelmap_texture);
+
+      /* With a little effort, we can do four pixel map look-ups with
+       * two TEX instructions:
+       */
+
+      /* TEX temp.rg, colorTemp.rgba, texture[1], 2D; */
+      temp_dst.writemask = WRITEMASK_XY; /* write R,G */
+      inst = v->emit(NULL, TGSI_OPCODE_TEX, temp_dst, src0);
+      inst->sampler = 1;
+      inst->tex_target = TEXTURE_2D_INDEX;
+
+      /* TEX temp.ba, colorTemp.baba, texture[1], 2D; */
+      src0.swizzle = MAKE_SWIZZLE4(SWIZZLE_Z, SWIZZLE_W, SWIZZLE_Z, SWIZZLE_W);
+      temp_dst.writemask = WRITEMASK_ZW; /* write B,A */
+      inst = v->emit(NULL, TGSI_OPCODE_TEX, temp_dst, src0);
+      inst->sampler = 1;
+      inst->tex_target = TEXTURE_2D_INDEX;
+
+      prog->SamplersUsed |= (1 << 1); /* mark sampler 1 as used */
+      v->samplers_used |= (1 << 1);
+
+      /* MOV colorTemp, temp; */
+      inst = v->emit(NULL, TGSI_OPCODE_MOV, dst0, temp);
+   }
+
+   /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
+    * new visitor. */
+   foreach_iter(exec_list_iterator, iter, original->instructions) {
+      glsl_to_tgsi_instruction *inst = (glsl_to_tgsi_instruction *)iter.get();
+      st_src_reg src_regs[3];
+
+      if (inst->dst.file == PROGRAM_OUTPUT)
+         prog->OutputsWritten |= BITFIELD64_BIT(inst->dst.index);
+
+      for (int i=0; i<3; i++) {
+         src_regs[i] = inst->src[i];
+         if (src_regs[i].file == PROGRAM_INPUT &&
+             src_regs[i].index == FRAG_ATTRIB_COL0)
+         {
+            src_regs[i].file = PROGRAM_TEMPORARY;
+            src_regs[i].index = src0.index;
+         }
+         else if (src_regs[i].file == PROGRAM_INPUT)
+            prog->InputsRead |= (1 << src_regs[i].index);
+      }
+
+      v->emit(NULL, inst->op, inst->dst, src_regs[0], src_regs[1], src_regs[2]);
+   }
+
+   /* Make modifications to fragment program info. */
+   prog->Parameters = _mesa_combine_parameter_lists(params,
+                                                    original->prog->Parameters);
+   prog->Attributes = _mesa_clone_parameter_list(original->prog->Attributes);
+   prog->Varying = _mesa_clone_parameter_list(original->prog->Varying);
+   _mesa_free_parameter_list(params);
+   count_resources(v, prog);
+   fp->glsl_to_tgsi = v;
+}
+
+/**
+ * Make fragment program for glBitmap:
+ *   Sample the texture and kill the fragment if the bit is 0.
+ * This program will be combined with the user's fragment program.
+ *
+ * Based on make_bitmap_fragment_program in st_cb_bitmap.c.
+ */
+extern "C" void
+get_bitmap_visitor(struct st_fragment_program *fp,
+                   glsl_to_tgsi_visitor *original, int samplerIndex)
+{
+   glsl_to_tgsi_visitor *v = new glsl_to_tgsi_visitor();
+   struct st_context *st = st_context(original->ctx);
+   struct gl_program *prog = &fp->Base.Base;
+   st_src_reg coord, src0;
+   st_dst_reg dst0;
+   glsl_to_tgsi_instruction *inst;
+
+   /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
+   v->ctx = original->ctx;
+   v->prog = prog;
+   v->glsl_version = original->glsl_version;
+   v->options = original->options;
+   v->next_temp = original->next_temp;
+   v->num_address_regs = original->num_address_regs;
+   v->samplers_used = prog->SamplersUsed = original->samplers_used;
+   v->indirect_addr_temps = original->indirect_addr_temps;
+   v->indirect_addr_consts = original->indirect_addr_consts;
+   memcpy(&v->immediates, &original->immediates, sizeof(v->immediates));
+
+   /* TEX tmp0, fragment.texcoord[0], texture[0], 2D; */
+   coord = st_src_reg(PROGRAM_INPUT, FRAG_ATTRIB_TEX0, glsl_type::vec2_type);
+   src0 = v->get_temp(glsl_type::vec4_type);
+   dst0 = st_dst_reg(src0);
+   inst = v->emit(NULL, TGSI_OPCODE_TEX, dst0, coord);
+   inst->sampler = samplerIndex;
+   inst->tex_target = TEXTURE_2D_INDEX;
+
+   prog->InputsRead |= (1 << FRAG_ATTRIB_TEX0);
+   prog->SamplersUsed |= (1 << samplerIndex); /* mark sampler as used */
+   v->samplers_used |= (1 << samplerIndex);
+
+   /* KIL if -tmp0 < 0 # texel=0 -> keep / texel=0 -> discard */
+   src0.negate = NEGATE_XYZW;
+   if (st->bitmap.tex_format == PIPE_FORMAT_L8_UNORM)
+      src0.swizzle = SWIZZLE_XXXX;
+   inst = v->emit(NULL, TGSI_OPCODE_KIL, undef_dst, src0);
+
+   /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
+    * new visitor. */
+   foreach_iter(exec_list_iterator, iter, original->instructions) {
+      glsl_to_tgsi_instruction *inst = (glsl_to_tgsi_instruction *)iter.get();
+      st_src_reg src_regs[3];
+
+      if (inst->dst.file == PROGRAM_OUTPUT)
+         prog->OutputsWritten |= BITFIELD64_BIT(inst->dst.index);
+
+      for (int i=0; i<3; i++) {
+         src_regs[i] = inst->src[i];
+         if (src_regs[i].file == PROGRAM_INPUT)
+            prog->InputsRead |= (1 << src_regs[i].index);
+      }
+
+      v->emit(NULL, inst->op, inst->dst, src_regs[0], src_regs[1], src_regs[2]);
+   }
+
+   /* Make modifications to fragment program info. */
+   prog->Parameters = _mesa_clone_parameter_list(original->prog->Parameters);
+   prog->Attributes = _mesa_clone_parameter_list(original->prog->Attributes);
+   prog->Varying = _mesa_clone_parameter_list(original->prog->Varying);
+   count_resources(v, prog);
+   fp->glsl_to_tgsi = v;
+}
+
+/* ------------------------- TGSI conversion stuff -------------------------- */
+struct label {
+   unsigned branch_target;
+   unsigned token;
+};
+
+/**
+ * Intermediate state used during shader translation.
+ */
+struct st_translate {
+   struct ureg_program *ureg;
+
+   struct ureg_dst temps[MAX_TEMPS];
+   struct ureg_src *constants;
+   struct ureg_src *immediates;
+   struct ureg_dst outputs[PIPE_MAX_SHADER_OUTPUTS];
+   struct ureg_src inputs[PIPE_MAX_SHADER_INPUTS];
+   struct ureg_dst address[1];
+   struct ureg_src samplers[PIPE_MAX_SAMPLERS];
+   struct ureg_src systemValues[SYSTEM_VALUE_MAX];
+
+   /* Extra info for handling point size clamping in vertex shader */
+   struct ureg_dst pointSizeResult; /**< Actual point size output register */
+   struct ureg_src pointSizeConst;  /**< Point size range constant register */
+   GLint pointSizeOutIndex;         /**< Temp point size output register */
+   GLboolean prevInstWrotePointSize;
+
+   const GLuint *inputMapping;
+   const GLuint *outputMapping;
+
+   /* For every instruction that contains a label (eg CALL), keep
+    * details so that we can go back afterwards and emit the correct
+    * tgsi instruction number for each label.
+    */
+   struct label *labels;
+   unsigned labels_size;
+   unsigned labels_count;
+
+   /* Keep a record of the tgsi instruction number that each mesa
+    * instruction starts at, will be used to fix up labels after
+    * translation.
+    */
+   unsigned *insn;
+   unsigned insn_size;
+   unsigned insn_count;
+
+   unsigned procType;  /**< TGSI_PROCESSOR_VERTEX/FRAGMENT */
+
+   boolean error;
+};
+
+/** Map Mesa's SYSTEM_VALUE_x to TGSI_SEMANTIC_x */
+static unsigned mesa_sysval_to_semantic[SYSTEM_VALUE_MAX] = {
+   TGSI_SEMANTIC_FACE,
+   TGSI_SEMANTIC_INSTANCEID
+};
+
+/**
+ * Make note of a branch to a label in the TGSI code.
+ * After we've emitted all instructions, we'll go over the list
+ * of labels built here and patch the TGSI code with the actual
+ * location of each label.
+ */
+static unsigned *get_label(struct st_translate *t, unsigned branch_target)
+{
+   unsigned i;
+
+   if (t->labels_count + 1 >= t->labels_size) {
+      t->labels_size = 1 << (util_logbase2(t->labels_size) + 1);
+      t->labels = (struct label *)realloc(t->labels, 
+                                          t->labels_size * sizeof(struct label));
+      if (t->labels == NULL) {
+         static unsigned dummy;
+         t->error = TRUE;
+         return &dummy;
+      }
+   }
+
+   i = t->labels_count++;
+   t->labels[i].branch_target = branch_target;
+   return &t->labels[i].token;
+}
+
+/**
+ * Called prior to emitting the TGSI code for each instruction.
+ * Allocate additional space for instructions if needed.
+ * Update the insn[] array so the next glsl_to_tgsi_instruction points to
+ * the next TGSI instruction.
+ */
+static void set_insn_start(struct st_translate *t, unsigned start)
+{
+   if (t->insn_count + 1 >= t->insn_size) {
+      t->insn_size = 1 << (util_logbase2(t->insn_size) + 1);
+      t->insn = (unsigned *)realloc(t->insn, t->insn_size * sizeof(t->insn[0]));
+      if (t->insn == NULL) {
+         t->error = TRUE;
+         return;
+      }
+   }
+
+   t->insn[t->insn_count++] = start;
+}
+
+/**
+ * Map a glsl_to_tgsi constant/immediate to a TGSI immediate.
+ */
+static struct ureg_src
+emit_immediate(struct st_translate *t,
+               gl_constant_value values[4],
+               int type, int size)
+{
+   struct ureg_program *ureg = t->ureg;
+
+   switch(type)
+   {
+   case GL_FLOAT:
+      return ureg_DECL_immediate(ureg, &values[0].f, size);
+   case GL_INT:
+      return ureg_DECL_immediate_int(ureg, &values[0].i, size);
+   case GL_UNSIGNED_INT:
+   case GL_BOOL:
+      return ureg_DECL_immediate_uint(ureg, &values[0].u, size);
+   default:
+      assert(!"should not get here - type must be float, int, uint, or bool");
+      return ureg_src_undef();
+   }
+}
+
+/**
+ * Map a glsl_to_tgsi dst register to a TGSI ureg_dst register.
+ */
+static struct ureg_dst
+dst_register(struct st_translate *t,
+             gl_register_file file,
+             GLuint index)
+{
+   switch(file) {
+   case PROGRAM_UNDEFINED:
+      return ureg_dst_undef();
+
+   case PROGRAM_TEMPORARY:
+      if (ureg_dst_is_undef(t->temps[index]))
+         t->temps[index] = ureg_DECL_temporary(t->ureg);
+
+      return t->temps[index];
+
+   case PROGRAM_OUTPUT:
+      if (t->procType == TGSI_PROCESSOR_VERTEX && index == VERT_RESULT_PSIZ)
+         t->prevInstWrotePointSize = GL_TRUE;
+
+      if (t->procType == TGSI_PROCESSOR_VERTEX)
+         assert(index < VERT_RESULT_MAX);
+      else if (t->procType == TGSI_PROCESSOR_FRAGMENT)
+         assert(index < FRAG_RESULT_MAX);
+      else
+         assert(index < GEOM_RESULT_MAX);
+
+      assert(t->outputMapping[index] < Elements(t->outputs));
+
+      return t->outputs[t->outputMapping[index]];
+
+   case PROGRAM_ADDRESS:
+      return t->address[index];
+
+   default:
+      assert(!"unknown dst register file");
+      return ureg_dst_undef();
+   }
+}
+
+/**
+ * Map a glsl_to_tgsi src register to a TGSI ureg_src register.
+ */
+static struct ureg_src
+src_register(struct st_translate *t,
+             gl_register_file file,
+             GLuint index)
+{
+   switch(file) {
+   case PROGRAM_UNDEFINED:
+      return ureg_src_undef();
+
+   case PROGRAM_TEMPORARY:
+      assert(index >= 0);
+      assert(index < Elements(t->temps));
+      if (ureg_dst_is_undef(t->temps[index]))
+         t->temps[index] = ureg_DECL_temporary(t->ureg);
+      return ureg_src(t->temps[index]);
+
+   case PROGRAM_NAMED_PARAM:
+   case PROGRAM_ENV_PARAM:
+   case PROGRAM_LOCAL_PARAM:
+   case PROGRAM_UNIFORM:
+      assert(index >= 0);
+      return t->constants[index];
+   case PROGRAM_STATE_VAR:
+   case PROGRAM_CONSTANT:       /* ie, immediate */
+      if (index < 0)
+         return ureg_DECL_constant(t->ureg, 0);
+      else
+         return t->constants[index];
+
+   case PROGRAM_IMMEDIATE:
+      return t->immediates[index];
+
+   case PROGRAM_INPUT:
+      assert(t->inputMapping[index] < Elements(t->inputs));
+      return t->inputs[t->inputMapping[index]];
+
+   case PROGRAM_OUTPUT:
+      assert(t->outputMapping[index] < Elements(t->outputs));
+      return ureg_src(t->outputs[t->outputMapping[index]]); /* not needed? */
+
+   case PROGRAM_ADDRESS:
+      return ureg_src(t->address[index]);
+
+   case PROGRAM_SYSTEM_VALUE:
+      assert(index < Elements(t->systemValues));
+      return t->systemValues[index];
+
+   default:
+      assert(!"unknown src register file");
+      return ureg_src_undef();
+   }
+}
+
+/**
+ * Create a TGSI ureg_dst register from an st_dst_reg.
+ */
+static struct ureg_dst
+translate_dst(struct st_translate *t,
+              const st_dst_reg *dst_reg,
+              bool saturate)
+{
+   struct ureg_dst dst = dst_register(t, 
+                                      dst_reg->file,
+                                      dst_reg->index);
+
+   dst = ureg_writemask(dst, dst_reg->writemask);
+   
+   if (saturate)
+      dst = ureg_saturate(dst);
+
+   if (dst_reg->reladdr != NULL)
+      dst = ureg_dst_indirect(dst, ureg_src(t->address[0]));
+
+   return dst;
+}
+
+/**
+ * Create a TGSI ureg_src register from an st_src_reg.
+ */
+static struct ureg_src
+translate_src(struct st_translate *t, const st_src_reg *src_reg)
+{
+   struct ureg_src src = src_register(t, src_reg->file, src_reg->index);
+
+   src = ureg_swizzle(src,
+                      GET_SWZ(src_reg->swizzle, 0) & 0x3,
+                      GET_SWZ(src_reg->swizzle, 1) & 0x3,
+                      GET_SWZ(src_reg->swizzle, 2) & 0x3,
+                      GET_SWZ(src_reg->swizzle, 3) & 0x3);
+
+   if ((src_reg->negate & 0xf) == NEGATE_XYZW)
+      src = ureg_negate(src);
+
+   if (src_reg->reladdr != NULL) {
+      /* Normally ureg_src_indirect() would be used here, but a stupid compiler 
+       * bug in g++ makes ureg_src_indirect (an inline C function) erroneously 
+       * set the bit for src.Negate.  So we have to do the operation manually
+       * here to work around the compiler's problems. */
+      /*src = ureg_src_indirect(src, ureg_src(t->address[0]));*/
+      struct ureg_src addr = ureg_src(t->address[0]);
+      src.Indirect = 1;
+      src.IndirectFile = addr.File;
+      src.IndirectIndex = addr.Index;
+      src.IndirectSwizzle = addr.SwizzleX;
+      
+      if (src_reg->file != PROGRAM_INPUT &&
+          src_reg->file != PROGRAM_OUTPUT) {
+         /* If src_reg->index was negative, it was set to zero in
+          * src_register().  Reassign it now.  But don't do this
+          * for input/output regs since they get remapped while
+          * const buffers don't.
+          */
+         src.Index = src_reg->index;
+      }
+   }
+
+   return src;
+}
+
+static void
+compile_tgsi_instruction(struct st_translate *t,
+                         const struct glsl_to_tgsi_instruction *inst)
+{
+   struct ureg_program *ureg = t->ureg;
+   GLuint i;
+   struct ureg_dst dst[1];
+   struct ureg_src src[4];
+   unsigned num_dst;
+   unsigned num_src;
+
+   num_dst = num_inst_dst_regs(inst->op);
+   num_src = num_inst_src_regs(inst->op);
+
+   if (num_dst) 
+      dst[0] = translate_dst(t, 
+                             &inst->dst,
+                             inst->saturate);
+
+   for (i = 0; i < num_src; i++) 
+      src[i] = translate_src(t, &inst->src[i]);
+
+   switch(inst->op) {
+   case TGSI_OPCODE_BGNLOOP:
+   case TGSI_OPCODE_CAL:
+   case TGSI_OPCODE_ELSE:
+   case TGSI_OPCODE_ENDLOOP:
+   case TGSI_OPCODE_IF:
+      assert(num_dst == 0);
+      ureg_label_insn(ureg,
+                      inst->op,
+                      src, num_src,
+                      get_label(t, 
+                                inst->op == TGSI_OPCODE_CAL ? inst->function->sig_id : 0));
+      return;
+
+   case TGSI_OPCODE_TEX:
+   case TGSI_OPCODE_TXB:
+   case TGSI_OPCODE_TXD:
+   case TGSI_OPCODE_TXL:
+   case TGSI_OPCODE_TXP:
+      src[num_src++] = t->samplers[inst->sampler];
+      ureg_tex_insn(ureg,
+                    inst->op,
+                    dst, num_dst, 
+                    translate_texture_target(inst->tex_target, inst->tex_shadow),
+                    src, num_src);
+      return;
+
+   case TGSI_OPCODE_SCS:
+      dst[0] = ureg_writemask(dst[0], TGSI_WRITEMASK_XY);
+      ureg_insn(ureg, inst->op, dst, num_dst, src, num_src);
+      break;
+
+   default:
+      ureg_insn(ureg,
+                inst->op,
+                dst, num_dst,
+                src, num_src);
+      break;
+   }
+}
+
+/**
+ * Emit the TGSI instructions to adjust the WPOS pixel center convention
+ * Basically, add (adjX, adjY) to the fragment position.
+ */
+static void
+emit_adjusted_wpos(struct st_translate *t,
+                   const struct gl_program *program,
+                   float adjX, float adjY)
+{
+   struct ureg_program *ureg = t->ureg;
+   struct ureg_dst wpos_temp = ureg_DECL_temporary(ureg);
+   struct ureg_src wpos_input = t->inputs[t->inputMapping[FRAG_ATTRIB_WPOS]];
+
+   /* Note that we bias X and Y and pass Z and W through unchanged.
+    * The shader might also use gl_FragCoord.w and .z.
+    */
+   ureg_ADD(ureg, wpos_temp, wpos_input,
+            ureg_imm4f(ureg, adjX, adjY, 0.0f, 0.0f));
+
+   t->inputs[t->inputMapping[FRAG_ATTRIB_WPOS]] = ureg_src(wpos_temp);
+}
+
+
+/**
+ * Emit the TGSI instructions for inverting the WPOS y coordinate.
+ * This code is unavoidable because it also depends on whether
+ * a FBO is bound (STATE_FB_WPOS_Y_TRANSFORM).
+ */
+static void
+emit_wpos_inversion(struct st_translate *t,
+                    const struct gl_program *program,
+                    bool invert)
+{
+   struct ureg_program *ureg = t->ureg;
+
+   /* Fragment program uses fragment position input.
+    * Need to replace instances of INPUT[WPOS] with temp T
+    * where T = INPUT[WPOS] by y is inverted.
+    */
+   static const gl_state_index wposTransformState[STATE_LENGTH]
+      = { STATE_INTERNAL, STATE_FB_WPOS_Y_TRANSFORM, 
+          (gl_state_index)0, (gl_state_index)0, (gl_state_index)0 };
+   
+   /* XXX: note we are modifying the incoming shader here!  Need to
+    * do this before emitting the constant decls below, or this
+    * will be missed:
+    */
+   unsigned wposTransConst = _mesa_add_state_reference(program->Parameters,
+                                                       wposTransformState);
+
+   struct ureg_src wpostrans = ureg_DECL_constant(ureg, wposTransConst);
+   struct ureg_dst wpos_temp;
+   struct ureg_src wpos_input = t->inputs[t->inputMapping[FRAG_ATTRIB_WPOS]];
+
+   /* MOV wpos_temp, input[wpos]
+    */
+   if (wpos_input.File == TGSI_FILE_TEMPORARY)
+      wpos_temp = ureg_dst(wpos_input);
+   else {
+      wpos_temp = ureg_DECL_temporary(ureg);
+      ureg_MOV(ureg, wpos_temp, wpos_input);
+   }
+
+   if (invert) {
+      /* MAD wpos_temp.y, wpos_input, wpostrans.xxxx, wpostrans.yyyy
+       */
+      ureg_MAD(ureg,
+               ureg_writemask(wpos_temp, TGSI_WRITEMASK_Y),
+               wpos_input,
+               ureg_scalar(wpostrans, 0),
+               ureg_scalar(wpostrans, 1));
+   } else {
+      /* MAD wpos_temp.y, wpos_input, wpostrans.zzzz, wpostrans.wwww
+       */
+      ureg_MAD(ureg,
+               ureg_writemask(wpos_temp, TGSI_WRITEMASK_Y),
+               wpos_input,
+               ureg_scalar(wpostrans, 2),
+               ureg_scalar(wpostrans, 3));
+   }
+
+   /* Use wpos_temp as position input from here on:
+    */
+   t->inputs[t->inputMapping[FRAG_ATTRIB_WPOS]] = ureg_src(wpos_temp);
+}
+
+
+/**
+ * Emit fragment position/ooordinate code.
+ */
+static void
+emit_wpos(struct st_context *st,
+          struct st_translate *t,
+          const struct gl_program *program,
+          struct ureg_program *ureg)
+{
+   const struct gl_fragment_program *fp =
+      (const struct gl_fragment_program *) program;
+   struct pipe_screen *pscreen = st->pipe->screen;
+   boolean invert = FALSE;
+
+   if (fp->OriginUpperLeft) {
+      /* Fragment shader wants origin in upper-left */
+      if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT)) {
+         /* the driver supports upper-left origin */
+      }
+      else if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT)) {
+         /* the driver supports lower-left origin, need to invert Y */
+         ureg_property_fs_coord_origin(ureg, TGSI_FS_COORD_ORIGIN_LOWER_LEFT);
+         invert = TRUE;
+      }
+      else
+         assert(0);
+   }
+   else {
+      /* Fragment shader wants origin in lower-left */
+      if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT))
+         /* the driver supports lower-left origin */
+         ureg_property_fs_coord_origin(ureg, TGSI_FS_COORD_ORIGIN_LOWER_LEFT);
+      else if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT))
+         /* the driver supports upper-left origin, need to invert Y */
+         invert = TRUE;
+      else
+         assert(0);
+   }
+   
+   if (fp->PixelCenterInteger) {
+      /* Fragment shader wants pixel center integer */
+      if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER))
+         /* the driver supports pixel center integer */
+         ureg_property_fs_coord_pixel_center(ureg, TGSI_FS_COORD_PIXEL_CENTER_INTEGER);
+      else if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER))
+         /* the driver supports pixel center half integer, need to bias X,Y */
+         emit_adjusted_wpos(t, program, 0.5f, invert ? 0.5f : -0.5f);
+      else
+         assert(0);
+   }
+   else {
+      /* Fragment shader wants pixel center half integer */
+      if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER)) {
+         /* the driver supports pixel center half integer */
+      }
+      else if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER)) {
+         /* the driver supports pixel center integer, need to bias X,Y */
+         ureg_property_fs_coord_pixel_center(ureg, TGSI_FS_COORD_PIXEL_CENTER_INTEGER);
+         emit_adjusted_wpos(t, program, 0.5f, invert ? -0.5f : 0.5f);
+      }
+      else
+         assert(0);
+   }
+
+   /* we invert after adjustment so that we avoid the MOV to temporary,
+    * and reuse the adjustment ADD instead */
+   emit_wpos_inversion(t, program, invert);
+}
+
+/**
+ * OpenGL's fragment gl_FrontFace input is 1 for front-facing, 0 for back.
+ * TGSI uses +1 for front, -1 for back.
+ * This function converts the TGSI value to the GL value.  Simply clamping/
+ * saturating the value to [0,1] does the job.
+ */
+static void
+emit_face_var(struct st_translate *t)
+{
+   struct ureg_program *ureg = t->ureg;
+   struct ureg_dst face_temp = ureg_DECL_temporary(ureg);
+   struct ureg_src face_input = t->inputs[t->inputMapping[FRAG_ATTRIB_FACE]];
+
+   /* MOV_SAT face_temp, input[face] */
+   face_temp = ureg_saturate(face_temp);
+   ureg_MOV(ureg, face_temp, face_input);
+
+   /* Use face_temp as face input from here on: */
+   t->inputs[t->inputMapping[FRAG_ATTRIB_FACE]] = ureg_src(face_temp);
+}
+
+static void
+emit_edgeflags(struct st_translate *t)
+{
+   struct ureg_program *ureg = t->ureg;
+   struct ureg_dst edge_dst = t->outputs[t->outputMapping[VERT_RESULT_EDGE]];
+   struct ureg_src edge_src = t->inputs[t->inputMapping[VERT_ATTRIB_EDGEFLAG]];
+
+   ureg_MOV(ureg, edge_dst, edge_src);
+}
+
+/**
+ * Translate intermediate IR (glsl_to_tgsi_instruction) to TGSI format.
+ * \param program  the program to translate
+ * \param numInputs  number of input registers used
+ * \param inputMapping  maps Mesa fragment program inputs to TGSI generic
+ *                      input indexes
+ * \param inputSemanticName  the TGSI_SEMANTIC flag for each input
+ * \param inputSemanticIndex  the semantic index (ex: which texcoord) for
+ *                            each input
+ * \param interpMode  the TGSI_INTERPOLATE_LINEAR/PERSP mode for each input
+ * \param numOutputs  number of output registers used
+ * \param outputMapping  maps Mesa fragment program outputs to TGSI
+ *                       generic outputs
+ * \param outputSemanticName  the TGSI_SEMANTIC flag for each output
+ * \param outputSemanticIndex  the semantic index (ex: which texcoord) for
+ *                             each output
+ *
+ * \return  PIPE_OK or PIPE_ERROR_OUT_OF_MEMORY
+ */
+extern "C" enum pipe_error
+st_translate_program(
+   struct gl_context *ctx,
+   uint procType,
+   struct ureg_program *ureg,
+   glsl_to_tgsi_visitor *program,
+   const struct gl_program *proginfo,
+   GLuint numInputs,
+   const GLuint inputMapping[],
+   const ubyte inputSemanticName[],
+   const ubyte inputSemanticIndex[],
+   const GLuint interpMode[],
+   GLuint numOutputs,
+   const GLuint outputMapping[],
+   const ubyte outputSemanticName[],
+   const ubyte outputSemanticIndex[],
+   boolean passthrough_edgeflags)
+{
+   struct st_translate translate, *t;
+   unsigned i;
+   enum pipe_error ret = PIPE_OK;
+
+   assert(numInputs <= Elements(t->inputs));
+   assert(numOutputs <= Elements(t->outputs));
+
+   t = &translate;
+   memset(t, 0, sizeof *t);
+
+   t->procType = procType;
+   t->inputMapping = inputMapping;
+   t->outputMapping = outputMapping;
+   t->ureg = ureg;
+   t->pointSizeOutIndex = -1;
+   t->prevInstWrotePointSize = GL_FALSE;
+
+   /*
+    * Declare input attributes.
+    */
+   if (procType == TGSI_PROCESSOR_FRAGMENT) {
+      for (i = 0; i < numInputs; i++) {
+         t->inputs[i] = ureg_DECL_fs_input(ureg,
+                                           inputSemanticName[i],
+                                           inputSemanticIndex[i],
+                                           interpMode[i]);
+      }
+
+      if (proginfo->InputsRead & FRAG_BIT_WPOS) {
+         /* Must do this after setting up t->inputs, and before
+          * emitting constant references, below:
+          */
+          emit_wpos(st_context(ctx), t, proginfo, ureg);
+      }
+
+      if (proginfo->InputsRead & FRAG_BIT_FACE)
+         emit_face_var(t);
+
+      /*
+       * Declare output attributes.
+       */
+      for (i = 0; i < numOutputs; i++) {
+         switch (outputSemanticName[i]) {
+         case TGSI_SEMANTIC_POSITION:
+            t->outputs[i] = ureg_DECL_output(ureg,
+                                             TGSI_SEMANTIC_POSITION, /* Z/Depth */
+                                             outputSemanticIndex[i]);
+            t->outputs[i] = ureg_writemask(t->outputs[i], TGSI_WRITEMASK_Z);
+            break;
+         case TGSI_SEMANTIC_STENCIL:
+            t->outputs[i] = ureg_DECL_output(ureg,
+                                             TGSI_SEMANTIC_STENCIL, /* Stencil */
+                                             outputSemanticIndex[i]);
+            t->outputs[i] = ureg_writemask(t->outputs[i], TGSI_WRITEMASK_Y);
+            break;
+         case TGSI_SEMANTIC_COLOR:
+            t->outputs[i] = ureg_DECL_output(ureg,
+                                             TGSI_SEMANTIC_COLOR,
+                                             outputSemanticIndex[i]);
+            break;
+         default:
+            assert(!"fragment shader outputs must be POSITION/STENCIL/COLOR");
+            return PIPE_ERROR_BAD_INPUT;
+         }
+      }
+   }
+   else if (procType == TGSI_PROCESSOR_GEOMETRY) {
+      for (i = 0; i < numInputs; i++) {
+         t->inputs[i] = ureg_DECL_gs_input(ureg,
+                                           i,
+                                           inputSemanticName[i],
+                                           inputSemanticIndex[i]);
+      }
+
+      for (i = 0; i < numOutputs; i++) {
+         t->outputs[i] = ureg_DECL_output(ureg,
+                                          outputSemanticName[i],
+                                          outputSemanticIndex[i]);
+      }
+   }
+   else {
+      assert(procType == TGSI_PROCESSOR_VERTEX);
+
+      for (i = 0; i < numInputs; i++) {
+         t->inputs[i] = ureg_DECL_vs_input(ureg, i);
+      }
+
+      for (i = 0; i < numOutputs; i++) {
+         t->outputs[i] = ureg_DECL_output(ureg,
+                                          outputSemanticName[i],
+                                          outputSemanticIndex[i]);
+         if ((outputSemanticName[i] == TGSI_SEMANTIC_PSIZE) && proginfo->Id) {
+            /* Writing to the point size result register requires special
+             * handling to implement clamping.
+             */
+            static const gl_state_index pointSizeClampState[STATE_LENGTH]
+               = { STATE_INTERNAL, STATE_POINT_SIZE_IMPL_CLAMP, (gl_state_index)0, (gl_state_index)0, (gl_state_index)0 };
+               /* XXX: note we are modifying the incoming shader here!  Need to
+               * do this before emitting the constant decls below, or this
+               * will be missed.
+               */
+            unsigned pointSizeClampConst =
+               _mesa_add_state_reference(proginfo->Parameters,
+                                         pointSizeClampState);
+            struct ureg_dst psizregtemp = ureg_DECL_temporary(ureg);
+            t->pointSizeConst = ureg_DECL_constant(ureg, pointSizeClampConst);
+            t->pointSizeResult = t->outputs[i];
+            t->pointSizeOutIndex = i;
+            t->outputs[i] = psizregtemp;
+         }
+      }
+      if (passthrough_edgeflags)
+         emit_edgeflags(t);
+   }
+
+   /* Declare address register.
+    */
+   if (program->num_address_regs > 0) {
+      assert(program->num_address_regs == 1);
+      t->address[0] = ureg_DECL_address(ureg);
+   }
+
+   /* Declare misc input registers
+    */
+   {
+      GLbitfield sysInputs = proginfo->SystemValuesRead;
+      unsigned numSys = 0;
+      for (i = 0; sysInputs; i++) {
+         if (sysInputs & (1 << i)) {
+            unsigned semName = mesa_sysval_to_semantic[i];
+            t->systemValues[i] = ureg_DECL_system_value(ureg, numSys, semName, 0);
+            numSys++;
+            sysInputs &= ~(1 << i);
+         }
+      }
+   }
+
+   if (program->indirect_addr_temps) {
+      /* If temps are accessed with indirect addressing, declare temporaries
+       * in sequential order.  Else, we declare them on demand elsewhere.
+       * (Note: the number of temporaries is equal to program->next_temp)
+       */
+      for (i = 0; i < (unsigned)program->next_temp; i++) {
+         /* XXX use TGSI_FILE_TEMPORARY_ARRAY when it's supported by ureg */
+         t->temps[i] = ureg_DECL_temporary(t->ureg);
+      }
+   }
+
+   /* Emit constants and uniforms.  TGSI uses a single index space for these, 
+    * so we put all the translated regs in t->constants.
+    */
+   if (proginfo->Parameters) {
+      t->constants = (struct ureg_src *)CALLOC(proginfo->Parameters->NumParameters * sizeof(t->constants[0]));
+      if (t->constants == NULL) {
+         ret = PIPE_ERROR_OUT_OF_MEMORY;
+         goto out;
+      }
+
+      for (i = 0; i < proginfo->Parameters->NumParameters; i++) {
+         switch (proginfo->Parameters->Parameters[i].Type) {
+         case PROGRAM_ENV_PARAM:
+         case PROGRAM_LOCAL_PARAM:
+         case PROGRAM_STATE_VAR:
+         case PROGRAM_NAMED_PARAM:
+         case PROGRAM_UNIFORM:
+            t->constants[i] = ureg_DECL_constant(ureg, i);
+            break;
+
+         /* Emit immediates for PROGRAM_CONSTANT only when there's no indirect
+          * addressing of the const buffer.
+          * FIXME: Be smarter and recognize param arrays:
+          * indirect addressing is only valid within the referenced
+          * array.
+          */
+         case PROGRAM_CONSTANT:
+            if (program->indirect_addr_consts)
+               t->constants[i] = ureg_DECL_constant(ureg, i);
+            else
+               t->constants[i] = emit_immediate(t,
+                                                proginfo->Parameters->ParameterValues[i],
+                                                proginfo->Parameters->Parameters[i].DataType,
+                                                4);
+            break;
+         default:
+            break;
+         }
+      }
+   }
+   
+   /* Emit immediate values.
+    */
+   t->immediates = (struct ureg_src *)CALLOC(program->num_immediates * sizeof(struct ureg_src));
+   if (t->immediates == NULL) {
+      ret = PIPE_ERROR_OUT_OF_MEMORY;
+      goto out;
+   }
+   i = 0;
+   foreach_iter(exec_list_iterator, iter, program->immediates) {
+      immediate_storage *imm = (immediate_storage *)iter.get();
+      t->immediates[i++] = emit_immediate(t, imm->values, imm->type, imm->size);
+   }
+
+   /* texture samplers */
+   for (i = 0; i < ctx->Const.MaxTextureImageUnits; i++) {
+      if (program->samplers_used & (1 << i)) {
+         t->samplers[i] = ureg_DECL_sampler(ureg, i);
+      }
+   }
+
+   /* Emit each instruction in turn:
+    */
+   foreach_iter(exec_list_iterator, iter, program->instructions) {
+      set_insn_start(t, ureg_get_instruction_number(ureg));
+      compile_tgsi_instruction(t, (glsl_to_tgsi_instruction *)iter.get());
+
+      if (t->prevInstWrotePointSize && proginfo->Id) {
+         /* The previous instruction wrote to the (fake) vertex point size
+          * result register.  Now we need to clamp that value to the min/max
+          * point size range, putting the result into the real point size
+          * register.
+          * Note that we can't do this easily at the end of program due to
+          * possible early return.
+          */
+         set_insn_start(t, ureg_get_instruction_number(ureg));
+         ureg_MAX(t->ureg,
+                  ureg_writemask(t->outputs[t->pointSizeOutIndex], WRITEMASK_X),
+                  ureg_src(t->outputs[t->pointSizeOutIndex]),
+                  ureg_swizzle(t->pointSizeConst, 1,1,1,1));
+         ureg_MIN(t->ureg, ureg_writemask(t->pointSizeResult, WRITEMASK_X),
+                  ureg_src(t->outputs[t->pointSizeOutIndex]),
+                  ureg_swizzle(t->pointSizeConst, 2,2,2,2));
+      }
+      t->prevInstWrotePointSize = GL_FALSE;
+   }
+
+   /* Fix up all emitted labels:
+    */
+   for (i = 0; i < t->labels_count; i++) {
+      ureg_fixup_label(ureg, t->labels[i].token,
+                       t->insn[t->labels[i].branch_target]);
+   }
+
+out:
+   FREE(t->insn);
+   FREE(t->labels);
+   FREE(t->constants);
+   FREE(t->immediates);
+
+   if (t->error) {
+      debug_printf("%s: translate error flag set\n", __FUNCTION__);
+   }
+
+   return ret;
+}
+/* ----------------------------- End TGSI code ------------------------------ */
+
+/**
+ * Convert a shader's GLSL IR into a Mesa gl_program, although without 
+ * generating Mesa IR.
+ */
+static struct gl_program *
+get_mesa_program(struct gl_context *ctx,
+                 struct gl_shader_program *shader_program,
+                struct gl_shader *shader)
+{
+   glsl_to_tgsi_visitor* v = new glsl_to_tgsi_visitor();
+   struct gl_program *prog;
+   GLenum target;
+   const char *target_string;
+   bool progress;
+   struct gl_shader_compiler_options *options =
+         &ctx->ShaderCompilerOptions[_mesa_shader_type_to_index(shader->Type)];
+
+   switch (shader->Type) {
+   case GL_VERTEX_SHADER:
+      target = GL_VERTEX_PROGRAM_ARB;
+      target_string = "vertex";
+      break;
+   case GL_FRAGMENT_SHADER:
+      target = GL_FRAGMENT_PROGRAM_ARB;
+      target_string = "fragment";
+      break;
+   case GL_GEOMETRY_SHADER:
+      target = GL_GEOMETRY_PROGRAM_NV;
+      target_string = "geometry";
+      break;
+   default:
+      assert(!"should not be reached");
+      return NULL;
+   }
+
+   validate_ir_tree(shader->ir);
+
+   prog = ctx->Driver.NewProgram(ctx, target, shader_program->Name);
+   if (!prog)
+      return NULL;
+   prog->Parameters = _mesa_new_parameter_list();
+   prog->Varying = _mesa_new_parameter_list();
+   prog->Attributes = _mesa_new_parameter_list();
+   v->ctx = ctx;
+   v->prog = prog;
+   v->shader_program = shader_program;
+   v->options = options;
+   v->glsl_version = ctx->Const.GLSLVersion;
+
+   add_uniforms_to_parameters_list(shader_program, shader, prog);
+
+   /* Emit intermediate IR for main(). */
+   visit_exec_list(shader->ir, v);
+
+   /* Now emit bodies for any functions that were used. */
+   do {
+      progress = GL_FALSE;
+
+      foreach_iter(exec_list_iterator, iter, v->function_signatures) {
+         function_entry *entry = (function_entry *)iter.get();
+
+         if (!entry->bgn_inst) {
+            v->current_function = entry;
+
+            entry->bgn_inst = v->emit(NULL, TGSI_OPCODE_BGNSUB);
+            entry->bgn_inst->function = entry;
+
+            visit_exec_list(&entry->sig->body, v);
+
+            glsl_to_tgsi_instruction *last;
+            last = (glsl_to_tgsi_instruction *)v->instructions.get_tail();
+            if (last->op != TGSI_OPCODE_RET)
+               v->emit(NULL, TGSI_OPCODE_RET);
+
+            glsl_to_tgsi_instruction *end;
+            end = v->emit(NULL, TGSI_OPCODE_ENDSUB);
+            end->function = entry;
+
+            progress = GL_TRUE;
+         }
+      }
+   } while (progress);
+
+#if 0
+   /* Print out some information (for debugging purposes) used by the 
+    * optimization passes. */
+   for (i=0; i < v->next_temp; i++) {
+      int fr = v->get_first_temp_read(i);
+      int fw = v->get_first_temp_write(i);
+      int lr = v->get_last_temp_read(i);
+      int lw = v->get_last_temp_write(i);
+      
+      printf("Temp %d: FR=%3d FW=%3d LR=%3d LW=%3d\n", i, fr, fw, lr, lw);
+      assert(fw <= fr);
+   }
+#endif
+
+   /* Remove reads to output registers, and to varyings in vertex shaders. */
+   v->remove_output_reads(PROGRAM_OUTPUT);
+   if (target == GL_VERTEX_PROGRAM_ARB)
+      v->remove_output_reads(PROGRAM_VARYING);
+   
+   /* Perform optimizations on the instructions in the glsl_to_tgsi_visitor. */
+   v->simplify_cmp();
+   v->copy_propagate();
+   while (v->eliminate_dead_code_advanced());
+
+   /* FIXME: These passes to optimize temporary registers don't work when there
+    * is indirect addressing of the temporary register space.  We need proper 
+    * array support so that we don't have to give up these passes in every 
+    * shader that uses arrays.
+    */
+   if (!v->indirect_addr_temps) {
+      v->eliminate_dead_code();
+      v->merge_registers();
+      v->renumber_registers();
+   }
+   
+   /* Write the END instruction. */
+   v->emit(NULL, TGSI_OPCODE_END);
+
+   if (ctx->Shader.Flags & GLSL_DUMP) {
+      printf("\n");
+      printf("GLSL IR for linked %s program %d:\n", target_string,
+             shader_program->Name);
+      _mesa_print_ir(shader->ir, NULL);
+      printf("\n");
+      printf("\n");
+   }
+
+   prog->Instructions = NULL;
+   prog->NumInstructions = 0;
+
+   do_set_program_inouts(shader->ir, prog);
+   count_resources(v, prog);
+
+   check_resources(ctx, shader_program, v, prog);
+
+   _mesa_reference_program(ctx, &shader->Program, prog);
+   
+   struct st_vertex_program *stvp;
+   struct st_fragment_program *stfp;
+   struct st_geometry_program *stgp;
+   
+   switch (shader->Type) {
+   case GL_VERTEX_SHADER:
+      stvp = (struct st_vertex_program *)prog;
+      stvp->glsl_to_tgsi = v;
+      break;
+   case GL_FRAGMENT_SHADER:
+      stfp = (struct st_fragment_program *)prog;
+      stfp->glsl_to_tgsi = v;
+      break;
+   case GL_GEOMETRY_SHADER:
+      stgp = (struct st_geometry_program *)prog;
+      stgp->glsl_to_tgsi = v;
+      break;
+   default:
+      assert(!"should not be reached");
+      return NULL;
+   }
+
+   return prog;
+}
+
+extern "C" {
+
+struct gl_shader *
+st_new_shader(struct gl_context *ctx, GLuint name, GLuint type)
+{
+   struct gl_shader *shader;
+   assert(type == GL_FRAGMENT_SHADER || type == GL_VERTEX_SHADER ||
+          type == GL_GEOMETRY_SHADER_ARB);
+   shader = rzalloc(NULL, struct gl_shader);
+   if (shader) {
+      shader->Type = type;
+      shader->Name = name;
+      _mesa_init_shader(ctx, shader);
+   }
+   return shader;
+}
+
+struct gl_shader_program *
+st_new_shader_program(struct gl_context *ctx, GLuint name)
+{
+   struct gl_shader_program *shProg;
+   shProg = rzalloc(NULL, struct gl_shader_program);
+   if (shProg) {
+      shProg->Name = name;
+      _mesa_init_shader_program(ctx, shProg);
+   }
+   return shProg;
+}
+
+/**
+ * Link a shader.
+ * Called via ctx->Driver.LinkShader()
+ * This actually involves converting GLSL IR into an intermediate TGSI-like IR 
+ * with code lowering and other optimizations.
+ */
+GLboolean
+st_link_shader(struct gl_context *ctx, struct gl_shader_program *prog)
+{
+   assert(prog->LinkStatus);
+
+   for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
+      if (prog->_LinkedShaders[i] == NULL)
+         continue;
+
+      bool progress;
+      exec_list *ir = prog->_LinkedShaders[i]->ir;
+      const struct gl_shader_compiler_options *options =
+            &ctx->ShaderCompilerOptions[_mesa_shader_type_to_index(prog->_LinkedShaders[i]->Type)];
+
+      do {
+         progress = false;
+
+         /* Lowering */
+         do_mat_op_to_vec(ir);
+         lower_instructions(ir, (MOD_TO_FRACT | DIV_TO_MUL_RCP | EXP_TO_EXP2
+                                | LOG_TO_LOG2
+                                | ((options->EmitNoPow) ? POW_TO_EXP2 : 0)));
+
+         progress = do_lower_jumps(ir, true, true, options->EmitNoMainReturn, options->EmitNoCont, options->EmitNoLoops) || progress;
+
+         progress = do_common_optimization(ir, true, options->MaxUnrollIterations) || progress;
+
+         progress = lower_quadop_vector(ir, false) || progress;
+
+         if (options->EmitNoIfs) {
+            progress = lower_discard(ir) || progress;
+            progress = lower_if_to_cond_assign(ir) || progress;
+         }
+
+         if (options->EmitNoNoise)
+            progress = lower_noise(ir) || progress;
+
+         /* If there are forms of indirect addressing that the driver
+          * cannot handle, perform the lowering pass.
+          */
+         if (options->EmitNoIndirectInput || options->EmitNoIndirectOutput
+             || options->EmitNoIndirectTemp || options->EmitNoIndirectUniform)
+           progress =
+             lower_variable_index_to_cond_assign(ir,
+                                                options->EmitNoIndirectInput,
+                                                options->EmitNoIndirectOutput,
+                                                options->EmitNoIndirectTemp,
+                                                options->EmitNoIndirectUniform)
+             || progress;
+
+         progress = do_vec_index_to_cond_assign(ir) || progress;
+      } while (progress);
+
+      validate_ir_tree(ir);
+   }
+
+   for (unsigned i = 0; i < MESA_SHADER_TYPES; i++) {
+      struct gl_program *linked_prog;
+
+      if (prog->_LinkedShaders[i] == NULL)
+         continue;
+
+      linked_prog = get_mesa_program(ctx, prog, prog->_LinkedShaders[i]);
+
+      if (linked_prog) {
+         bool ok = true;
+
+         switch (prog->_LinkedShaders[i]->Type) {
+         case GL_VERTEX_SHADER:
+            _mesa_reference_vertprog(ctx, &prog->VertexProgram,
+                                     (struct gl_vertex_program *)linked_prog);
+            ok = ctx->Driver.ProgramStringNotify(ctx, GL_VERTEX_PROGRAM_ARB,
+                                                 linked_prog);
+            break;
+         case GL_FRAGMENT_SHADER:
+            _mesa_reference_fragprog(ctx, &prog->FragmentProgram,
+                                     (struct gl_fragment_program *)linked_prog);
+            ok = ctx->Driver.ProgramStringNotify(ctx, GL_FRAGMENT_PROGRAM_ARB,
+                                                 linked_prog);
+            break;
+         case GL_GEOMETRY_SHADER:
+            _mesa_reference_geomprog(ctx, &prog->GeometryProgram,
+                                     (struct gl_geometry_program *)linked_prog);
+            ok = ctx->Driver.ProgramStringNotify(ctx, GL_GEOMETRY_PROGRAM_NV,
+                                                 linked_prog);
+            break;
+         }
+         if (!ok) {
+            return GL_FALSE;
+         }
+      }
+
+      _mesa_reference_program(ctx, &linked_prog, NULL);
+   }
+
+   return GL_TRUE;
+}
+
+
+/**
+ * Link a GLSL shader program.  Called via glLinkProgram().
+ */
+void
+st_glsl_link_shader(struct gl_context *ctx, struct gl_shader_program *prog)
+{
+   unsigned int i;
+
+   _mesa_clear_shader_program_data(ctx, prog);
+
+   prog->LinkStatus = GL_TRUE;
+
+   for (i = 0; i < prog->NumShaders; i++) {
+      if (!prog->Shaders[i]->CompileStatus) {
+         fail_link(prog, "linking with uncompiled shader");
+         prog->LinkStatus = GL_FALSE;
+      }
+   }
+
+   prog->Varying = _mesa_new_parameter_list();
+   _mesa_reference_vertprog(ctx, &prog->VertexProgram, NULL);
+   _mesa_reference_fragprog(ctx, &prog->FragmentProgram, NULL);
+   _mesa_reference_geomprog(ctx, &prog->GeometryProgram, NULL);
+
+   if (prog->LinkStatus) {
+      link_shaders(ctx, prog);
+   }
+
+   if (prog->LinkStatus) {
+      if (!ctx->Driver.LinkShader(ctx, prog)) {
+         prog->LinkStatus = GL_FALSE;
+      }
+   }
+
+   set_uniform_initializers(ctx, prog);
+
+   if (ctx->Shader.Flags & GLSL_DUMP) {
+      if (!prog->LinkStatus) {
+         printf("GLSL shader program %d failed to link\n", prog->Name);
+      }
+
+      if (prog->InfoLog && prog->InfoLog[0] != 0) {
+         printf("GLSL shader program %d info log:\n", prog->Name);
+         printf("%s\n", prog->InfoLog);
+      }
+   }
+}
+
+} /* extern "C" */
diff --git a/src/mesa/state_tracker/st_glsl_to_tgsi.h b/src/mesa/state_tracker/st_glsl_to_tgsi.h
new file mode 100644 (file)
index 0000000..d877471
--- /dev/null
@@ -0,0 +1,72 @@
+/*
+ * Copyright © 2010 Intel Corporation
+ * Copyright © 2011 Bryan Cain
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
+ */
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "main/glheader.h"
+#include "tgsi/tgsi_ureg.h"
+
+struct gl_context;
+struct gl_shader;
+struct gl_shader_program;
+struct glsl_to_tgsi_visitor;
+
+enum pipe_error st_translate_program(
+   struct gl_context *ctx,
+   uint procType,
+   struct ureg_program *ureg,
+   struct glsl_to_tgsi_visitor *program,
+   const struct gl_program *proginfo,
+   GLuint numInputs,
+   const GLuint inputMapping[],
+   const ubyte inputSemanticName[],
+   const ubyte inputSemanticIndex[],
+   const GLuint interpMode[],
+   GLuint numOutputs,
+   const GLuint outputMapping[],
+   const ubyte outputSemanticName[],
+   const ubyte outputSemanticIndex[],
+   boolean passthrough_edgeflags);
+
+void free_glsl_to_tgsi_visitor(struct glsl_to_tgsi_visitor *v);
+void get_pixel_transfer_visitor(struct st_fragment_program *fp,
+                                struct glsl_to_tgsi_visitor *original,
+                                int scale_and_bias, int pixel_maps);
+void get_bitmap_visitor(struct st_fragment_program *fp,
+                        struct glsl_to_tgsi_visitor *original,
+                        int samplerIndex);
+
+struct gl_shader *st_new_shader(struct gl_context *ctx, GLuint name, GLuint type);
+
+struct gl_shader_program *
+st_new_shader_program(struct gl_context *ctx, GLuint name);
+
+void st_glsl_link_shader(struct gl_context *ctx, struct gl_shader_program *prog);
+GLboolean st_link_shader(struct gl_context *ctx, struct gl_shader_program *prog);
+
+#ifdef __cplusplus
+}
+#endif
index a41e5b16a85efb405ff52ae9f188024bbd747579..656c985d78ffd00d77b22d1ff94a9924beff3845 100644 (file)
@@ -267,7 +267,7 @@ src_register( struct st_translate *t,
 /**
  * Map mesa texture target to TGSI texture target.
  */
-static unsigned
+unsigned
 translate_texture_target( GLuint textarget,
                           GLboolean shadow )
 {
@@ -511,7 +511,7 @@ static void emit_ddy( struct st_translate *t,
 
 
 
-static unsigned
+unsigned
 translate_opcode( unsigned op )
 {
    switch( op ) {
@@ -1207,7 +1207,7 @@ st_translate_mesa_program(
             else
                t->constants[i] = 
                   ureg_DECL_immediate( ureg,
-                                       program->Parameters->ParameterValues[i],
+                                       (const float*) program->Parameters->ParameterValues[i],
                                        4 );
             break;
          default:
index 0615e52ef6297bd5a64c8b078e62b2cc2ddac3fe..0dbdf5f6159bacc914e817fbe41d730d267e7c37 100644 (file)
@@ -64,6 +64,12 @@ st_translate_mesa_program(
 void
 st_free_tokens(const struct tgsi_token *tokens);
 
+unsigned
+translate_opcode(unsigned op);
+
+unsigned
+translate_texture_target(GLuint textarget, GLboolean shadow);
+
 
 #if defined __cplusplus
 } /* extern "C" */
index 132ebdbadc9464e38c7e1ae31f9a354b83320b7b..ca01d2e1976bbbd6fa21d95aa4cb702d0b4669cd 100644 (file)
@@ -174,8 +174,8 @@ st_release_gp_variants(struct st_context *st, struct st_geometry_program *stgp)
  * \param tokensOut  destination for TGSI tokens
  * \return  pointer to cached pipe_shader object.
  */
-static void
-st_prepare_vertex_program(struct st_context *st,
+void
+st_prepare_vertex_program(struct gl_context *ctx,
                             struct st_vertex_program *stvp)
 {
    GLuint attr;
@@ -184,9 +184,10 @@ st_prepare_vertex_program(struct st_context *st,
    stvp->num_outputs = 0;
 
    if (stvp->Base.IsPositionInvariant)
-      _mesa_insert_mvp_code(st->ctx, &stvp->Base);
+      _mesa_insert_mvp_code(ctx, &stvp->Base);
 
-   assert(stvp->Base.Base.NumInstructions > 1);
+   if (!stvp->glsl_to_tgsi)
+      assert(stvp->Base.Base.NumInstructions > 1);
 
    /*
     * Determine number of inputs, the mappings between VERT_ATTRIB_x
@@ -292,10 +293,13 @@ st_translate_vertex_program(struct st_context *st,
    enum pipe_error error;
    unsigned num_outputs;
 
-   st_prepare_vertex_program( st, stvp );
+   st_prepare_vertex_program(st->ctx, stvp);
 
-   _mesa_remove_output_reads(&stvp->Base.Base, PROGRAM_OUTPUT);
-   _mesa_remove_output_reads(&stvp->Base.Base, PROGRAM_VARYING);
+   if (!stvp->glsl_to_tgsi)
+   {
+      _mesa_remove_output_reads(&stvp->Base.Base, PROGRAM_OUTPUT);
+      _mesa_remove_output_reads(&stvp->Base.Base, PROGRAM_VARYING);
+   }
 
    ureg = ureg_create( TGSI_PROCESSOR_VERTEX );
    if (ureg == NULL) {
@@ -318,22 +322,41 @@ st_translate_vertex_program(struct st_context *st,
       debug_printf("\n");
    }
 
-   error = st_translate_mesa_program(st->ctx,
-                                     TGSI_PROCESSOR_VERTEX,
-                                     ureg,
-                                     &stvp->Base.Base,
-                                     /* inputs */
-                                     vpv->num_inputs,
-                                     stvp->input_to_index,
-                                     NULL, /* input semantic name */
-                                     NULL, /* input semantic index */
-                                     NULL,
-                                     /* outputs */
-                                     num_outputs,
-                                     stvp->result_to_output,
-                                     stvp->output_semantic_name,
-                                     stvp->output_semantic_index,
-                                     key->passthrough_edgeflags );
+   if (stvp->glsl_to_tgsi)
+      error = st_translate_program(st->ctx,
+                                   TGSI_PROCESSOR_VERTEX,
+                                   ureg,
+                                   stvp->glsl_to_tgsi,
+                                   &stvp->Base.Base,
+                                   /* inputs */
+                                   stvp->num_inputs,
+                                   stvp->input_to_index,
+                                   NULL, /* input semantic name */
+                                   NULL, /* input semantic index */
+                                   NULL, /* interp mode */
+                                   /* outputs */
+                                   stvp->num_outputs,
+                                   stvp->result_to_output,
+                                   stvp->output_semantic_name,
+                                   stvp->output_semantic_index,
+                                   key->passthrough_edgeflags );
+   else
+      error = st_translate_mesa_program(st->ctx,
+                                        TGSI_PROCESSOR_VERTEX,
+                                        ureg,
+                                        &stvp->Base.Base,
+                                        /* inputs */
+                                        vpv->num_inputs,
+                                        stvp->input_to_index,
+                                        NULL, /* input semantic name */
+                                        NULL, /* input semantic index */
+                                        NULL,
+                                        /* outputs */
+                                        num_outputs,
+                                        stvp->result_to_output,
+                                        stvp->output_semantic_name,
+                                        stvp->output_semantic_index,
+                                        key->passthrough_edgeflags );
 
    if (error)
       goto fail;
@@ -393,6 +416,151 @@ st_get_vp_variant(struct st_context *st,
    return vpv;
 }
 
+/**
+ * Translate Mesa fragment shader attributes to TGSI attributes.
+ * \return GL_TRUE if color output should be written to all render targets, 
+ *         GL_FALSE if not
+ */
+GLboolean
+st_prepare_fragment_program(struct gl_context *ctx,
+                            struct st_fragment_program *stfp)
+{
+   GLuint attr;
+   const GLbitfield inputsRead = stfp->Base.Base.InputsRead;
+   GLboolean write_all = GL_FALSE;
+
+   /*
+    * Convert Mesa program inputs to TGSI input register semantics.
+    */
+   for (attr = 0; attr < FRAG_ATTRIB_MAX; attr++) {
+      if (inputsRead & (1 << attr)) {
+         const GLuint slot = stfp->num_inputs++;
+
+         stfp->input_to_index[attr] = slot;
+
+         switch (attr) {
+         case FRAG_ATTRIB_WPOS:
+            stfp->input_semantic_name[slot] = TGSI_SEMANTIC_POSITION;
+            stfp->input_semantic_index[slot] = 0;
+            stfp->interp_mode[slot] = TGSI_INTERPOLATE_LINEAR;
+            break;
+         case FRAG_ATTRIB_COL0:
+            stfp->input_semantic_name[slot] = TGSI_SEMANTIC_COLOR;
+            stfp->input_semantic_index[slot] = 0;
+            stfp->interp_mode[slot] = TGSI_INTERPOLATE_LINEAR;
+            break;
+         case FRAG_ATTRIB_COL1:
+            stfp->input_semantic_name[slot] = TGSI_SEMANTIC_COLOR;
+            stfp->input_semantic_index[slot] = 1;
+            stfp->interp_mode[slot] = TGSI_INTERPOLATE_LINEAR;
+            break;
+         case FRAG_ATTRIB_FOGC:
+            stfp->input_semantic_name[slot] = TGSI_SEMANTIC_FOG;
+            stfp->input_semantic_index[slot] = 0;
+            stfp->interp_mode[slot] = TGSI_INTERPOLATE_PERSPECTIVE;
+            break;
+         case FRAG_ATTRIB_FACE:
+            stfp->input_semantic_name[slot] = TGSI_SEMANTIC_FACE;
+            stfp->input_semantic_index[slot] = 0;
+            stfp->interp_mode[slot] = TGSI_INTERPOLATE_CONSTANT;
+            break;
+            /* In most cases, there is nothing special about these
+             * inputs, so adopt a convention to use the generic
+             * semantic name and the mesa FRAG_ATTRIB_ number as the
+             * index. 
+             * 
+             * All that is required is that the vertex shader labels
+             * its own outputs similarly, and that the vertex shader
+             * generates at least every output required by the
+             * fragment shader plus fixed-function hardware (such as
+             * BFC).
+             * 
+             * There is no requirement that semantic indexes start at
+             * zero or be restricted to a particular range -- nobody
+             * should be building tables based on semantic index.
+             */
+         case FRAG_ATTRIB_PNTC:
+         case FRAG_ATTRIB_TEX0:
+         case FRAG_ATTRIB_TEX1:
+         case FRAG_ATTRIB_TEX2:
+         case FRAG_ATTRIB_TEX3:
+         case FRAG_ATTRIB_TEX4:
+         case FRAG_ATTRIB_TEX5:
+         case FRAG_ATTRIB_TEX6:
+         case FRAG_ATTRIB_TEX7:
+         case FRAG_ATTRIB_VAR0:
+         default:
+            /* Actually, let's try and zero-base this just for
+             * readability of the generated TGSI.
+             */
+            assert(attr >= FRAG_ATTRIB_TEX0);
+            stfp->input_semantic_index[slot] = (attr - FRAG_ATTRIB_TEX0);
+            stfp->input_semantic_name[slot] = TGSI_SEMANTIC_GENERIC;
+            if (attr == FRAG_ATTRIB_PNTC)
+               stfp->interp_mode[slot] = TGSI_INTERPOLATE_LINEAR;
+            else
+               stfp->interp_mode[slot] = TGSI_INTERPOLATE_PERSPECTIVE;
+            break;
+         }
+      }
+      else {
+         stfp->input_to_index[attr] = -1;
+      }
+   }
+
+   /*
+    * Semantics and mapping for outputs
+    */
+   {
+      uint numColors = 0;
+      GLbitfield64 outputsWritten = stfp->Base.Base.OutputsWritten;
+
+      /* if z is written, emit that first */
+      if (outputsWritten & BITFIELD64_BIT(FRAG_RESULT_DEPTH)) {
+         stfp->output_semantic_name[stfp->num_outputs] = TGSI_SEMANTIC_POSITION;
+         stfp->output_semantic_index[stfp->num_outputs] = 0;
+         stfp->result_to_output[FRAG_RESULT_DEPTH] = stfp->num_outputs;
+         stfp->num_outputs++;
+         outputsWritten &= ~(1 << FRAG_RESULT_DEPTH);
+      }
+
+      if (outputsWritten & BITFIELD64_BIT(FRAG_RESULT_STENCIL)) {
+         stfp->output_semantic_name[stfp->num_outputs] = TGSI_SEMANTIC_STENCIL;
+         stfp->output_semantic_index[stfp->num_outputs] = 0;
+         stfp->result_to_output[FRAG_RESULT_STENCIL] = stfp->num_outputs;
+         stfp->num_outputs++;
+         outputsWritten &= ~(1 << FRAG_RESULT_STENCIL);
+      }
+
+      /* handle remaning outputs (color) */
+      for (attr = 0; attr < FRAG_RESULT_MAX; attr++) {
+         if (outputsWritten & BITFIELD64_BIT(attr)) {
+            switch (attr) {
+            case FRAG_RESULT_DEPTH:
+            case FRAG_RESULT_STENCIL:
+               /* handled above */
+               assert(0);
+               break;
+            case FRAG_RESULT_COLOR:
+               write_all = GL_TRUE; /* fallthrough */
+            default:
+               assert(attr == FRAG_RESULT_COLOR ||
+                      (FRAG_RESULT_DATA0 <= attr && attr < FRAG_RESULT_MAX));
+               stfp->output_semantic_name[stfp->num_outputs] = TGSI_SEMANTIC_COLOR;
+               stfp->output_semantic_index[stfp->num_outputs] = numColors;
+               stfp->result_to_output[attr] = stfp->num_outputs;
+               numColors++;
+               break;
+            }
+
+            stfp->num_outputs++;
+         }
+      }
+   }
+   
+   return write_all;
+}
+
 
 /**
  * Translate a Mesa fragment shader into a TGSI shader using extra info in
@@ -445,153 +613,11 @@ st_translate_fragment_program(struct st_context *st,
 
    if (!stfp->tgsi.tokens) {
       /* need to translate Mesa instructions to TGSI now */
-      GLuint outputMapping[FRAG_RESULT_MAX];
-      GLuint inputMapping[FRAG_ATTRIB_MAX];
-      GLuint interpMode[PIPE_MAX_SHADER_INPUTS];  /* XXX size? */
-      GLuint attr;
-      const GLbitfield inputsRead = stfp->Base.Base.InputsRead;
       struct ureg_program *ureg;
-      GLboolean write_all = GL_FALSE;
-
-      ubyte input_semantic_name[PIPE_MAX_SHADER_INPUTS];
-      ubyte input_semantic_index[PIPE_MAX_SHADER_INPUTS];
-      uint fs_num_inputs = 0;
-
-      ubyte fs_output_semantic_name[PIPE_MAX_SHADER_OUTPUTS];
-      ubyte fs_output_semantic_index[PIPE_MAX_SHADER_OUTPUTS];
-      uint fs_num_outputs = 0;
-
-
-      _mesa_remove_output_reads(&stfp->Base.Base, PROGRAM_OUTPUT);
-
-      /*
-       * Convert Mesa program inputs to TGSI input register semantics.
-       */
-      for (attr = 0; attr < FRAG_ATTRIB_MAX; attr++) {
-         if (inputsRead & (1 << attr)) {
-            const GLuint slot = fs_num_inputs++;
-
-            inputMapping[attr] = slot;
-
-            switch (attr) {
-            case FRAG_ATTRIB_WPOS:
-               input_semantic_name[slot] = TGSI_SEMANTIC_POSITION;
-               input_semantic_index[slot] = 0;
-               interpMode[slot] = TGSI_INTERPOLATE_LINEAR;
-               break;
-            case FRAG_ATTRIB_COL0:
-               input_semantic_name[slot] = TGSI_SEMANTIC_COLOR;
-               input_semantic_index[slot] = 0;
-               interpMode[slot] = TGSI_INTERPOLATE_LINEAR;
-               break;
-            case FRAG_ATTRIB_COL1:
-               input_semantic_name[slot] = TGSI_SEMANTIC_COLOR;
-               input_semantic_index[slot] = 1;
-               interpMode[slot] = TGSI_INTERPOLATE_LINEAR;
-               break;
-            case FRAG_ATTRIB_FOGC:
-               input_semantic_name[slot] = TGSI_SEMANTIC_FOG;
-               input_semantic_index[slot] = 0;
-               interpMode[slot] = TGSI_INTERPOLATE_PERSPECTIVE;
-               break;
-            case FRAG_ATTRIB_FACE:
-               input_semantic_name[slot] = TGSI_SEMANTIC_FACE;
-               input_semantic_index[slot] = 0;
-               interpMode[slot] = TGSI_INTERPOLATE_CONSTANT;
-               break;
-               /* In most cases, there is nothing special about these
-                * inputs, so adopt a convention to use the generic
-                * semantic name and the mesa FRAG_ATTRIB_ number as the
-                * index. 
-                * 
-                * All that is required is that the vertex shader labels
-                * its own outputs similarly, and that the vertex shader
-                * generates at least every output required by the
-                * fragment shader plus fixed-function hardware (such as
-                * BFC).
-                * 
-                * There is no requirement that semantic indexes start at
-                * zero or be restricted to a particular range -- nobody
-                * should be building tables based on semantic index.
-                */
-            case FRAG_ATTRIB_PNTC:
-            case FRAG_ATTRIB_TEX0:
-            case FRAG_ATTRIB_TEX1:
-            case FRAG_ATTRIB_TEX2:
-            case FRAG_ATTRIB_TEX3:
-            case FRAG_ATTRIB_TEX4:
-            case FRAG_ATTRIB_TEX5:
-            case FRAG_ATTRIB_TEX6:
-            case FRAG_ATTRIB_TEX7:
-            case FRAG_ATTRIB_VAR0:
-            default:
-               /* Actually, let's try and zero-base this just for
-                * readability of the generated TGSI.
-                */
-               assert(attr >= FRAG_ATTRIB_TEX0);
-               input_semantic_index[slot] = (attr - FRAG_ATTRIB_TEX0);
-               input_semantic_name[slot] = TGSI_SEMANTIC_GENERIC;
-               if (attr == FRAG_ATTRIB_PNTC)
-                  interpMode[slot] = TGSI_INTERPOLATE_LINEAR;
-               else
-                  interpMode[slot] = TGSI_INTERPOLATE_PERSPECTIVE;
-               break;
-            }
-         }
-         else {
-            inputMapping[attr] = -1;
-         }
-      }
-
-      /*
-       * Semantics and mapping for outputs
-       */
-      {
-         uint numColors = 0;
-         GLbitfield64 outputsWritten = stfp->Base.Base.OutputsWritten;
-
-         /* if z is written, emit that first */
-         if (outputsWritten & BITFIELD64_BIT(FRAG_RESULT_DEPTH)) {
-            fs_output_semantic_name[fs_num_outputs] = TGSI_SEMANTIC_POSITION;
-            fs_output_semantic_index[fs_num_outputs] = 0;
-            outputMapping[FRAG_RESULT_DEPTH] = fs_num_outputs;
-            fs_num_outputs++;
-            outputsWritten &= ~(1 << FRAG_RESULT_DEPTH);
-         }
-
-         if (outputsWritten & BITFIELD64_BIT(FRAG_RESULT_STENCIL)) {
-            fs_output_semantic_name[fs_num_outputs] = TGSI_SEMANTIC_STENCIL;
-            fs_output_semantic_index[fs_num_outputs] = 0;
-            outputMapping[FRAG_RESULT_STENCIL] = fs_num_outputs;
-            fs_num_outputs++;
-            outputsWritten &= ~(1 << FRAG_RESULT_STENCIL);
-         }
-
-         /* handle remaning outputs (color) */
-         for (attr = 0; attr < FRAG_RESULT_MAX; attr++) {
-            if (outputsWritten & BITFIELD64_BIT(attr)) {
-               switch (attr) {
-               case FRAG_RESULT_DEPTH:
-               case FRAG_RESULT_STENCIL:
-                  /* handled above */
-                  assert(0);
-                  break;
-               case FRAG_RESULT_COLOR:
-                  write_all = GL_TRUE; /* fallthrough */
-               default:
-                  assert(attr == FRAG_RESULT_COLOR ||
-                         (FRAG_RESULT_DATA0 <= attr && attr < FRAG_RESULT_MAX));
-                  fs_output_semantic_name[fs_num_outputs] = TGSI_SEMANTIC_COLOR;
-                  fs_output_semantic_index[fs_num_outputs] = numColors;
-                  outputMapping[attr] = fs_num_outputs;
-                  numColors++;
-                  break;
-               }
-
-               fs_num_outputs++;
-            }
-         }
-      }
+      GLboolean write_all = st_prepare_fragment_program(st->ctx, stfp);
+      
+      if (!stfp->glsl_to_tgsi)
+         _mesa_remove_output_reads(&stfp->Base.Base, PROGRAM_OUTPUT);
 
       ureg = ureg_create( TGSI_PROCESSOR_FRAGMENT );
       if (ureg == NULL)
@@ -605,21 +631,39 @@ st_translate_fragment_program(struct st_context *st,
       if (write_all == GL_TRUE)
          ureg_property_fs_color0_writes_all_cbufs(ureg, 1);
 
-      st_translate_mesa_program(st->ctx,
-                                TGSI_PROCESSOR_FRAGMENT,
-                                ureg,
-                                &stfp->Base.Base,
-                                /* inputs */
-                                fs_num_inputs,
-                                inputMapping,
-                                input_semantic_name,
-                                input_semantic_index,
-                                interpMode,
-                                /* outputs */
-                                fs_num_outputs,
-                                outputMapping,
-                                fs_output_semantic_name,
-                                fs_output_semantic_index, FALSE );
+      if (stfp->glsl_to_tgsi)
+         st_translate_program(st->ctx,
+                              TGSI_PROCESSOR_FRAGMENT,
+                              ureg,
+                              stfp->glsl_to_tgsi,
+                              &stfp->Base.Base,
+                              /* inputs */
+                              stfp->num_inputs,
+                              stfp->input_to_index,
+                              stfp->input_semantic_name,
+                              stfp->input_semantic_index,
+                              stfp->interp_mode,
+                              /* outputs */
+                              stfp->num_outputs,
+                              stfp->result_to_output,
+                              stfp->output_semantic_name,
+                              stfp->output_semantic_index, FALSE );
+      else
+         st_translate_mesa_program(st->ctx,
+                                   TGSI_PROCESSOR_FRAGMENT,
+                                   ureg,
+                                   &stfp->Base.Base,
+                                   /* inputs */
+                                   stfp->num_inputs,
+                                   stfp->input_to_index,
+                                   stfp->input_semantic_name,
+                                   stfp->input_semantic_index,
+                                   stfp->interp_mode,
+                                   /* outputs */
+                                   stfp->num_outputs,
+                                   stfp->result_to_output,
+                                   stfp->output_semantic_name,
+                                   stfp->output_semantic_index, FALSE );
 
       stfp->tgsi.tokens = ureg_get_tokens( ureg, NULL );
       ureg_destroy( ureg );
index c4244df939e01a5092bd266d2579fac654225c2d..67723de6d53a52d1959173d3682ef0b044510780 100644 (file)
@@ -38,6 +38,7 @@
 #include "program/program.h"
 #include "pipe/p_state.h"
 #include "st_context.h"
+#include "st_glsl_to_tgsi.h"
 
 
 /** Fragment program variant key */
@@ -83,6 +84,22 @@ struct st_fp_variant
 struct st_fragment_program
 {
    struct gl_fragment_program Base;
+   struct glsl_to_tgsi_visitor* glsl_to_tgsi;
+   
+   /** maps a Mesa FRAG_ATTRIB_x to a packed TGSI input index */
+   GLuint input_to_index[FRAG_ATTRIB_MAX];
+   /** maps a TGSI input index back to a Mesa FRAG_ATTRIB_x */
+   GLuint index_to_input[PIPE_MAX_SHADER_INPUTS];
+   ubyte input_semantic_name[PIPE_MAX_SHADER_INPUTS];
+   ubyte input_semantic_index[PIPE_MAX_SHADER_INPUTS];
+   GLuint num_inputs;
+   GLuint interp_mode[PIPE_MAX_SHADER_INPUTS];  /* XXX size? */
+
+   /** Maps FRAG_RESULT_x to slot */
+   GLuint result_to_output[FRAG_RESULT_MAX];
+   ubyte output_semantic_name[FRAG_RESULT_MAX];
+   ubyte output_semantic_index[FRAG_RESULT_MAX];
+   GLuint num_outputs;
 
    struct pipe_shader_state tgsi;
 
@@ -136,6 +153,7 @@ struct st_vp_variant
 struct st_vertex_program
 {
    struct gl_vertex_program Base;  /**< The Mesa vertex program */
+   struct glsl_to_tgsi_visitor* glsl_to_tgsi;
 
    /** maps a Mesa VERT_ATTRIB_x to a packed TGSI input index */
    GLuint input_to_index[VERT_ATTRIB_MAX];
@@ -184,6 +202,7 @@ struct st_gp_variant
 struct st_geometry_program
 {
    struct gl_geometry_program Base;  /**< The Mesa geometry program */
+   struct glsl_to_tgsi_visitor* glsl_to_tgsi;
 
    /** map GP input back to VP output */
    GLuint input_map[PIPE_MAX_SHADER_INPUTS];
@@ -276,6 +295,14 @@ st_get_gp_variant(struct st_context *st,
                   const struct st_gp_variant_key *key);
 
 
+extern void
+st_prepare_vertex_program(struct gl_context *ctx,
+                          struct st_vertex_program *stvp);
+
+extern GLboolean
+st_prepare_fragment_program(struct gl_context *ctx,
+                            struct st_fragment_program *stfp);
+
 
 extern void
 st_release_vp_variants( struct st_context *st,
index ffe7e256a56163363605440edfae5f7d68169e16..0e857fddcddedb59a5a9b037dd5b6d51ef6baa41 100644 (file)
@@ -396,3 +396,23 @@ st_texture_image_copy(struct pipe_context *pipe,
    }
 }
 
+
+struct pipe_resource *
+st_create_color_map_texture(struct gl_context *ctx)
+{
+   struct st_context *st = st_context(ctx);
+   struct pipe_context *pipe = st->pipe;
+   struct pipe_resource *pt;
+   enum pipe_format format;
+   const uint texSize = 256; /* simple, and usually perfect */
+
+   /* find an RGBA texture format */
+   format = st_choose_format(pipe->screen, GL_RGBA, GL_NONE, GL_NONE,
+                             PIPE_TEXTURE_2D, 0, PIPE_BIND_SAMPLER_VIEW);
+
+   /* create texture for color map/table */
+   pt = st_texture_create(st, PIPE_TEXTURE_2D, format, 0,
+                          texSize, texSize, 1, 1, PIPE_BIND_SAMPLER_VIEW);
+   return pt;
+}
+
index d50c3c9af79a01b743d0f109b1afa58b548bd5d8..b822f47cf9ea322a0989e792fb6d194fbd7ccb2e 100644 (file)
@@ -232,4 +232,8 @@ st_texture_image_copy(struct pipe_context *pipe,
                       struct pipe_resource *src, GLuint srcLevel,
                       GLuint face);
 
+
+extern struct pipe_resource *
+st_create_color_map_texture(struct gl_context *ctx);
+
 #endif