1 files changed, 103 insertions, 123 deletions

diff --git a/src/gallium/drivers/r300/r300_vs.c b/src/gallium/drivers/r300/r300_vs.c
index 60a04bbfeda..bfab9c3b014 100644
--- a/src/gallium/drivers/r300/r300_vs.c
+++ b/src/gallium/drivers/r300/r300_vs.c
@@ -31,11 +31,10 @@
 
 #include "tgsi/tgsi_dump.h"
 #include "tgsi/tgsi_parse.h"
+#include "tgsi/tgsi_ureg.h"
 
 #include "radeon_compiler.h"
 
-#include "util/u_math.h"
-
 /* Convert info about VS output semantics into r300_shader_semantics. */
 static void r300_shader_read_vs_outputs(
     struct tgsi_shader_info* info,
@@ -82,102 +81,49 @@ static void r300_shader_read_vs_outputs(
 
             case TGSI_SEMANTIC_EDGEFLAG:
                 assert(index == 0);
-                fprintf(stderr, "r300 VP: cannot handle edgeflag output\n");
-                assert(0);
+                fprintf(stderr, "r300 VP: cannot handle edgeflag output.\n");
                 break;
+
             default:
-                assert(0);
+                fprintf(stderr, "r300 VP: unknown vertex output semantic: %i.\n",
+                        info->output_semantic_name[i]);
         }
     }
+
+    /* WPOS is a straight copy of POSITION and it's always emitted. */
+    vs_outputs->wpos = i;
 }
 
-static void r300_shader_vap_output_fmt(struct r300_vertex_shader* vs)
+/* This function sets up:
+ * - VAP mapping, which maps VS registers to output semantics and
+ *   at the same time it indicates which attributes are enabled and should
+ *   be rasterized.
+ * - Stream mapping to VS outputs if TCL is not present. */
+static void r300_init_vs_output_mapping(struct r300_vertex_shader* vs)
 {
     struct r300_shader_semantics* vs_outputs = &vs->outputs;
-    uint32_t* hwfmt = vs->hwfmt;
-    int i, gen_count;
+    struct r300_vap_output_state *vap_out = &vs->vap_out;
+    int *stream_loc = vs->stream_loc_notcl;
+    int i, gen_count, tabi = 0;
     boolean any_bcolor_used = vs_outputs->bcolor[0] != ATTR_UNUSED ||
                               vs_outputs->bcolor[1] != ATTR_UNUSED;
 
-    /* Do the actual vertex_info setup.
-     *
-     * vertex_info has four uints of hardware-specific data in it.
-     * vinfo.hwfmt[0] is R300_VAP_VTX_STATE_CNTL
-     * vinfo.hwfmt[1] is R300_VAP_VSM_VTX_ASSM
-     * vinfo.hwfmt[2] is R300_VAP_OUTPUT_VTX_FMT_0
-     * vinfo.hwfmt[3] is R300_VAP_OUTPUT_VTX_FMT_1 */
-
-    hwfmt[0] = 0x5555; /* XXX this is classic Mesa bonghits */
+    vap_out->vap_vtx_state_cntl = 0x5555; /* XXX this is classic Mesa bonghits */
 
     /* Position. */
     if (vs_outputs->pos != ATTR_UNUSED) {
-        hwfmt[1] |= R300_INPUT_CNTL_POS;
-        hwfmt[2] |= R300_VAP_OUTPUT_VTX_FMT_0__POS_PRESENT;
+        vap_out->vap_vsm_vtx_assm |= R300_INPUT_CNTL_POS;
+        vap_out->vap_out_vtx_fmt[0] |= R300_VAP_OUTPUT_VTX_FMT_0__POS_PRESENT;
+
+        stream_loc[tabi++] = 0;
     } else {
         assert(0);
     }
 
     /* Point size. */
     if (vs_outputs->psize != ATTR_UNUSED) {
-        hwfmt[2] |= R300_VAP_OUTPUT_VTX_FMT_0__PT_SIZE_PRESENT;
-    }
-
-    /* Colors. */
-    for (i = 0; i < ATTR_COLOR_COUNT; i++) {
-        if (vs_outputs->color[i] != ATTR_UNUSED || any_bcolor_used ||
-            vs_outputs->color[1] != ATTR_UNUSED) {
-            hwfmt[1] |= R300_INPUT_CNTL_COLOR;
-            hwfmt[2] |= R300_VAP_OUTPUT_VTX_FMT_0__COLOR_0_PRESENT << i;
-        }
-    }
-
-    /* Back-face colors. */
-    if (any_bcolor_used) {
-        for (i = 0; i < ATTR_COLOR_COUNT; i++) {
-            hwfmt[1] |= R300_INPUT_CNTL_COLOR;
-            hwfmt[2] |= R300_VAP_OUTPUT_VTX_FMT_0__COLOR_0_PRESENT << (2+i);
-        }
-    }
+        vap_out->vap_out_vtx_fmt[0] |= R300_VAP_OUTPUT_VTX_FMT_0__PT_SIZE_PRESENT;
 
-    /* Texture coordinates. */
-    gen_count = 0;
-    for (i = 0; i < ATTR_GENERIC_COUNT; i++) {
-        if (vs_outputs->generic[i] != ATTR_UNUSED) {
-            hwfmt[1] |= (R300_INPUT_CNTL_TC0 << gen_count);
-            hwfmt[3] |= (4 << (3 * gen_count));
-            gen_count++;
-        }
-    }
-
-    /* Fog coordinates. */
-    if (vs_outputs->fog != ATTR_UNUSED) {
-        hwfmt[1] |= (R300_INPUT_CNTL_TC0 << gen_count);
-        hwfmt[3] |= (4 << (3 * gen_count));
-        gen_count++;
-    }
-
-    /* XXX magic */
-    assert(gen_count <= 8);
-
-    /* WPOS. */
-    vs->wpos_tex_output = gen_count;
-}
-
-/* Sets up stream mapping to equivalent VS outputs if TCL is bypassed
- * or isn't present. */
-static void r300_stream_locations_notcl(
-    struct r300_shader_semantics* vs_outputs,
-    int* stream_loc)
-{
-    int i, tabi = 0, gen_count;
-    boolean any_bcolor_used = vs_outputs->bcolor[0] != ATTR_UNUSED ||
-                              vs_outputs->bcolor[1] != ATTR_UNUSED;
-
-    /* Position. */
-    stream_loc[tabi++] = 0;
-
-    /* Point size. */
-    if (vs_outputs->psize != ATTR_UNUSED) {
         stream_loc[tabi++] = 1;
     }
 
@@ -185,6 +131,9 @@ static void r300_stream_locations_notcl(
     for (i = 0; i < ATTR_COLOR_COUNT; i++) {
         if (vs_outputs->color[i] != ATTR_UNUSED || any_bcolor_used ||
             vs_outputs->color[1] != ATTR_UNUSED) {
+            vap_out->vap_vsm_vtx_assm |= R300_INPUT_CNTL_COLOR;
+            vap_out->vap_out_vtx_fmt[0] |= R300_VAP_OUTPUT_VTX_FMT_0__COLOR_0_PRESENT << i;
+
             stream_loc[tabi++] = 2 + i;
         }
     }
@@ -192,32 +141,40 @@ static void r300_stream_locations_notcl(
     /* Back-face colors. */
     if (any_bcolor_used) {
         for (i = 0; i < ATTR_COLOR_COUNT; i++) {
+            vap_out->vap_vsm_vtx_assm |= R300_INPUT_CNTL_COLOR;
+            vap_out->vap_out_vtx_fmt[0] |= R300_VAP_OUTPUT_VTX_FMT_0__COLOR_0_PRESENT << (2+i);
+
             stream_loc[tabi++] = 4 + i;
         }
     }
 
     /* Texture coordinates. */
     gen_count = 0;
-    for (i = 0; i < ATTR_GENERIC_COUNT; i++) {
+    for (i = 0; i < ATTR_GENERIC_COUNT && gen_count < 8; i++) {
         if (vs_outputs->generic[i] != ATTR_UNUSED) {
-            assert(tabi < 16);
+            vap_out->vap_vsm_vtx_assm |= (R300_INPUT_CNTL_TC0 << gen_count);
+            vap_out->vap_out_vtx_fmt[1] |= (4 << (3 * gen_count));
+
             stream_loc[tabi++] = 6 + gen_count;
             gen_count++;
         }
     }
 
     /* Fog coordinates. */
-    if (vs_outputs->fog != ATTR_UNUSED) {
-        assert(tabi < 16);
+    if (gen_count < 8 && vs_outputs->fog != ATTR_UNUSED) {
+        vap_out->vap_vsm_vtx_assm |= (R300_INPUT_CNTL_TC0 << gen_count);
+        vap_out->vap_out_vtx_fmt[1] |= (4 << (3 * gen_count));
+
         stream_loc[tabi++] = 6 + gen_count;
         gen_count++;
     }
 
     /* WPOS. */
-    if (vs_outputs->wpos != ATTR_UNUSED) {
-        assert(tabi < 16);
+    if (gen_count < 8) {
+        vs->wpos_tex_output = gen_count;
         stream_loc[tabi++] = 6 + gen_count;
-        gen_count++;
+    } else {
+        vs_outputs->wpos = ATTR_UNUSED;
     }
 
     for (; tabi < 16;) {
@@ -294,36 +251,42 @@ static void set_vertex_inputs_outputs(struct r300_vertex_program_compiler * c)
     }
 }
 
-static void r300_insert_wpos(struct r300_vertex_program_compiler* c,
-                             struct r300_shader_semantics* outputs)
+static void r300_dummy_vertex_shader(
+    struct r300_context* r300,
+    struct r300_vertex_shader* shader)
 {
-    int i, lastOutput = 0;
+    struct pipe_shader_state state;
+    struct ureg_program *ureg;
+    struct ureg_dst dst;
+    struct ureg_src imm;
 
-    /* Find the max output index. */
-    lastOutput = MAX2(lastOutput, outputs->psize);
-    for (i = 0; i < ATTR_COLOR_COUNT; i++) {
-        lastOutput = MAX2(lastOutput, outputs->color[i]);
-        lastOutput = MAX2(lastOutput, outputs->bcolor[i]);
-    }
-    for (i = 0; i < ATTR_GENERIC_COUNT; i++) {
-        lastOutput = MAX2(lastOutput, outputs->generic[i]);
-    }
-    lastOutput = MAX2(lastOutput, outputs->fog);
+    /* Make a simple vertex shader which outputs (0, 0, 0, 1),
+     * effectively rendering nothing. */
+    ureg = ureg_create(TGSI_PROCESSOR_VERTEX);
+    dst = ureg_DECL_output(ureg, TGSI_SEMANTIC_POSITION, 0);
+    imm = ureg_imm4f(ureg, 0, 0, 0, 1);
+
+    ureg_MOV(ureg, dst, imm);
+    ureg_END(ureg);
 
-    /* Set WPOS after the last output. */
-    lastOutput++;
-    rc_copy_output(&c->Base, 0, lastOutput); /* out[lastOutput] = out[0]; */
-    outputs->wpos = lastOutput;
+    state.tokens = ureg_finalize(ureg);
+
+    shader->dummy = TRUE;
+    r300_translate_vertex_shader(r300, shader, state.tokens);
+
+    ureg_destroy(ureg);
 }
 
 void r300_translate_vertex_shader(struct r300_context* r300,
-                                  struct r300_vertex_shader* vs)
+                                  struct r300_vertex_shader* vs,
+                                  const struct tgsi_token *tokens)
 {
     struct r300_vertex_program_compiler compiler;
     struct tgsi_to_rc ttr;
 
-    /* Initialize. */
+    tgsi_scan_shader(tokens, &vs->info);
     r300_shader_read_vs_outputs(&vs->info, &vs->outputs);
+    r300_init_vs_output_mapping(vs);
 
     /* Setup the compiler */
     rc_init(&compiler.Base);
@@ -334,7 +297,7 @@ void r300_translate_vertex_shader(struct r300_context* r300,
 
     if (compiler.Base.Debug) {
         debug_printf("r300: Initial vertex program\n");
-        tgsi_dump(vs->state.tokens, 0);
+        tgsi_dump(tokens, 0);
     }
 
     /* Translate TGSI to our internal representation */
@@ -342,51 +305,68 @@ void r300_translate_vertex_shader(struct r300_context* r300,
     ttr.info = &vs->info;
     ttr.use_half_swizzles = FALSE;
 
-    r300_tgsi_to_rc(&ttr, vs->state.tokens);
+    r300_tgsi_to_rc(&ttr, tokens);
+
+    compiler.RequiredOutputs =
+        ~(~0 << (vs->info.num_outputs +
+                 (vs->outputs.wpos != ATTR_UNUSED ? 1 : 0)));
 
-    compiler.RequiredOutputs = ~(~0 << (vs->info.num_outputs+1));
     compiler.SetHwInputOutput = &set_vertex_inputs_outputs;
 
     /* Insert the WPOS output. */
-    r300_insert_wpos(&compiler, &vs->outputs);
-
-    r300_shader_vap_output_fmt(vs);
-    r300_stream_locations_notcl(&vs->outputs, vs->stream_loc_notcl);
+    if (vs->outputs.wpos != ATTR_UNUSED) {
+        rc_copy_output(&compiler.Base, 0, vs->outputs.wpos);
+    }
 
     /* Invoke the compiler */
     r3xx_compile_vertex_program(&compiler);
     if (compiler.Base.Error) {
         /* XXX We should fallback using Draw. */
-        fprintf(stderr, "r300 VP: Compiler error\n");
-        abort();
+        fprintf(stderr, "r300 VP: Compiler error:\n%sUsing a dummy shader"
+                " instead.\n", compiler.Base.ErrorMsg);
+
+        if (vs->dummy) {
+            fprintf(stderr, "r300 VP: Cannot compile the dummy shader! "
+                    "Giving up...\n");
+            abort();
+        }
+
+        rc_destroy(&compiler.Base);
+        r300_dummy_vertex_shader(r300, vs);
+        return;
     }
 
+    /* Initialize numbers of constants for each type. */
+    vs->externals_count = ttr.immediate_offset;
+    vs->immediates_count = vs->code.constants.Count - vs->externals_count;
+
     /* And, finally... */
     rc_destroy(&compiler.Base);
-    vs->translated = TRUE;
 }
 
 boolean r300_vertex_shader_setup_wpos(struct r300_context* r300)
 {
     struct r300_vertex_shader* vs = r300->vs_state.state;
+    struct r300_vap_output_state *vap_out = &vs->vap_out;
     int tex_output = vs->wpos_tex_output;
     uint32_t tex_fmt = R300_INPUT_CNTL_TC0 << tex_output;
-    uint32_t* hwfmt = vs->hwfmt;
 
-    if (r300->fs->inputs.wpos != ATTR_UNUSED) {
-        /* Enable WPOS in VAP. */
-        if (!(hwfmt[1] & tex_fmt)) {
-            hwfmt[1] |= tex_fmt;
-            hwfmt[3] |= (4 << (3 * tex_output));
+    if (vs->outputs.wpos == ATTR_UNUSED) {
+        return FALSE;
+    }
 
-            assert(tex_output < 8);
+    if (r300_fs(r300)->shader->inputs.wpos != ATTR_UNUSED) {
+        /* Enable WPOS in VAP. */
+        if (!(vap_out->vap_vsm_vtx_assm & tex_fmt)) {
+            vap_out->vap_vsm_vtx_assm |= tex_fmt;
+            vap_out->vap_out_vtx_fmt[1] |= (4 << (3 * tex_output));
             return TRUE;
         }
     } else {
         /* Disable WPOS in VAP. */
-        if (hwfmt[1] & tex_fmt) {
-            hwfmt[1] &= ~tex_fmt;
-            hwfmt[3] &= ~(4 << (3 * tex_output));
+        if (vap_out->vap_vsm_vtx_assm & tex_fmt) {
+            vap_out->vap_vsm_vtx_assm &= ~tex_fmt;
+            vap_out->vap_out_vtx_fmt[1] &= ~(4 << (3 * tex_output));
             return TRUE;
         }
     }