st/nine: Prepare constant compaction in nine_shader

When indirect addressing is not used, we know exactly
which constants are accessed, and thus can
have them located in consecutive slots.
We thus parse again the shader with a slot map
for compaction.

The path contains the work inside nine_shader.c for this
path, but it needs some other commits to work, and thus
is not enabled yet by this commit.

Signed-off-by: Axel Davy <[email protected]>

2 files changed, 118 insertions, 28 deletions

diff --git a/src/gallium/state_trackers/nine/nine_shader.c b/src/gallium/state_trackers/nine/nine_shader.c
index b651f065585..130d63b61dd 100644
--- a/src/gallium/state_trackers/nine/nine_shader.c
+++ b/src/gallium/state_trackers/nine/nine_shader.c
@@ -509,6 +509,7 @@ struct shader_translator
     unsigned num_lconstb;
 
     boolean slots_used[NINE_MAX_CONST_ALL];
+    unsigned *slot_map;
     unsigned num_slots;
 
     boolean indirect_const_access;
@@ -556,6 +557,9 @@ nine_record_outputs(struct shader_translator *tx, BYTE Usage, BYTE UsageIndex,
 static struct ureg_src nine_float_constant_src(struct shader_translator *tx, int idx)
 {
     struct ureg_src src;
+
+    if (tx->slot_map)
+        idx = tx->slot_map[idx];
     /* vswp constant handling: we use two buffers
      * to fit all the float constants. The special handling
      * doesn't need to be elsewhere, because all the instructions
@@ -589,6 +593,8 @@ static struct ureg_src nine_integer_constant_src(struct shader_translator *tx, i
         src = ureg_src_dimension(src, 2);
     } else {
         unsigned slot_idx = tx->info->const_i_base + idx;
+        if (tx->slot_map)
+            slot_idx = tx->slot_map[slot_idx];
         src = ureg_src_register(TGSI_FILE_CONSTANT, slot_idx);
         src = ureg_src_dimension(src, 0);
         tx->slots_used[slot_idx] = TRUE;
@@ -615,6 +621,8 @@ static struct ureg_src nine_boolean_constant_src(struct shader_translator *tx, i
         src = ureg_src_dimension(src, 3);
     } else {
         unsigned slot_idx = tx->info->const_b_base + r;
+        if (tx->slot_map)
+            slot_idx = tx->slot_map[slot_idx];
         src = ureg_src_register(TGSI_FILE_CONSTANT, slot_idx);
         src = ureg_src_dimension(src, 0);
         tx->slots_used[slot_idx] = TRUE;
@@ -3656,6 +3664,8 @@ tx_ctor(struct shader_translator *tx, struct pipe_screen *screen, struct nine_sh
 static void
 tx_dtor(struct shader_translator *tx)
 {
+    if (tx->slot_map)
+        FREE(tx->slot_map);
     if (tx->num_inst_labels)
         FREE(tx->inst_labels);
     FREE(tx->lconstf);
@@ -3736,6 +3746,48 @@ shader_add_ps_fog_stage(struct shader_translator *tx, struct ureg_src src_col)
     ureg_MOV(ureg, ureg_writemask(oCol0, TGSI_WRITEMASK_W), src_col);
 }
 
+static void parse_shader(struct shader_translator *tx)
+{
+    struct nine_shader_info *info = tx->info;
+
+    while (!sm1_parse_eof(tx) && !tx->failure)
+        sm1_parse_instruction(tx);
+    tx->parse++; /* for byte_size */
+
+    if (tx->failure)
+        return;
+
+    if (IS_PS && tx->version.major < 3) {
+        if (tx->version.major < 2) {
+            assert(tx->num_temp); /* there must be color output */
+            info->rt_mask |= 0x1;
+            shader_add_ps_fog_stage(tx, ureg_src(tx->regs.r[0]));
+        } else {
+            shader_add_ps_fog_stage(tx, ureg_src(tx->regs.oCol[0]));
+        }
+    }
+
+    if (IS_VS && tx->version.major < 3 && ureg_dst_is_undef(tx->regs.oFog) && info->fog_enable) {
+        tx->regs.oFog = ureg_DECL_output(tx->ureg, TGSI_SEMANTIC_FOG, 0);
+        ureg_MOV(tx->ureg, ureg_writemask(tx->regs.oFog, TGSI_WRITEMASK_X), ureg_imm1f(tx->ureg, 0.0f));
+    }
+
+    if (info->position_t)
+        ureg_property(tx->ureg, TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION, TRUE);
+
+    if (IS_VS && !ureg_dst_is_undef(tx->regs.oPts)) {
+        struct ureg_dst oPts = ureg_DECL_output(tx->ureg, TGSI_SEMANTIC_PSIZE, 0);
+        ureg_MAX(tx->ureg, tx->regs.oPts, ureg_src(tx->regs.oPts), ureg_imm1f(tx->ureg, info->point_size_min));
+        ureg_MIN(tx->ureg, oPts, ureg_src(tx->regs.oPts), ureg_imm1f(tx->ureg, info->point_size_max));
+        info->point_size = TRUE;
+    }
+
+    if (info->process_vertices)
+        shader_add_vs_viewport_transform(tx);
+
+    ureg_END(tx->ureg);
+}
+
 HRESULT
 nine_translate_shader(struct NineDevice9 *device, struct nine_shader_info *info, struct pipe_context *pipe)
 {
@@ -3743,6 +3795,7 @@ nine_translate_shader(struct NineDevice9 *device, struct nine_shader_info *info,
     HRESULT hr = D3D_OK;
     const unsigned processor = info->type;
     struct pipe_screen *screen = info->process_vertices ? device->screen_sw : device->screen;
+    unsigned *const_ranges = NULL;
 
     user_assert(processor != ~0, D3DERR_INVALIDCALL);
 
@@ -3771,9 +3824,7 @@ nine_translate_shader(struct NineDevice9 *device, struct nine_shader_info *info,
     DUMP("%s%u.%u\n", processor == PIPE_SHADER_VERTEX ? "VS" : "PS",
          tx->version.major, tx->version.minor);
 
-    while (!sm1_parse_eof(tx) && !tx->failure)
-        sm1_parse_instruction(tx);
-    tx->parse++; /* for byte_size */
+    parse_shader(tx);
 
     if (tx->failure) {
         /* For VS shaders, we print the warning later,
@@ -3785,36 +3836,69 @@ nine_translate_shader(struct NineDevice9 *device, struct nine_shader_info *info,
         goto out;
     }
 
-    if (IS_PS && tx->version.major < 3) {
-        if (tx->version.major < 2) {
-            assert(tx->num_temp); /* there must be color output */
-            info->rt_mask |= 0x1;
-            shader_add_ps_fog_stage(tx, ureg_src(tx->regs.r[0]));
-        } else {
-            shader_add_ps_fog_stage(tx, ureg_src(tx->regs.oCol[0]));
-        }
-    }
+    /* Recompile after compacting constant slots if possible */
+    if (!tx->indirect_const_access && !info->swvp_on && tx->num_slots > 0 && 0) {
+        unsigned *slot_map;
+        unsigned c;
+        int i, j, num_ranges, prev;
 
-    if (IS_VS && tx->version.major < 3 && ureg_dst_is_undef(tx->regs.oFog) && info->fog_enable) {
-        tx->regs.oFog = ureg_DECL_output(tx->ureg, TGSI_SEMANTIC_FOG, 0);
-        ureg_MOV(tx->ureg, ureg_writemask(tx->regs.oFog, TGSI_WRITEMASK_X), ureg_imm1f(tx->ureg, 0.0f));
-    }
+        DBG("Recompiling shader for constant compaction\n");
+        ureg_destroy(tx->ureg);
 
-    if (info->position_t)
-        ureg_property(tx->ureg, TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION, TRUE);
+        if (tx->num_inst_labels)
+            FREE(tx->inst_labels);
+        FREE(tx->lconstf);
+        FREE(tx->regs.r);
+
+        num_ranges = 0;
+        prev = -2;
+        for (i = 0; i < NINE_MAX_CONST_ALL; i++) {
+            if (tx->slots_used[i]) {
+                if (prev != i - 1)
+                    num_ranges++;
+                prev = i;
+            }
+        }
+        slot_map = MALLOC(NINE_MAX_CONST_ALL * sizeof(unsigned));
+        const_ranges = CALLOC(num_ranges + 1, 2 * sizeof(unsigned)); /* ranges stop when last is of size 0 */
+        if (!slot_map || !const_ranges) {
+            hr = E_OUTOFMEMORY;
+            goto out;
+        }
+        c = 0;
+        j = -1;
+        prev = -2;
+        for (i = 0; i < NINE_MAX_CONST_ALL; i++) {
+            if (tx->slots_used[i]) {
+                if (prev != i - 1)
+                    j++;
+                /* Initialize first slot of the range */
+                if (!const_ranges[2*j+1])
+                    const_ranges[2*j] = i;
+                const_ranges[2*j+1]++;
+                prev = i;
+                slot_map[i] = c++;
+            }
+        }
 
-    if (IS_VS && !ureg_dst_is_undef(tx->regs.oPts)) {
-        struct ureg_dst oPts = ureg_DECL_output(tx->ureg, TGSI_SEMANTIC_PSIZE, 0);
-        ureg_MAX(tx->ureg, tx->regs.oPts, ureg_src(tx->regs.oPts), ureg_imm1f(tx->ureg, info->point_size_min));
-        ureg_MIN(tx->ureg, oPts, ureg_src(tx->regs.oPts), ureg_imm1f(tx->ureg, info->point_size_max));
-        info->point_size = TRUE;
+        if (tx_ctor(tx, screen, info) == E_OUTOFMEMORY) {
+            hr = E_OUTOFMEMORY;
+            goto out;
+        }
+        tx->slot_map = slot_map;
+        parse_shader(tx);
+        assert(!tx->failure);
+#if !defined(NDEBUG)
+        i = 0;
+        j = 0;
+        while (const_ranges[i*2+1] != 0) {
+            j += const_ranges[i*2+1];
+            i++;
+        }
+        assert(j == tx->num_slots);
+#endif
     }
 
-    if (info->process_vertices)
-        shader_add_vs_viewport_transform(tx);
-
-    ureg_END(tx->ureg);
-
     /* record local constants */
     if (tx->num_lconstf && tx->indirect_const_access) {
         struct nine_range *ranges;
@@ -3920,8 +4004,12 @@ nine_translate_shader(struct NineDevice9 *device, struct nine_shader_info *info,
         goto out;
     }
 
+    info->const_ranges = const_ranges;
+    const_ranges = NULL;
     info->byte_size = (tx->parse - tx->byte_code) * sizeof(DWORD);
 out:
+    if (const_ranges)
+        FREE(const_ranges);
     tx_dtor(tx);
     return hr;
 }
diff --git a/src/gallium/state_trackers/nine/nine_shader.h b/src/gallium/state_trackers/nine/nine_shader.h
index a2eb91d0d8c..6eb97127b9c 100644
--- a/src/gallium/state_trackers/nine/nine_shader.h
+++ b/src/gallium/state_trackers/nine/nine_shader.h
@@ -82,6 +82,8 @@ struct nine_shader_info
     unsigned const_int_slots;
     unsigned const_bool_slots;
 
+    unsigned *const_ranges;
+
     struct nine_lconstf lconstf; /* out, NOTE: members to be free'd by user */
     uint8_t bumpenvmat_needed;