vkd3d-shader: Add analysis of indexableTemps.

Try to reduce component count as much as possible.

Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>

3 files changed, 64 insertions, 2 deletions

diff --git a/libs/vkd3d-shader/spirv.c b/libs/vkd3d-shader/spirv.c
index 0fa1e318..eccf6b2e 100644
--- a/libs/vkd3d-shader/spirv.c
+++ b/libs/vkd3d-shader/spirv.c
@@ -6075,6 +6075,7 @@ static void vkd3d_dxbc_compiler_emit_dcl_indexable_temp(struct vkd3d_dxbc_compil
     struct vkd3d_spirv_builder *builder = &compiler->spirv_builder;
     struct vkd3d_shader_register reg;
     struct vkd3d_symbol reg_symbol;
+    unsigned int component_count;
     size_t function_location;
     uint32_t id;
 
@@ -6088,11 +6089,13 @@ static void vkd3d_dxbc_compiler_emit_dcl_indexable_temp(struct vkd3d_dxbc_compil
     reg.idx[0].offset = temp->register_idx;
     reg.idx[1].offset = ~0u;
 
+    component_count = vkd3d_shader_scan_get_idxtemp_components(compiler->scan_info, &reg);
+
     function_location = vkd3d_dxbc_compiler_get_current_function_location(compiler);
     vkd3d_spirv_begin_function_stream_insertion(builder, function_location);
 
     id = vkd3d_dxbc_compiler_emit_array_variable(compiler, &builder->function_stream,
-            SpvStorageClassFunction, VKD3D_TYPE_FLOAT, VKD3D_VEC4_SIZE, temp->register_size);
+            SpvStorageClassFunction, VKD3D_TYPE_FLOAT, component_count, temp->register_size);
 
     vkd3d_dxbc_compiler_emit_register_debug_name(builder, id, &reg);
 
@@ -6100,7 +6103,8 @@ static void vkd3d_dxbc_compiler_emit_dcl_indexable_temp(struct vkd3d_dxbc_compil
 
     vkd3d_symbol_make_register(&reg_symbol, &reg);
     vkd3d_symbol_set_register_info(&reg_symbol, id,
-            SpvStorageClassFunction, VKD3D_TYPE_FLOAT, VKD3DSP_WRITEMASK_ALL);
+            SpvStorageClassFunction, VKD3D_TYPE_FLOAT,
+            vkd3d_write_mask_from_component_count(component_count));
     reg_symbol.info.reg.indexable_count = temp->register_size;
     vkd3d_dxbc_compiler_put_symbol(compiler, &reg_symbol);
 }
diff --git a/libs/vkd3d-shader/vkd3d_shader_main.c b/libs/vkd3d-shader/vkd3d_shader_main.c
index b8192e49..2b4e26ca 100644
--- a/libs/vkd3d-shader/vkd3d_shader_main.c
+++ b/libs/vkd3d-shader/vkd3d_shader_main.c
@@ -239,6 +239,7 @@ struct vkd3d_shader_scan_entry
     struct hash_map_entry entry;
     struct vkd3d_shader_scan_key key;
     unsigned int flags;
+    unsigned required_components;
 };
 
 static uint32_t vkd3d_shader_scan_entry_hash(const void *key)
@@ -267,6 +268,19 @@ unsigned int vkd3d_shader_scan_get_register_flags(const struct vkd3d_shader_scan
     return e ? e->flags : 0u;
 }
 
+unsigned int vkd3d_shader_scan_get_idxtemp_components(const struct vkd3d_shader_scan_info *scan_info,
+        const struct vkd3d_shader_register *reg)
+{
+    const struct vkd3d_shader_scan_entry *e;
+    struct vkd3d_shader_scan_key key;
+
+    key.register_type = reg->type;
+    key.register_id = reg->idx[0].offset;
+
+    e = (const struct vkd3d_shader_scan_entry *)hash_map_find(&scan_info->register_map, &key);
+    return e ? e->required_components : 4u;
+}
+
 static void vkd3d_shader_scan_set_register_flags(struct vkd3d_shader_scan_info *scan_info,
         enum vkd3d_shader_register_type type, unsigned int id, unsigned int flags)
 {
@@ -285,6 +299,30 @@ static void vkd3d_shader_scan_set_register_flags(struct vkd3d_shader_scan_info *
     {
         entry.key = key;
         entry.flags = flags;
+        entry.required_components = 0;
+        hash_map_insert(&scan_info->register_map, &key, &entry.entry);
+    }
+}
+
+static void vkd3d_shader_scan_record_idxtemp_components(struct vkd3d_shader_scan_info *scan_info,
+        const struct vkd3d_shader_register *reg, unsigned int required_components)
+{
+    struct vkd3d_shader_scan_entry entry;
+    struct vkd3d_shader_scan_entry *e;
+    struct vkd3d_shader_scan_key key;
+
+    key.register_type = reg->type;
+    key.register_id = reg->idx[0].offset;
+
+    if ((e = (struct vkd3d_shader_scan_entry *)hash_map_find(&scan_info->register_map, &key)))
+    {
+        e->required_components = max(required_components, e->required_components);
+    }
+    else
+    {
+        entry.key = key;
+        entry.flags = 0;
+        entry.required_components = required_components;
         hash_map_insert(&scan_info->register_map, &key, &entry.entry);
     }
 }
@@ -556,6 +594,24 @@ static void vkd3d_shader_scan_instruction(struct vkd3d_shader_scan_info *scan_in
             break;
     }
 
+    /* If we do nothing, we will have to assume that IDXTEMP is an array of vec4.
+     * This is problematic for performance if shader only accesses the first 1, 2 or 3 components.
+     * The dcl_indexableTemp instruction specifies number of components but FXC does not seem to
+     * care, so we have to analyze write masks instead. */
+    for (i = 0; i < instruction->dst_count; ++i)
+    {
+        if (instruction->dst[i].reg.type == VKD3DSPR_IDXTEMP)
+        {
+            unsigned int write_mask, required_components;
+            write_mask = instruction->dst[i].write_mask;
+            write_mask |= write_mask >> 2;
+            write_mask |= write_mask >> 1;
+            required_components = vkd3d_write_mask_component_count(write_mask);
+            vkd3d_shader_scan_record_idxtemp_components(scan_info,
+                    &instruction->dst[i].reg, required_components);
+        }
+    }
+
     if (vkd3d_shader_instruction_is_uav_read(instruction))
     {
         is_atomic = vkd3d_shader_instruction_is_uav_atomic(instruction);
diff --git a/libs/vkd3d-shader/vkd3d_shader_private.h b/libs/vkd3d-shader/vkd3d_shader_private.h
index c79f355a..279d7655 100644
--- a/libs/vkd3d-shader/vkd3d_shader_private.h
+++ b/libs/vkd3d-shader/vkd3d_shader_private.h
@@ -912,6 +912,8 @@ static inline unsigned int vkd3d_compact_swizzle(unsigned int swizzle, unsigned
 
 unsigned int vkd3d_shader_scan_get_register_flags(const struct vkd3d_shader_scan_info *scan_info,
         enum vkd3d_shader_register_type type, unsigned int id);
+unsigned int vkd3d_shader_scan_get_idxtemp_components(const struct vkd3d_shader_scan_info *scan_info,
+        const struct vkd3d_shader_register *reg);
 
 /* DXIL support */
 int vkd3d_shader_compile_dxil(const struct vkd3d_shader_code *dxbc,