vkd3d-shader/hlsl: Allocate temporary registers separately for each entry function.

Approved-by: Elizabeth Figura (@zfigura)
Approved-by: Henri Verbeet (@hverbeet)
Merge-Request: https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/1154

libs/vkd3d-shader/hlsl.h

@@ -1089,9 +1089,6 @@ struct hlsl_ctx
     } constant_defs;
     /* 'c' registers where the constants expected by SM2 sincos are stored. */
     struct hlsl_reg d3dsincosconst1, d3dsincosconst2;
-    /* Number of temp. registers required for the shader to run, i.e. the largest temp register
-     *   index that will be used in the output bytecode (+1). */
-    uint32_t temp_count;
 
     /* Number of threads to be executed (on the X, Y, and Z dimensions) in a single thread group in
      *   compute shader profiles. It is set using the numthreads() attribute in the entry point. */
@@ -1430,6 +1427,8 @@ struct hlsl_state_block_entry *clone_stateblock_entry(struct hlsl_ctx *ctx,
 
 void hlsl_lower_index_loads(struct hlsl_ctx *ctx, struct hlsl_block *body);
 void hlsl_run_const_passes(struct hlsl_ctx *ctx, struct hlsl_block *body);
+uint32_t allocate_temp_registers(struct hlsl_ctx *ctx, struct hlsl_ir_function_decl *entry_func);
+void mark_indexable_vars(struct hlsl_ctx *ctx, struct hlsl_ir_function_decl *entry_func);
 int hlsl_emit_bytecode(struct hlsl_ctx *ctx, struct hlsl_ir_function_decl *entry_func,
         enum vkd3d_shader_target_type target_type, struct vkd3d_shader_code *out);
 int hlsl_emit_effect_binary(struct hlsl_ctx *ctx, struct vkd3d_shader_code *out);

libs/vkd3d-shader/hlsl_codegen.c

@@ -4114,7 +4114,7 @@ static void dump_function(struct rb_entry *entry, void *context)
     }
 }
 
-static bool mark_indexable_vars(struct hlsl_ctx *ctx, struct hlsl_deref *deref,
+static bool mark_indexable_var(struct hlsl_ctx *ctx, struct hlsl_deref *deref,
         struct hlsl_ir_node *instr)
 {
     if (!deref->rel_offset.node)
@@ -4127,6 +4127,20 @@ static bool mark_indexable_vars(struct hlsl_ctx *ctx, struct hlsl_deref *deref,
     return true;
 }
 
+void mark_indexable_vars(struct hlsl_ctx *ctx, struct hlsl_ir_function_decl *entry_func)
+{
+    struct hlsl_scope *scope;
+    struct hlsl_ir_var *var;
+
+    LIST_FOR_EACH_ENTRY(scope, &ctx->scopes, struct hlsl_scope, entry)
+    {
+        LIST_FOR_EACH_ENTRY(var, &scope->vars, struct hlsl_ir_var, scope_entry)
+            var->indexable = false;
+    }
+
+    transform_derefs(ctx, mark_indexable_var, &entry_func->body);
+}
+
 static char get_regset_name(enum hlsl_regset regset)
 {
     switch (regset)
@@ -5116,9 +5130,21 @@ static void allocate_const_registers(struct hlsl_ctx *ctx, struct hlsl_ir_functi
  * index to all (simultaneously live) variables or intermediate values. Agnostic
  * as to how many registers are actually available for the current backend, and
  * does not handle constants. */
-static void allocate_temp_registers(struct hlsl_ctx *ctx, struct hlsl_ir_function_decl *entry_func)
+uint32_t allocate_temp_registers(struct hlsl_ctx *ctx, struct hlsl_ir_function_decl *entry_func)
 {
     struct register_allocator allocator = {0};
+    struct hlsl_scope *scope;
+    struct hlsl_ir_var *var;
+
+    /* Reset variable temp register allocations. */
+    LIST_FOR_EACH_ENTRY(scope, &ctx->scopes, struct hlsl_scope, entry)
+    {
+        LIST_FOR_EACH_ENTRY(var, &scope->vars, struct hlsl_ir_var, scope_entry)
+        {
+            if (!(var->is_input_semantic || var->is_output_semantic || var->is_uniform))
+                memset(var->regs, 0, sizeof(var->regs));
+        }
+    }
 
     /* ps_1_* outputs are special and go in temp register 0. */
     if (ctx->profile->major_version == 1 && ctx->profile->type == VKD3D_SHADER_TYPE_PIXEL)
@@ -5127,8 +5153,7 @@ static void allocate_temp_registers(struct hlsl_ctx *ctx, struct hlsl_ir_functio
 
         for (i = 0; i < entry_func->parameters.count; ++i)
         {
-            const struct hlsl_ir_var *var = entry_func->parameters.vars[i];
+            var = entry_func->parameters.vars[i];
-
             if (var->is_output_semantic)
             {
                 record_allocation(ctx, &allocator, 0, VKD3DSP_WRITEMASK_ALL, var->first_write, var->last_read);
@@ -5138,8 +5163,9 @@ static void allocate_temp_registers(struct hlsl_ctx *ctx, struct hlsl_ir_functio
     }
 
     allocate_temp_registers_recurse(ctx, &entry_func->body, &allocator);
-    ctx->temp_count = allocator.reg_count;
     vkd3d_free(allocator.allocations);
+
+    return allocator.reg_count;
 }
 
 static void allocate_semantic_register(struct hlsl_ctx *ctx, struct hlsl_ir_var *var, unsigned int *counter, bool output)
@@ -7743,13 +7769,10 @@ static void process_entry_function(struct hlsl_ctx *ctx, struct hlsl_ir_function
     compute_liveness(ctx, entry_func);
     mark_vars_usage(ctx);
 
-    transform_derefs(ctx, mark_indexable_vars, body);
-
     calculate_resource_register_counts(ctx);
 
     allocate_register_reservations(ctx, &ctx->extern_vars);
     allocate_register_reservations(ctx, &entry_func->extern_vars);
-    allocate_temp_registers(ctx, entry_func);
     allocate_semantic_registers(ctx, entry_func);
 }
 
@@ -7774,6 +7797,8 @@ int hlsl_emit_bytecode(struct hlsl_ctx *ctx, struct hlsl_ir_function_decl *entry
 
     if (profile->major_version < 4)
     {
+        mark_indexable_vars(ctx, entry_func);
+        allocate_temp_registers(ctx, entry_func);
         allocate_const_registers(ctx, entry_func);
     }
     else

libs/vkd3d-shader/tpf.c

@@ -6387,7 +6387,7 @@ static void write_sm4_block(const struct tpf_writer *tpf, const struct hlsl_bloc
     }
 }
 
-static void write_sm4_shdr(struct hlsl_ctx *ctx, const struct hlsl_ir_function_decl *entry_func,
+static void write_sm4_shdr(struct hlsl_ctx *ctx, struct hlsl_ir_function_decl *entry_func,
         struct sm4_stat *stat, struct dxbc_writer *dxbc)
 {
     const struct hlsl_profile_info *profile = ctx->profile;
@@ -6399,6 +6399,7 @@ static void write_sm4_shdr(struct hlsl_ctx *ctx, const struct hlsl_ir_function_d
     const struct hlsl_ir_var *var;
     size_t token_count_position;
     struct tpf_writer tpf;
+    uint32_t temp_count;
 
     static const uint16_t shader_types[VKD3D_SHADER_TYPE_COUNT] =
     {
@@ -6413,6 +6414,11 @@ static void write_sm4_shdr(struct hlsl_ctx *ctx, const struct hlsl_ir_function_d
         VKD3D_SM4_LIB,
     };
 
+    mark_indexable_vars(ctx, entry_func);
+    temp_count = allocate_temp_registers(ctx, entry_func);
+    if (ctx->result)
+        return;
+
     tpf_writer_init(&tpf, ctx, stat, &buffer);
 
     extern_resources = sm4_get_extern_resources(ctx, &extern_resources_count);
@@ -6450,8 +6456,8 @@ static void write_sm4_shdr(struct hlsl_ctx *ctx, const struct hlsl_ir_function_d
     if (profile->type == VKD3D_SHADER_TYPE_COMPUTE)
         write_sm4_dcl_thread_group(&tpf, ctx->thread_count);
 
-    if (ctx->temp_count)
+    if (temp_count)
-        write_sm4_dcl_temps(&tpf, ctx->temp_count);
+        write_sm4_dcl_temps(&tpf, temp_count);
 
     LIST_FOR_EACH_ENTRY(scope, &ctx->scopes, struct hlsl_scope, entry)
     {