glsl/lower_vector_derefs: Don't use a temporary for TCS outputs
authorJason Ekstrand <jason.ekstrand@intel.com>
Tue, 12 Mar 2019 01:43:15 +0000 (20:43 -0500)
committerJason Ekstrand <jason@jlekstrand.net>
Wed, 13 Mar 2019 02:10:31 +0000 (02:10 +0000)
Tessellation control shader outputs act as if they have memory backing
them and you can have multiple writes to different components of the
same vector in-flight at the same time.  When this happens, the load vec
store pattern that gets used by ir_triop_vector_insert doesn't yield the
correct results.  Instead, just emit a sequence of conditional
assignments.

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Cc: mesa-stable@lists.freedesktop.org
src/compiler/glsl/lower_vector_derefs.cpp

index 6cd9a2d819a3728497abd49bc59bc30e5040a73e..2aae30d8201101d2c4afcf6ce4a632aad3eda02a 100644 (file)
@@ -32,8 +32,9 @@ namespace {
 
 class vector_deref_visitor : public ir_rvalue_enter_visitor {
 public:
-   vector_deref_visitor()
-      : progress(false)
+   vector_deref_visitor(void *mem_ctx, gl_shader_stage shader_stage)
+      : progress(false), shader_stage(shader_stage),
+        factory(&factory_instructions, mem_ctx)
    {
    }
 
@@ -45,6 +46,9 @@ public:
    virtual ir_visitor_status visit_enter(ir_assignment *ir);
 
    bool progress;
+   gl_shader_stage shader_stage;
+   exec_list factory_instructions;
+   ir_factory factory;
 };
 
 } /* anonymous namespace */
@@ -65,13 +69,63 @@ vector_deref_visitor::visit_enter(ir_assignment *ir)
    ir_constant *old_index_constant =
       deref->array_index->constant_expression_value(mem_ctx);
    if (!old_index_constant) {
-      ir->rhs = new(mem_ctx) ir_expression(ir_triop_vector_insert,
-                                           new_lhs->type,
-                                           new_lhs->clone(mem_ctx, NULL),
-                                           ir->rhs,
-                                           deref->array_index);
-      ir->write_mask = (1 << new_lhs->type->vector_elements) - 1;
-      ir->set_lhs(new_lhs);
+      if (shader_stage == MESA_SHADER_TESS_CTRL &&
+          deref->variable_referenced()->data.mode == ir_var_shader_out) {
+         /* Tessellation control shader outputs act as if they have memory
+          * backing them and if we have writes from multiple threads
+          * targeting the same vec4 (this can happen for patch outputs), the
+          * load-vec-store pattern of ir_triop_vector_insert doesn't work.
+          * Instead, we have to lower to a series of conditional write-masked
+          * assignments.
+          */
+         ir_variable *const src_temp =
+            factory.make_temp(ir->rhs->type, "scalar_tmp");
+
+         /* The newly created variable declaration goes before the assignment
+          * because we're going to set it as the new LHS.
+          */
+         ir->insert_before(factory.instructions);
+         ir->set_lhs(new(mem_ctx) ir_dereference_variable(src_temp));
+
+         ir_variable *const arr_index =
+            factory.make_temp(deref->array_index->type, "index_tmp");
+         factory.emit(assign(arr_index, deref->array_index));
+
+         for (unsigned i = 0; i < new_lhs->type->vector_elements; i++) {
+            ir_constant *const cmp_index =
+               ir_constant::zero(factory.mem_ctx, deref->array_index->type);
+            cmp_index->value.u[0] = i;
+
+            ir_rvalue *const lhs_clone = new_lhs->clone(factory.mem_ctx, NULL);
+            ir_dereference_variable *const src_temp_deref =
+               new(mem_ctx) ir_dereference_variable(src_temp);
+
+            if (new_lhs->ir_type != ir_type_swizzle) {
+               assert(lhs_clone->as_dereference());
+               ir_assignment *cond_assign =
+                  new(mem_ctx) ir_assignment(lhs_clone->as_dereference(),
+                                             src_temp_deref,
+                                             equal(arr_index, cmp_index),
+                                             WRITEMASK_X << i);
+               factory.emit(cond_assign);
+            } else {
+               ir_assignment *cond_assign =
+                  new(mem_ctx) ir_assignment(swizzle(lhs_clone, i, 1),
+                                             src_temp_deref,
+                                             equal(arr_index, cmp_index));
+               factory.emit(cond_assign);
+            }
+         }
+         ir->insert_after(factory.instructions);
+      } else {
+         ir->rhs = new(mem_ctx) ir_expression(ir_triop_vector_insert,
+                                              new_lhs->type,
+                                              new_lhs->clone(mem_ctx, NULL),
+                                              ir->rhs,
+                                              deref->array_index);
+         ir->write_mask = (1 << new_lhs->type->vector_elements) - 1;
+         ir->set_lhs(new_lhs);
+      }
    } else if (new_lhs->ir_type != ir_type_swizzle) {
       ir->set_lhs(new_lhs);
       ir->write_mask = 1 << old_index_constant->get_uint_component(0);
@@ -105,7 +159,7 @@ vector_deref_visitor::handle_rvalue(ir_rvalue **rv)
 bool
 lower_vector_derefs(gl_linked_shader *shader)
 {
-   vector_deref_visitor v;
+   vector_deref_visitor v(shader->ir, shader->Stage);
 
    visit_list_elements(&v, shader->ir);