i965: Fix variable indexing of UBO arrays under non-uniform control flow.

[mesa.git] / src / mesa / drivers / dri / i965 / brw_fs_visitor.cpp

diff --git a/src/mesa/drivers/dri/i965/brw_fs_visitor.cpp b/src/mesa/drivers/dri/i965/brw_fs_visitor.cpp
index 01d4cbd1ac7ed7b92a7530f9af6a5876d7abf9fa..2128795c7ddc2c7942e6109281abee4a658f331f 100644 (file)
--- a/src/mesa/drivers/dri/i965/brw_fs_visitor.cpp
+++ b/src/mesa/drivers/dri/i965/brw_fs_visitor.cpp
@@ -39,6 +39,7 @@
 #include "brw_context.h"
 #include "brw_eu.h"
 #include "brw_wm.h"
+#include "brw_cs.h"
 #include "brw_vec4.h"
 #include "brw_fs.h"
 #include "main/uniforms.h"
@@ -334,6 +335,18 @@ fs_visitor::emit_minmax(enum brw_conditional_mod conditionalmod, const fs_reg &d
    }
 }
 
+void
+fs_visitor::emit_uniformize(const fs_reg &dst, const fs_reg &src)
+{
+   const fs_reg chan_index = vgrf(glsl_type::uint_type);
+
+   emit(SHADER_OPCODE_FIND_LIVE_CHANNEL, component(chan_index, 0))
+      ->force_writemask_all = true;
+   emit(SHADER_OPCODE_BROADCAST, component(dst, 0),
+        src, component(chan_index, 0))
+      ->force_writemask_all = true;
+}
+
 bool
 fs_visitor::try_emit_saturate(ir_expression *ir)
 {
@@ -1194,13 +1207,13 @@ fs_visitor::visit(ir_expression *ir)
                                  const_uniform_block->value.u[0]);
       } else {
          /* The block index is not a constant. Evaluate the index expression
-          * per-channel and add the base UBO index; the generator will select
-          * a value from any live channel.
+          * per-channel and add the base UBO index; we have to select a value
+          * from any live channel.
           */
          surf_index = vgrf(glsl_type::uint_type);
          emit(ADD(surf_index, op[0],
-                  fs_reg(stage_prog_data->binding_table.ubo_start)))
-            ->force_writemask_all = true;
+                  fs_reg(stage_prog_data->binding_table.ubo_start)));
+         emit_uniformize(surf_index, surf_index);
 
          /* Assume this may touch any UBO. It would be nice to provide
           * a tighter bound, but the array information is already lowered away.
@@ -3286,7 +3299,7 @@ fs_visitor::emit_untyped_atomic(unsigned atomic_op, unsigned surf_index,
 
    /* Emit the instruction. */
    fs_inst *inst = emit(SHADER_OPCODE_UNTYPED_ATOMIC, dst, src_payload,
-                        fs_reg(atomic_op), fs_reg(surf_index));
+                        fs_reg(surf_index), fs_reg(atomic_op));
    inst->mlen = mlen;
 }
 
@@ -3334,7 +3347,7 @@ fs_visitor::emit_untyped_surface_read(unsigned surf_index, fs_reg dst,
 
    /* Emit the instruction. */
    inst = emit(SHADER_OPCODE_UNTYPED_SURFACE_READ, dst, src_payload,
-               fs_reg(surf_index));
+               fs_reg(surf_index), fs_reg(1));
    inst->mlen = mlen;
 }
 
@@ -4237,6 +4250,25 @@ fs_visitor::fs_visitor(struct brw_context *brw,
    init();
 }
 
+fs_visitor::fs_visitor(struct brw_context *brw,
+                       void *mem_ctx,
+                       const struct brw_cs_prog_key *key,
+                       struct brw_cs_prog_data *prog_data,
+                       struct gl_shader_program *shader_prog,
+                       struct gl_compute_program *cp,
+                       unsigned dispatch_width)
+   : backend_visitor(brw, shader_prog, &cp->Base, &prog_data->base,
+                     MESA_SHADER_COMPUTE),
+     reg_null_f(retype(brw_null_vec(dispatch_width), BRW_REGISTER_TYPE_F)),
+     reg_null_d(retype(brw_null_vec(dispatch_width), BRW_REGISTER_TYPE_D)),
+     reg_null_ud(retype(brw_null_vec(dispatch_width), BRW_REGISTER_TYPE_UD)),
+     key(key), prog_data(&prog_data->base),
+     dispatch_width(dispatch_width)
+{
+   this->mem_ctx = mem_ctx;
+   init();
+}
+
 void
 fs_visitor::init()
 {
@@ -4248,6 +4280,9 @@ fs_visitor::init()
    case MESA_SHADER_GEOMETRY:
       key_tex = &((const brw_vue_prog_key *) key)->tex;
       break;
+   case MESA_SHADER_COMPUTE:
+      key_tex = &((const brw_cs_prog_key*) key)->tex;
+      break;
    default:
       unreachable("unhandled shader stage");
    }