glsl: Skip type conversions in loop-size estimates on float-only arches.loop

On float-only architectures (such as i915), type conversions are all
no-ops, so we shouldn't include them when estimating how many
instructions a loop contains.

Overestimating can result in the unroller deciding not to unroll the
loop, which can be detrimental on these architectures.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>

15 files changed, 51 insertions, 18 deletions

diff --git a/src/glsl/glsl_parser_extras.cpp b/src/glsl/glsl_parser_extras.cpp
index ae7a365f4b2..a67e8f5e378 100644
--- a/src/glsl/glsl_parser_extras.cpp
+++ b/src/glsl/glsl_parser_extras.cpp
@@ -1011,7 +1011,8 @@ ast_struct_specifier::ast_struct_specifier(const char *identifier,
 bool
 do_common_optimization(exec_list *ir, bool linked,
 		       bool uniform_locations_assigned,
-		       unsigned max_unroll_iterations)
+		       unsigned max_unroll_iterations,
+		       bool native_integers)
 {
    GLboolean progress = GL_FALSE;
 
@@ -1050,7 +1051,7 @@ do_common_optimization(exec_list *ir, bool linked,
    loop_state *ls = analyze_loop_variables(ir);
    if (ls->loop_found) {
       progress = set_loop_controls(ir, ls) || progress;
-      progress = unroll_loops(ir, ls, max_unroll_iterations) || progress;
+      progress = unroll_loops(ir, ls, max_unroll_iterations, native_integers) || progress;
    }
    delete ls;
 
diff --git a/src/glsl/ir.cpp b/src/glsl/ir.cpp
index 1ba87515ea7..b601c7b11fd 100644
--- a/src/glsl/ir.cpp
+++ b/src/glsl/ir.cpp
@@ -507,6 +507,12 @@ ir_expression::get_operator(const char *str)
    return (ir_expression_operation) -1;
 }
 
+bool
+ir_expression::is_type_conversion()
+{
+   return (operation >= ir_unop_f2i && operation <= ir_unop_u2i);
+}
+
 ir_constant::ir_constant()
 {
    this->ir_type = ir_type_constant;
diff --git a/src/glsl/ir.h b/src/glsl/ir.h
index d6c6a607ae8..1ffa5f4ad39 100644
--- a/src/glsl/ir.h
+++ b/src/glsl/ir.h
@@ -1021,6 +1021,9 @@ public:
     */
    static ir_expression_operation get_operator(const char *);
 
+   /** Whether or not this expression is a type conversion (i.e. i2f) */
+   bool is_type_conversion();
+
    virtual void accept(ir_visitor *v)
    {
       v->visit(this);
diff --git a/src/glsl/ir_optimization.h b/src/glsl/ir_optimization.h
index 35678358307..b5c5ee17f4e 100644
--- a/src/glsl/ir_optimization.h
+++ b/src/glsl/ir_optimization.h
@@ -39,7 +39,8 @@
 
 bool do_common_optimization(exec_list *ir, bool linked,
 			    bool uniform_locations_assigned,
-			    unsigned max_unroll_iterations);
+			    unsigned max_unroll_iterations,
+			    bool native_integers);
 
 bool do_algebraic(exec_list *instructions);
 bool do_constant_folding(exec_list *instructions);
diff --git a/src/glsl/linker.cpp b/src/glsl/linker.cpp
index 6ba297237c7..e082d3e4bde 100644
--- a/src/glsl/linker.cpp
+++ b/src/glsl/linker.cpp
@@ -2313,7 +2313,7 @@ link_shaders(struct gl_context *ctx, struct gl_shader_program *prog)
 
       unsigned max_unroll = ctx->ShaderCompilerOptions[i].MaxUnrollIterations;
 
-      while (do_common_optimization(prog->_LinkedShaders[i]->ir, true, false, max_unroll))
+      while (do_common_optimization(prog->_LinkedShaders[i]->ir, true, false, max_unroll, ctx->Const.NativeIntegers))
 	 ;
    }
 
diff --git a/src/glsl/loop_analysis.cpp b/src/glsl/loop_analysis.cpp
index 6a0e4da5100..8bbc0caa20f 100644
--- a/src/glsl/loop_analysis.cpp
+++ b/src/glsl/loop_analysis.cpp
@@ -238,6 +238,10 @@ loop_analysis::visit_leave(ir_loop *ir)
    if (ls->contains_calls)
       return visit_continue;
 
+   printf("KAYDEN: about to analyze:\n");
+   ir->print();
+   printf("KAYDEN: -----------------\n");
+
    foreach_list(node, &ir->body_instructions) {
       /* Skip over declarations at the start of a loop.
        */
diff --git a/src/glsl/loop_analysis.h b/src/glsl/loop_analysis.h
index 8bed1db0210..9c88867b84a 100644
--- a/src/glsl/loop_analysis.h
+++ b/src/glsl/loop_analysis.h
@@ -57,7 +57,8 @@ set_loop_controls(exec_list *instructions, loop_state *ls);
 
 
 extern bool
-unroll_loops(exec_list *instructions, loop_state *ls, unsigned max_iterations);
+unroll_loops(exec_list *instructions, loop_state *ls, unsigned max_iterations,
+	     bool native_integers);
 
 
 /**
diff --git a/src/glsl/loop_unroll.cpp b/src/glsl/loop_unroll.cpp
index 3434fde6292..3d5cbcf8046 100644
--- a/src/glsl/loop_unroll.cpp
+++ b/src/glsl/loop_unroll.cpp
@@ -27,11 +27,13 @@
 
 class loop_unroll_visitor : public ir_hierarchical_visitor {
 public:
-   loop_unroll_visitor(loop_state *state, unsigned max_iterations)
+   loop_unroll_visitor(loop_state *state, unsigned max_iterations,
+		       bool native_integers)
    {
       this->state = state;
       this->progress = false;
       this->max_iterations = max_iterations;
+      this->native_integers = native_integers;
    }
 
    virtual ir_visitor_status visit_leave(ir_loop *ir);
@@ -40,6 +42,7 @@ public:
 
    bool progress;
    unsigned max_iterations;
+   bool native_integers;
 };
 
 
@@ -54,11 +57,13 @@ class loop_unroll_count : public ir_hierarchical_visitor {
 public:
    int nodes;
    bool fail;
+   bool native_integers;
 
-   loop_unroll_count(exec_list *list)
+   loop_unroll_count(exec_list *list, bool native_ints)
    {
       nodes = 0;
       fail = false;
+      native_integers = native_ints;
 
       run(list);
    }
@@ -71,6 +76,10 @@ public:
 
    virtual ir_visitor_status visit_enter(ir_expression *ir)
    {
+      /* On float-only architectures, type conversions are no-ops. */
+      if (ir->is_type_conversion() && !native_integers)
+	 return visit_continue;
+
       nodes++;
       return visit_continue;
    }
@@ -112,7 +121,7 @@ loop_unroll_visitor::visit_leave(ir_loop *ir)
 
    /* Don't try to unroll nested loops and loops with a huge body.
     */
-   loop_unroll_count count(&ir->body_instructions);
+   loop_unroll_count count(&ir->body_instructions, this->native_integers);
 
    if (count.fail || count.nodes * iterations > (int)max_iterations * 5)
       return visit_continue;
@@ -243,9 +252,10 @@ loop_unroll_visitor::visit_leave(ir_loop *ir)
 
 
 bool
-unroll_loops(exec_list *instructions, loop_state *ls, unsigned max_iterations)
+unroll_loops(exec_list *instructions, loop_state *ls, unsigned max_iterations,
+	     bool native_integers)
 {
-   loop_unroll_visitor v(ls, max_iterations);
+   loop_unroll_visitor v(ls, max_iterations, native_integers);
 
    v.run(instructions);
 
diff --git a/src/glsl/main.cpp b/src/glsl/main.cpp
index d43bf1a7463..acb6a3c68f7 100644
--- a/src/glsl/main.cpp
+++ b/src/glsl/main.cpp
@@ -166,7 +166,7 @@ compile_shader(struct gl_context *ctx, struct gl_shader *shader)
    if (!state->error && !shader->ir->is_empty()) {
       bool progress;
       do {
-	 progress = do_common_optimization(shader->ir, false, false, 32);
+	 progress = do_common_optimization(shader->ir, false, false, 32, true);
       } while (progress);
 
       validate_ir_tree(shader->ir);
diff --git a/src/glsl/test_optpass.cpp b/src/glsl/test_optpass.cpp
index 6abafb5d311..ef4af252474 100644
--- a/src/glsl/test_optpass.cpp
+++ b/src/glsl/test_optpass.cpp
@@ -64,7 +64,7 @@ do_optimization(struct exec_list *ir, const char *optimization)
 
    if (sscanf(optimization, "do_common_optimization ( %d , %d ) ",
               &int_0, &int_1) == 2) {
-      return do_common_optimization(ir, int_0 != 0, false, int_1);
+      return do_common_optimization(ir, int_0 != 0, false, int_1, true);
    } else if (strcmp(optimization, "do_algebraic") == 0) {
       return do_algebraic(ir);
    } else if (strcmp(optimization, "do_constant_folding") == 0) {
diff --git a/src/mesa/drivers/dri/i965/brw_shader.cpp b/src/mesa/drivers/dri/i965/brw_shader.cpp
index aa1bfdbd230..7775bd3cc58 100644
--- a/src/mesa/drivers/dri/i965/brw_shader.cpp
+++ b/src/mesa/drivers/dri/i965/brw_shader.cpp
@@ -184,7 +184,7 @@ brw_link_shader(struct gl_context *ctx, struct gl_shader_program *shProg)
 				   false /* loops */
 				   ) || progress;
 
-	 progress = do_common_optimization(shader->ir, true, true, 32)
+	 progress = do_common_optimization(shader->ir, true, true, 32, true)
 	   || progress;
       } while (progress);
 
diff --git a/src/mesa/drivers/dri/intel/intel_tex_layout.c b/src/mesa/drivers/dri/intel/intel_tex_layout.c
index 65645bc46a4..db7b84f8a82 100644
--- a/src/mesa/drivers/dri/intel/intel_tex_layout.c
+++ b/src/mesa/drivers/dri/intel/intel_tex_layout.c
@@ -115,17 +115,22 @@ intel_vertical_texture_alignment_unit(struct intel_context *intel,
    if (_mesa_is_format_compressed(format))
       return 4;
 
-   if (format == MESA_FORMAT_S8)
+   if (format == MESA_FORMAT_S8) {
+      printf("vertical align 8/4 (S8): format %s\n", _mesa_get_format_name(format));
       return intel->gen >= 7 ? 8 : 4;
+   }
 
    GLenum base_format = _mesa_get_format_base_format(format);
 
    if (intel->gen >= 6 &&
        (base_format == GL_DEPTH_COMPONENT ||
 	base_format == GL_DEPTH_STENCIL)) {
+      printf("vertical align 4: format %s\n", _mesa_get_format_name(format));
       return 4;
    }
 
+   printf("vertical align 2: format %s\n", _mesa_get_format_name(format));
+
    return 2;
 }
 
diff --git a/src/mesa/main/ff_fragment_shader.cpp b/src/mesa/main/ff_fragment_shader.cpp
index 3c91b1a42ad..f4a6072d407 100644
--- a/src/mesa/main/ff_fragment_shader.cpp
+++ b/src/mesa/main/ff_fragment_shader.cpp
@@ -1324,7 +1324,7 @@ create_new_program(struct gl_context *ctx, struct state_key *key)
 
    validate_ir_tree(p.shader->ir);
 
-   while (do_common_optimization(p.shader->ir, false, false, 32))
+   while (do_common_optimization(p.shader->ir, false, false, 32, ctx->Const.NativeIntegers))
       ;
    reparent_ir(p.shader->ir, p.shader->ir);
 
diff --git a/src/mesa/program/ir_to_mesa.cpp b/src/mesa/program/ir_to_mesa.cpp
index 840648e0449..a8873466886 100644
--- a/src/mesa/program/ir_to_mesa.cpp
+++ b/src/mesa/program/ir_to_mesa.cpp
@@ -3061,7 +3061,8 @@ _mesa_ir_link_shader(struct gl_context *ctx, struct gl_shader_program *prog)
 	 progress = do_lower_jumps(ir, true, true, options->EmitNoMainReturn, options->EmitNoCont, options->EmitNoLoops) || progress;
 
 	 progress = do_common_optimization(ir, true, true,
-					   options->MaxUnrollIterations)
+					   options->MaxUnrollIterations,
+					   false)
 	   || progress;
 
 	 progress = lower_quadop_vector(ir, true) || progress;
@@ -3171,7 +3172,7 @@ _mesa_glsl_compile_shader(struct gl_context *ctx, struct gl_shader *shader)
       /* Do some optimization at compile time to reduce shader IR size
        * and reduce later work if the same shader is linked multiple times
        */
-      while (do_common_optimization(shader->ir, false, false, 32))
+      while (do_common_optimization(shader->ir, false, false, 32, false))
 	 ;
 
       validate_ir_tree(shader->ir);
diff --git a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
index 9e68deb3471..fdc782e4bba 100644
--- a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
+++ b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
@@ -5002,7 +5002,8 @@ st_link_shader(struct gl_context *ctx, struct gl_shader_program *prog)
          progress = do_lower_jumps(ir, true, true, options->EmitNoMainReturn, options->EmitNoCont, options->EmitNoLoops) || progress;
 
          progress = do_common_optimization(ir, true, true,
-					   options->MaxUnrollIterations)
+					   options->MaxUnrollIterations,
+					   ctx->Const.NativeIntegers)
 	   || progress;
 
          progress = lower_quadop_vector(ir, false) || progress;