ir_to_mesa: Get temps allocated at the right times.

The alloced_vec4/vec4 distinction was an experiment to expose the cost
of temps to the codegen.  But the problem is that the temporary
production rule gets called after the emit rule that was using the
temp.  We could have the args to emit_op be pointers to where the temp
would get allocated later, but that seems overly hard while just
trying to bring this thing up.  Besides, the temps used in expressions
bear only the vaguest relation to how many temps will be used after
register allocation.

diff --git a/ir_to_mesa.cpp b/ir_to_mesa.cpp
index f36dea5f314b9f882a930506a9a54522c849a70f..eb55f82e27ff5effd62cf14aaa8999e2e1ca097c 100644 (file)
--- a/ir_to_mesa.cpp
+++ b/ir_to_mesa.cpp
@@ -293,6 +293,9 @@ ir_to_mesa_visitor::visit(ir_expression *ir)
       ir->accept(&v);
       exit(1);
    }
+
+   /* Allocate a temporary for the result. */
+   this->get_temp(this->result);
 }
 
 
@@ -309,6 +312,7 @@ ir_to_mesa_visitor::visit(ir_swizzle *ir)
    assert(this->result);
 
    tree = this->create_tree(MB_TERM_swizzle_vec4, ir, this->result, NULL);
+   this->get_temp(tree);
 
    for (i = 0; i < 4; i++) {
       if (i < ir->type->vector_elements) {

diff --git a/mesa_codegen.brg b/mesa_codegen.brg
index 9f2761b08e278057a86d0fd0956197dfb4d58009..f1f24dab84ffd1d1fdccc52b6e471e386882f8e8 100644 (file)
--- a/mesa_codegen.brg
+++ b/mesa_codegen.brg
@@ -76,25 +76,10 @@
 # produced at the cost of "cost".  We measure "cost" in approximate
 # instruction count.  The BURG should then more or less minimize the
 # number of instructions.
-#
-# A reference of a variable has an allocated register already, so it
-# can be used as an argument for pretty much anything.
-alloced_vec4: reference_vec4 0
-
-# If something produces a vec4 with a location already, then we don't need
-# to allocate a temp reg for it.
-vec4: alloced_vec4 0
-
-# If something produces a vec4 result that needs a place to live,
-# then there's a cost with allocating a temporary for it.  We
-# approximate that as one instruction's cost, even though sometimes
-# that temp might not be a newly-allocated temp due to later
-# live-dead analysis.
-alloced_vec4: vec4 1
-{
-       /* FINISHME */
-       tree->v->get_temp(tree);
-}
+
+# A reference of a variable is just a vec4 register location,
+# so it can be used as an argument for pretty much anything.
+vec4: reference_vec4 0
 
 # Here's the rule everyone will hit: Moving the result of an
 # expression into a variable-dereference register location.
@@ -102,14 +87,14 @@ alloced_vec4: vec4 1
 # Note that this is likely a gratuitous move.  We could make variants
 # of each of the following rules, e.g:
 #
-# vec4: add_vec4_vec4(alloced_vec4, alloced_vec4) 1
+# vec4: add_vec4_vec4(vec4, vec4) 1
 # {
 #         emit(ADD, tree, tree->left, tree->right);
 # }
 #
 # becoming
 #
-# vec4: assign(alloced_vec4_vec4, add_vec4_vec4(alloced_vec4, alloced_vec4) 1
+# vec4: assign(vec4_vec4, add_vec4_vec4(vec4, vec4) 1
 # {
 #         emit(ADD, tree->left, tree->right->left, tree->right->right);
 # }
@@ -117,7 +102,7 @@ alloced_vec4: vec4 1
 # But it seems like a lot of extra typing and duped code, when we
 # probably want copy propagation and dead code after codegen anyway,
 # which would clean these up.
-stmt: assign(alloced_vec4, alloced_vec4) 1
+stmt: assign(vec4, vec4) 1
 {
        ir_to_mesa_emit_op1(tree, OPCODE_MOV,
                            ir_to_mesa_dst_reg_from_src(tree->left->src_reg),
@@ -126,7 +111,7 @@ stmt: assign(alloced_vec4, alloced_vec4) 1
 
 # Perform a swizzle by composing our swizzle with the swizzle
 # required to get at the src reg.
-vec4: swizzle_vec4(alloced_vec4) 1
+vec4: swizzle_vec4(vec4) 1
 {
        ir_to_mesa_src_reg reg = tree->left->src_reg;
        int swiz[4];
@@ -145,7 +130,7 @@ vec4: swizzle_vec4(alloced_vec4) 1
                           reg);
 }
 
-vec4: add_vec4_vec4(alloced_vec4, alloced_vec4) 1
+vec4: add_vec4_vec4(vec4, vec4) 1
 {
        ir_to_mesa_emit_op2(tree, OPCODE_ADD,
                            ir_to_mesa_dst_reg_from_src(tree->src_reg),
@@ -153,7 +138,7 @@ vec4: add_vec4_vec4(alloced_vec4, alloced_vec4) 1
                            tree->right->src_reg);
 }
 
-vec4: sub_vec4_vec4(alloced_vec4, alloced_vec4) 1
+vec4: sub_vec4_vec4(vec4, vec4) 1
 {
        ir_to_mesa_emit_op2(tree, OPCODE_SUB,
                            ir_to_mesa_dst_reg_from_src(tree->src_reg),
@@ -161,7 +146,7 @@ vec4: sub_vec4_vec4(alloced_vec4, alloced_vec4) 1
                            tree->right->src_reg);
 }
 
-vec4: mul_vec4_vec4(alloced_vec4, alloced_vec4) 1
+vec4: mul_vec4_vec4(vec4, vec4) 1
 {
        ir_to_mesa_emit_op2(tree, OPCODE_MUL,
                            ir_to_mesa_dst_reg_from_src(tree->src_reg),
@@ -169,7 +154,7 @@ vec4: mul_vec4_vec4(alloced_vec4, alloced_vec4) 1
                            tree->right->src_reg);
 }
 
-vec4: dp4_vec4_vec4(alloced_vec4, alloced_vec4) 1
+vec4: dp4_vec4_vec4(vec4, vec4) 1
 {
        ir_to_mesa_emit_op2(tree, OPCODE_DP4,
                            ir_to_mesa_dst_reg_from_src(tree->src_reg),
@@ -178,7 +163,7 @@ vec4: dp4_vec4_vec4(alloced_vec4, alloced_vec4) 1
        tree->src_reg.swizzle = SWIZZLE_XXXX;
 }
 
-vec4: dp3_vec4_vec4(alloced_vec4, alloced_vec4) 1
+vec4: dp3_vec4_vec4(vec4, vec4) 1
 {
        ir_to_mesa_emit_op2(tree, OPCODE_DP3,
                            ir_to_mesa_dst_reg_from_src(tree->src_reg),
@@ -188,7 +173,7 @@ vec4: dp3_vec4_vec4(alloced_vec4, alloced_vec4) 1
 }
 
 
-vec4: dp2_vec4_vec4(alloced_vec4, alloced_vec4) 1
+vec4: dp2_vec4_vec4(vec4, vec4) 1
 {
        ir_to_mesa_emit_op2(tree, OPCODE_DP2,
                            ir_to_mesa_dst_reg_from_src(tree->src_reg),
@@ -197,7 +182,7 @@ vec4: dp2_vec4_vec4(alloced_vec4, alloced_vec4) 1
        tree->src_reg.swizzle = SWIZZLE_XXXX;
 }
 
-vec4: div_vec4_vec4(alloced_vec4, alloced_vec4) 1
+vec4: div_vec4_vec4(vec4, vec4) 1
 {
        /* FINISHME: Mesa RCP only uses the X channel, this node is for vec4. */
        ir_to_mesa_emit_op1(tree, OPCODE_RCP,
@@ -210,7 +195,7 @@ vec4: div_vec4_vec4(alloced_vec4, alloced_vec4) 1
                            tree->left->src_reg);
 }
 
-vec4: sqrt_vec4(alloced_vec4) 1
+vec4: sqrt_vec4(vec4) 1
 {
        /* FINISHME: Mesa RSQ only uses the X channel, this node is for vec4. */
        ir_to_mesa_emit_op1(tree, OPCODE_RSQ,