i965/fs: Clear result before visiting shadow comparitor and LOD info.
authorKenneth Graunke <kenneth@whitecape.org>
Fri, 22 Jul 2011 04:17:10 +0000 (21:17 -0700)
committerKenneth Graunke <kenneth@whitecape.org>
Mon, 25 Jul 2011 17:45:56 +0000 (10:45 -0700)
Commit 53c89c67f33639afef951e178f93f4e29acc5d53 ("i965: Avoid generating
MOVs for assignments of expressions.") added the line "this->result =
reg_undef" all over the code.  Unfortunately, since Eric developed his
patch before I landed Ivybridge support, he missed adding it to
fs_visitor::emit_texture_gen7() after rebasing.

Furthermore, since I developed TXD support before Eric's patch, I
neglected to add it to the gradient handling when I rebased.

Neglecting to set this causes the visitor to use this->result as storage
rather than generating a new temporary.  These missing statements
resulted in the same register being used to store several different
values.

Fixes the following piglit tests on Ivybridge:
- glsl-fs-shadow2dproj.shader_test
- glsl-fs-shadow2dproj-bias.shader_test

NOTE: This is a candidate for the 7.11 branch.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
src/mesa/drivers/dri/i965/brw_fs_visitor.cpp

index cbe5cf428c59be577196ae4c757801f3be35e6d7..9632aae64b063ab1faddd623a7c243e825311e80 100644 (file)
@@ -603,9 +603,11 @@ fs_visitor::emit_texture_gen4(ir_texture *ir, fs_reg dst, fs_reg coordinate,
       /* gen4's SIMD8 sampler always has the slots for u,v,r present. */
       mlen += 3;
    } else if (ir->op == ir_txd) {
+      this->result = reg_undef;
       ir->lod_info.grad.dPdx->accept(this);
       fs_reg dPdx = this->result;
 
+      this->result = reg_undef;
       ir->lod_info.grad.dPdy->accept(this);
       fs_reg dPdy = this->result;
 
@@ -786,9 +788,11 @@ fs_visitor::emit_texture_gen5(ir_texture *ir, fs_reg dst, fs_reg coordinate,
       inst = emit(FS_OPCODE_TXL, dst);
       break;
    case ir_txd: {
+      this->result = reg_undef;
       ir->lod_info.grad.dPdx->accept(this);
       fs_reg dPdx = this->result;
 
+      this->result = reg_undef;
       ir->lod_info.grad.dPdy->accept(this);
       fs_reg dPdy = this->result;
 
@@ -850,6 +854,7 @@ fs_visitor::emit_texture_gen7(ir_texture *ir, fs_reg dst, fs_reg coordinate,
    }
 
    if (ir->shadow_comparitor && ir->op != ir_txd) {
+      this->result = reg_undef;
       ir->shadow_comparitor->accept(this);
       emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
       mlen += reg_width;
@@ -860,11 +865,13 @@ fs_visitor::emit_texture_gen7(ir_texture *ir, fs_reg dst, fs_reg coordinate,
    case ir_tex:
       break;
    case ir_txb:
+      this->result = reg_undef;
       ir->lod_info.bias->accept(this);
       emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
       mlen += reg_width;
       break;
    case ir_txl:
+      this->result = reg_undef;
       ir->lod_info.lod->accept(this);
       emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
       mlen += reg_width;
@@ -873,9 +880,11 @@ fs_visitor::emit_texture_gen7(ir_texture *ir, fs_reg dst, fs_reg coordinate,
       if (c->dispatch_width == 16)
         fail("Gen7 does not support sample_d/sample_d_c in SIMD16 mode.");
 
+      this->result = reg_undef;
       ir->lod_info.grad.dPdx->accept(this);
       fs_reg dPdx = this->result;
 
+      this->result = reg_undef;
       ir->lod_info.grad.dPdy->accept(this);
       fs_reg dPdy = this->result;
 
@@ -1070,6 +1079,7 @@ fs_visitor::visit(ir_texture *ir)
       if (hw_compare_supported) {
         inst->shadow_compare = true;
       } else {
+        this->result = reg_undef;
         ir->shadow_comparitor->accept(this);
         fs_reg ref = this->result;