st/glsl_to_tgsi: use correct writemask when converting generic intrinsics
authorNicolai Hähnle <nicolai.haehnle@amd.com>
Mon, 12 Jun 2017 08:53:07 +0000 (10:53 +0200)
committerNicolai Hähnle <nicolai.haehnle@amd.com>
Mon, 19 Jun 2017 10:07:05 +0000 (12:07 +0200)
This fixes a bug when lowering ballotARB: previously, using writemask 0xf,
emit_asm would create TGSI_OPCODE_BALLOT instructions that span two registers
to cover 4 64-bit channels. This could trample over other a neighbouring
temporary.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=101360
Cc: 17.1 <mesa-stable@lists.freedesktop.org>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
src/mesa/state_tracker/st_glsl_to_tgsi.cpp

index 24d417d..7852941 100644 (file)
@@ -3978,6 +3978,8 @@ glsl_to_tgsi_visitor::visit_generic_intrinsic(ir_call *ir, unsigned op)
    ir->return_deref->accept(this);
    st_dst_reg dst = st_dst_reg(this->result);
 
+   dst.writemask = u_bit_consecutive(0, ir->return_deref->var->type->vector_elements);
+
    st_src_reg src[4] = { undef_src, undef_src, undef_src, undef_src };
    unsigned num_src = 0;
    foreach_in_list(ir_rvalue, param, &ir->actual_parameters) {