aco: Make a better guess at which instructions need the VCC hint.

Previously, bool_to_vector_condition would always set the VCC hint
on its result. This commit improves it by having the optimizer set
the VCC hint only when the result really needs to be in the VCC.

Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Tested-by: Marge Bot <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3451>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3451>

src/amd/compiler/aco_instruction_selection.cpp

@@ -356,7 +356,7 @@ Temp bool_to_vector_condition(isel_context *ctx, Temp val, Temp dst = Temp(0, s2
    assert(val.regClass() == s1);
    assert(dst.regClass() == bld.lm);
 
-   return bld.sop2(Builder::s_cselect, bld.hint_vcc(Definition(dst)), Operand((uint32_t) -1), Operand(0u), bld.scc(val));
+   return bld.sop2(Builder::s_cselect, Definition(dst), Operand((uint32_t) -1), Operand(0u), bld.scc(val));
 }
 
 Temp bool_to_scalar_condition(isel_context *ctx, Temp val, Temp dst = Temp(0, s1))

src/amd/compiler/aco_optimizer.cpp

@@ -85,6 +85,7 @@ enum Label {
    label_constant_64bit = 1 << 22,
    label_uniform_bitwise = 1 << 23,
    label_scc_invert = 1 << 24,
+   label_vcc_hint = 1 << 25,
 };
 
 static constexpr uint32_t instr_labels = label_vec | label_mul | label_mad | label_omod_success | label_clamp_success |
@@ -405,6 +406,15 @@ struct ssa_info {
       return label & label_uniform_bool;
    }
 
+   void set_vcc_hint()
+   {
+      add_label(label_vcc_hint);
+   }
+
+   bool is_vcc_hint()
+   {
+      return label & label_vcc_hint;
+   }
 };
 
 struct opt_ctx {
@@ -1087,6 +1097,8 @@ void label_instruction(opt_ctx &ctx, Block& block, aco_ptr<Instruction>& instr)
                instr->operands[1].constantEquals(0x3f800000u) &&
                instr->operands[2].isTemp())
          ctx.info[instr->definitions[0].tempId()].set_b2f(instr->operands[2].getTemp());
+
+      ctx.info[instr->operands[2].tempId()].set_vcc_hint();
       break;
    case aco_opcode::v_cmp_lg_u32:
       if (instr->format == Format::VOPC && /* don't optimize VOP3 / SDWA / DPP */
@@ -2235,6 +2247,10 @@ void combine_instruction(opt_ctx &ctx, Block& block, aco_ptr<Instruction>& instr
          return;
    }
 
+   if (ctx.info[instr->definitions[0].tempId()].is_vcc_hint()) {
+      instr->definitions[0].setHint(vcc);
+   }
+
    /* TODO: There are still some peephole optimizations that could be done:
     * - abs(a - b) -> s_absdiff_i32
     * - various patterns for s_bitcmp{0,1}_b32 and s_bitset{0,1}_b32