aco: Fix how p_elect interacts with optimizations.
authorTimur Kristóf <timur.kristof@gmail.com>
Sat, 17 Jul 2021 21:25:52 +0000 (23:25 +0200)
committerDaniel Schürmann <daniel@schuermann.dev>
Sat, 17 Jul 2021 22:48:06 +0000 (00:48 +0200)
Since p_elect doesn't have any operands, ACO's value numbering and/or
the pre-RA optimizer could currently recognize two p_elect instructions
in two different blocks as the same.

This patch adds exec as an operand to p_elect in order to achieve
correct behavior.

Fixes: e66f54e5c83fd545e1a4062e683b584a35dacc00
Closes: #5080
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11943>

src/amd/compiler/aco_instruction_selection.cpp

index 70bf5e8..300356d 100644 (file)
@@ -8722,7 +8722,11 @@ visit_intrinsic(isel_context* ctx, nir_intrinsic_instr* instr)
       break;
    }
    case nir_intrinsic_elect: {
-      Temp elected = bld.pseudo(aco_opcode::p_elect, bld.def(bld.lm));
+      /* p_elect is lowered in aco_insert_exec_mask.
+       * Use exec as an operand so value numbering and the pre-RA optimizer won't recognize
+       * two p_elect with different exec masks as the same.
+       */
+      Temp elected = bld.pseudo(aco_opcode::p_elect, bld.def(bld.lm), Operand(exec, bld.lm));
       emit_wqm(bld, elected, get_ssa_temp(ctx, &instr->dest.ssa));
       ctx->block->kind |= block_kind_needs_lowering;
       break;