From fdb97d3d2914c8f887a7968432db4fdbd35d8376 Mon Sep 17 00:00:00 2001 From: =?utf8?q?Daniel=20Sch=C3=BCrmann?= Date: Mon, 10 Aug 2020 16:56:25 +0200 Subject: [PATCH] aco: execute branch instructions in WQM if necessary It could happen that only the branch condition was computed in WQM and not the branch instruction. There is now some rendundancy which should be cleaned up. Fixes: 3817fa7a4d1f51c385b28a2e45a1edf227526028 ('aco: fix WQM handling in nested loops') Reviewed-by: Rhys Perry Part-of: --- src/amd/compiler/aco_insert_exec_mask.cpp | 11 +++++------ 1 file changed, 5 insertions(+), 6 deletions(-) diff --git a/src/amd/compiler/aco_insert_exec_mask.cpp b/src/amd/compiler/aco_insert_exec_mask.cpp index 7246fb74e0c..ccdd5c7d2aa 100644 --- a/src/amd/compiler/aco_insert_exec_mask.cpp +++ b/src/amd/compiler/aco_insert_exec_mask.cpp @@ -165,12 +165,6 @@ void mark_block_wqm(wqm_ctx &ctx, unsigned block_idx) ctx.branch_wqm[block_idx] = true; Block& block = ctx.program->blocks[block_idx]; - aco_ptr& branch = block.instructions.back(); - - if (branch->opcode != aco_opcode::p_branch) { - assert(!branch->operands.empty() && branch->operands[0].isTemp()); - set_needs_wqm(ctx, branch->operands[0].getTemp()); - } /* TODO: this sets more branch conditions to WQM than it needs to * it should be enough to stop at the "exec mask top level" */ @@ -233,6 +227,11 @@ void get_block_needs(wqm_ctx &ctx, exec_ctx &exec_ctx, Block* block) } } + if (instr->format == Format::PSEUDO_BRANCH && ctx.branch_wqm[block->index]) { + needs = WQM; + propagate_wqm = true; + } + if (propagate_wqm) { for (const Operand& op : instr->operands) { if (op.isTemp()) { -- 2.30.2