From 1f9a0490c641a6754ccaf77d627830a372ce9a18 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Daniel=20Sch=C3=BCrmann?= Date: Thu, 5 Mar 2026 22:26:19 +0100 Subject: [PATCH] nir/opt_loop: Don't peel initial break from do-while loops As the main purpose of this optimization is to transform while- into do-while loops, don't apply on loops which are already in do-while form. Also set nir_loop::do_while after this transformation, so that it is only applied once. Totals from 576 (0.28% of 202440) affected shaders: (Navi48) Instrs: 1337529 -> 1253438 (-6.29%); split: -6.36%, +0.07% CodeSize: 8390852 -> 7837328 (-6.60%); split: -6.61%, +0.01% VGPRs: 50856 -> 50844 (-0.02%) SpillSGPRs: 42198 -> 35395 (-16.12%); split: -16.13%, +0.01% SpillVGPRs: 47608 -> 44620 (-6.28%) Latency: 31043828 -> 44143753 (+42.20%); split: -0.06%, +42.26% InvThroughput: 6973433 -> 10079000 (+44.53%); split: -0.08%, +44.61% VClause: 26839 -> 24718 (-7.90%); split: -7.91%, +0.00% SClause: 21831 -> 21583 (-1.14%); split: -1.52%, +0.38% Copies: 183503 -> 150040 (-18.24%); split: -18.84%, +0.61% Branches: 27738 -> 26848 (-3.21%); split: -5.12%, +1.91% PreSGPRs: 40233 -> 39083 (-2.86%); split: -2.88%, +0.02% PreVGPRs: 38745 -> 38903 (+0.41%); split: -0.02%, +0.43% VALU: 688396 -> 645948 (-6.17%); split: -6.17%, +0.01% SALU: 189792 -> 177642 (-6.40%); split: -6.97%, +0.57% VMEM: 121500 -> 112748 (-7.20%) SMEM: 38765 -> 37767 (-2.57%); split: -2.58%, +0.00% VOPD: 102488 -> 89071 (-13.09%); split: +0.24%, -13.33% Reviewed-by: Timothy Arceri Part-of: --- src/compiler/nir/nir_opt_loop.c | 48 ++++++++++++++----- .../drivers/d3d12/ci/d3d12-quick_shader.txt | 1 - 2 files changed, 35 insertions(+), 14 deletions(-) diff --git a/src/compiler/nir/nir_opt_loop.c b/src/compiler/nir/nir_opt_loop.c index b5465354f19..bd4f44aa395 100644 --- a/src/compiler/nir/nir_opt_loop.c +++ b/src/compiler/nir/nir_opt_loop.c @@ -339,6 +339,37 @@ has_phi_with_constant_src(nir_block *block, nir_block *pred) return false; } +/* For now, we restrict this optimization to cases where the outer IF + * can be constant-folded or where at least one phi at the loop-header + * has a constant loop-carried source. If it can be constant-folded, + * we additionally require that there is actual work to be done after + * the initial break. This is to avoid unconditionally unrolling long + * loops. + * + * Note: If this restriction is lifted, it might recurse infinitely. + * Prevent by e.g. restricting to single-exit loops. + */ +static bool +should_peel_initial_break(nir_loop *loop) +{ + nir_block *header_block = nir_loop_first_block(loop); + nir_cf_node *if_node = nir_cf_node_next(&header_block->cf_node); + nir_if *nif = nir_cf_node_as_if(if_node); + + /* This loop is already in do-while form. */ + if (loop->do_while) + return false; + + /* If there is a phi with a constant loop-carried value, we can remove + * the phi by peeling one iteration. + */ + if (has_phi_with_constant_src(header_block, nir_loop_last_block(loop))) + return true; + + return nir_block_contains_work(nir_cf_node_cf_tree_next(if_node)) && + can_constant_fold(nir_get_scalar(nif->condition.ssa, 0), header_block); +} + /** * This optimization tries to peel the first loop break. * @@ -399,19 +430,7 @@ opt_loop_peel_initial_break(nir_loop *loop) if (nir_block_ends_in_jump(nir_loop_last_block(loop))) return false; - /* For now, we restrict this optimization to cases where the outer IF - * can be constant-folded or where at least one phi at the loop-header - * has a constant loop-carried source. If it can be constant-folded, - * we additionally require that there is actual work to be done after - * the initial break. This is to avoid unconditionally unrolling long - * loops. - * - * Note: If this restriction is lifted, it might recurse infinitely. - * Prevent by e.g. restricting to single-exit loops. - */ - if (!has_phi_with_constant_src(header_block, nir_loop_last_block(loop)) && - (!nir_block_contains_work(nir_cf_node_cf_tree_next(if_node)) || - !can_constant_fold(nir_get_scalar(nif->condition.ssa, 0), header_block))) + if (!should_peel_initial_break(loop)) return false; /* Even though this if statement has a jump on one side, we may still have @@ -456,6 +475,9 @@ opt_loop_peel_initial_break(nir_loop *loop) nir_after_cf_node(&loop->cf_node)); nir_cf_reinsert(&tmp, nir_after_block(nir_if_first_else_block(nif))); + /* This loop is now in do-while form. */ + loop->do_while = true; + return true; } diff --git a/src/gallium/drivers/d3d12/ci/d3d12-quick_shader.txt b/src/gallium/drivers/d3d12/ci/d3d12-quick_shader.txt index bf2e2118232..5d1a3e63db8 100644 --- a/src/gallium/drivers/d3d12/ci/d3d12-quick_shader.txt +++ b/src/gallium/drivers/d3d12/ci/d3d12-quick_shader.txt @@ -12,7 +12,6 @@ spec@glsl-1.50@execution@primitive-id-no-gs-quad-strip,Fail spec@glsl-1.50@execution@primitive-id-no-gs-quads,Fail spec@glsl-1.50@execution@primitive-id-no-gs-point,Fail spec@glsl-1.50@execution@variable-indexing@gs-output-array-vec4-index-wr,Crash -shaders@ssa@fs-if-def-else-break,Fail # These tests use a TCS output variable only as temporary storage. Since the output # is unused by the TES, we remove it.