[llvm-branch-commits] [llvm] [AMDGPU] Let -align-loops override the target loop alignment (PR #242684)
Dmitry Sidorov via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Sat Nov 14 10:49:49 PST 2026
https://github.com/MrSidims created https://github.com/llvm/llvm-project/pull/242684
On subtargets that align loops themselves, -align-loops=N removed the loop
alignment instead of setting it to N.
Let the option override the target heuristics. -align-loops=1 now disables
loop alignment, as on other targets.
---
<sub>Stack created with <a href="https://github.com/github/gh-stack">GitHub Stacks CLI</a> • <a href="https://gh.io/stacks-feedback">Give Feedback 💬</a></sub>
>From 51bf36a707aa38c605ebdafd2c0d7eea90615bfd Mon Sep 17 00:00:00 2001
From: Dmitry Sidorov <Dmitry.Sidorov at amd.com>
Date: Sat, 14 Nov 2026 14:19:37 +0100
Subject: [PATCH] [AMDGPU] Let -align-loops override the target loop alignment
On subtargets that align loops themselves, -align-loops=N removed the loop
alignment instead of setting it to N.
Let the option override the target heuristics. -align-loops=1 now disables
loop alignment, as on other targets.
---
llvm/lib/Target/AMDGPU/SIISelLowering.cpp | 6 +++++-
.../test/CodeGen/AMDGPU/loop-header-align-gfx950.mir | 12 ++++++------
llvm/test/CodeGen/AMDGPU/loop-prefetch.ll | 9 ++++++++-
3 files changed, 19 insertions(+), 8 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index c22723e391093a..9cf22b5b97df38 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -20819,6 +20819,9 @@ Align SITargetLowering::getPrefLoopAlignment(
const Align PrefAlign = TargetLowering::getPrefLoopAlignment(ML);
const Align CacheLineAlign = Align(64);
+ if (getTargetMachine().Options.LoopAlignment)
+ return PrefAlign;
+
// GFX950: Prevent an 8-byte instruction at the block being aligned from being
// split by the 32-byte instruction fetch window boundary. This avoids a
// significant fetch delay after a backward branch. We use 32-byte alignment
@@ -20913,7 +20916,8 @@ unsigned SITargetLowering::getMaxPermittedBytesForAlignment(
// GFX950: Limit padding to 4 bytes (one s_nop) for blocks where an 8-byte
// instruction could be split by the 32-byte fetch window boundary.
// See getPrefLoopAlignment() for context.
- if (needsFetchWindowAlignment(*MBB))
+ if (!getTargetMachine().Options.LoopAlignment &&
+ needsFetchWindowAlignment(*MBB))
return 4;
return TargetLowering::getMaxPermittedBytesForAlignment(MBB);
}
diff --git a/llvm/test/CodeGen/AMDGPU/loop-header-align-gfx950.mir b/llvm/test/CodeGen/AMDGPU/loop-header-align-gfx950.mir
index d4ac12b1643f39..d032cd43b48877 100644
--- a/llvm/test/CodeGen/AMDGPU/loop-header-align-gfx950.mir
+++ b/llvm/test/CodeGen/AMDGPU/loop-header-align-gfx950.mir
@@ -48,7 +48,7 @@ body: |
; ALIGN1-NEXT: {{ $}}
; ALIGN1-NEXT: renamable $sgpr0_sgpr1 = S_MOV_B64 0
; ALIGN1-NEXT: {{ $}}
- ; ALIGN1-NEXT: bb.1 (align 32, max-bytes-for-alignment 4):
+ ; ALIGN1-NEXT: bb.1:
; ALIGN1-NEXT: successors: %bb.2(0x04000000), %bb.1(0x7c000000)
; ALIGN1-NEXT: liveins: $sgpr0_sgpr1, $vgpr0_vgpr1
; ALIGN1-NEXT: {{ $}}
@@ -71,7 +71,7 @@ body: |
; ALIGN16-NEXT: {{ $}}
; ALIGN16-NEXT: renamable $sgpr0_sgpr1 = S_MOV_B64 0
; ALIGN16-NEXT: {{ $}}
- ; ALIGN16-NEXT: bb.1:
+ ; ALIGN16-NEXT: bb.1 (align 16):
; ALIGN16-NEXT: successors: %bb.2(0x04000000), %bb.1(0x7c000000)
; ALIGN16-NEXT: liveins: $sgpr0_sgpr1, $vgpr0_vgpr1
; ALIGN16-NEXT: {{ $}}
@@ -166,7 +166,7 @@ body: |
; ALIGN16-NEXT: {{ $}}
; ALIGN16-NEXT: renamable $sgpr0_sgpr1 = S_MOV_B64 0
; ALIGN16-NEXT: {{ $}}
- ; ALIGN16-NEXT: bb.1:
+ ; ALIGN16-NEXT: bb.1 (align 16):
; ALIGN16-NEXT: successors: %bb.2(0x04000000), %bb.1(0x7c000000)
; ALIGN16-NEXT: liveins: $sgpr0_sgpr1, $vgpr0
; ALIGN16-NEXT: {{ $}}
@@ -321,7 +321,7 @@ body: |
; ALIGN16-NEXT: {{ $}}
; ALIGN16-NEXT: S_BRANCH %bb.2
; ALIGN16-NEXT: {{ $}}
- ; ALIGN16-NEXT: bb.6:
+ ; ALIGN16-NEXT: bb.6 (align 16):
; ALIGN16-NEXT: successors: %bb.7(0x04000000), %bb.2(0x7c000000)
; ALIGN16-NEXT: liveins: $vgpr0_vgpr1
; ALIGN16-NEXT: {{ $}}
@@ -483,7 +483,7 @@ body: |
; ALIGN1-NEXT: {{ $}}
; ALIGN1-NEXT: S_BRANCH %bb.2
; ALIGN1-NEXT: {{ $}}
- ; ALIGN1-NEXT: bb.6 (align 32, max-bytes-for-alignment 4):
+ ; ALIGN1-NEXT: bb.6:
; ALIGN1-NEXT: successors: %bb.7(0x04000000), %bb.2(0x7c000000)
; ALIGN1-NEXT: liveins: $vgpr0_vgpr1
; ALIGN1-NEXT: {{ $}}
@@ -534,7 +534,7 @@ body: |
; ALIGN16-NEXT: {{ $}}
; ALIGN16-NEXT: S_BRANCH %bb.2
; ALIGN16-NEXT: {{ $}}
- ; ALIGN16-NEXT: bb.6:
+ ; ALIGN16-NEXT: bb.6 (align 16):
; ALIGN16-NEXT: successors: %bb.7(0x04000000), %bb.2(0x7c000000)
; ALIGN16-NEXT: liveins: $vgpr0_vgpr1
; ALIGN16-NEXT: {{ $}}
diff --git a/llvm/test/CodeGen/AMDGPU/loop-prefetch.ll b/llvm/test/CodeGen/AMDGPU/loop-prefetch.ll
index 30697437a346a4..078cbe17c204e4 100644
--- a/llvm/test/CodeGen/AMDGPU/loop-prefetch.ll
+++ b/llvm/test/CodeGen/AMDGPU/loop-prefetch.ll
@@ -1,12 +1,15 @@
; RUN: llc -mtriple=amdgpu10.30 -asm-verbose=0 < %s | FileCheck --check-prefixes=GCN,GFX10,GFX10-ASM %s
; RUN: llc -mtriple=amdgpu10.30 < %s -filetype=obj | llvm-objdump -d --triple=amdgpu10.30 --symbolize-operands - | FileCheck --check-prefixes=GCN,GFX10,GFX10-DIS %s
-; RUN: llc -mtriple=amdgpu10.30 -asm-verbose=0 -align-loops=1 < %s | FileCheck --check-prefixes=GCN,GFX10,GFX10-ASM %s
+; RUN: llc -mtriple=amdgpu10.30 -asm-verbose=0 -align-loops=1 < %s | FileCheck --check-prefix=ALIGN1 %s
; RUN: llc -mtriple=amdgpu10.30 -asm-verbose=0 -align-loops=16 < %s | FileCheck --check-prefixes=GCN,ALIGN16 %s
; RUN: llc -mtriple=amdgpu8.02 < %s | FileCheck --check-prefix=GFX8 %s
; GFX8-NOT: s_inst_prefetch
; GFX8-NOT: .palign 6
+; ALIGN1-NOT: s_inst_prefetch
+; ALIGN1-NOT: .p2align 6
+
; GCN-LABEL: test_loop_64
; GFX10: s_movk_i32 s{{[0-9]+}}, 0x400
; GFX10-DIS-NEXT: {{^$}}
@@ -16,6 +19,7 @@
; GFX10: s_cbranch_scc0 [[L1]]
; GFX10-NEXT: s_endpgm
; ALIGN16: s_movk_i32 s{{[0-9]+}}, 0x400
+; ALIGN16-NEXT: .p2align 4
; ALIGN16-NEXT: [[L1:.LBB[0-9_]+]]:
; ALIGN16: s_cbranch_scc0 [[L1]]
; ALIGN16-NEXT: s_endpgm
@@ -45,6 +49,7 @@ bb2: ; preds = %bb2, %bb
; GFX10: s_cbranch_scc0 [[L1]]
; GFX10-NEXT: s_endpgm
; ALIGN16: s_movk_i32 s{{[0-9]+}}, 0x400
+; ALIGN16-NEXT: .p2align 4
; ALIGN16-NEXT: [[L1:.LBB[0-9_]+]]:
; ALIGN16: s_cbranch_scc0 [[L1]]
; ALIGN16-NEXT: s_endpgm
@@ -91,6 +96,7 @@ bb2: ; preds = %bb2, %bb
; GFX10-NEXT: s_inst_prefetch 0x2
; GFX10-NEXT: s_endpgm
; ALIGN16: s_movk_i32 s{{[0-9]+}}, 0x400
+; ALIGN16-NEXT: .p2align 4
; ALIGN16-NEXT: [[L1:.LBB[0-9_]+]]:
; ALIGN16: s_cbranch_scc0 [[L1]]
; ALIGN16-NEXT: s_endpgm
@@ -151,6 +157,7 @@ bb2: ; preds = %bb2, %bb
; GFX10: s_cbranch_scc0 [[L1]]
; GFX10-NEXT: s_endpgm
; ALIGN16: s_movk_i32 s{{[0-9]+}}, 0x400
+; ALIGN16-NEXT: .p2align 4
; ALIGN16-NEXT: [[L1:.LBB[0-9_]+]]:
; ALIGN16: s_cbranch_scc0 [[L1]]
; ALIGN16-NEXT: s_endpgm
More information about the llvm-branch-commits
mailing list