[llvm] Unroll loops apple (PR #149358)
Ahmad Yasin via llvm-commits
llvm-commits at lists.llvm.org
Wed Jul 23 02:51:23 PDT 2025
https://github.com/ayasin-a updated https://github.com/llvm/llvm-project/pull/149358
>From 7b46981ceca0054efdedf6c52ebb784018471ef0 Mon Sep 17 00:00:00 2001
From: Ahmad Yasin <ahmad.yasin at apple.com>
Date: Wed, 9 Jul 2025 16:27:30 +0300
Subject: [PATCH 1/3] tune budget for runtime-unroll
---
llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
index 90d3d92d6bbf5..1eb6589ed8ca9 100644
--- a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
+++ b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
@@ -4808,7 +4808,7 @@ getAppleRuntimeUnrollPreferences(Loop *L, ScalarEvolution &SE,
if (Header == L->getLoopLatch()) {
// Estimate the size of the loop.
unsigned Size;
- if (!isLoopSizeWithinBudget(L, TTI, 8, &Size))
+ if (!isLoopSizeWithinBudget(L, TTI, 9, &Size))
return;
SmallPtrSet<Value *, 8> LoadedValues;
>From eb21acc1bca702f935f408e64a0fb34f4870ea37 Mon Sep 17 00:00:00 2001
From: Ahmad Yasin <ahmad.yasin at apple.com>
Date: Wed, 9 Jul 2025 22:21:59 +0300
Subject: [PATCH 2/3] Include immediate users of loaded values in the
load/stores dependencies predicate
---
.../AArch64/AArch64TargetTransformInfo.cpp | 19 +++++++++++++------
1 file changed, 13 insertions(+), 6 deletions(-)
diff --git a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
index 1eb6589ed8ca9..6d97ae7c8c5e7 100644
--- a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
+++ b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
@@ -4811,7 +4811,7 @@ getAppleRuntimeUnrollPreferences(Loop *L, ScalarEvolution &SE,
if (!isLoopSizeWithinBudget(L, TTI, 9, &Size))
return;
- SmallPtrSet<Value *, 8> LoadedValues;
+ SmallPtrSet<Value *, 8> LoadedValuesPlus;
SmallVector<StoreInst *> Stores;
for (auto *BB : L->blocks()) {
for (auto &I : *BB) {
@@ -4821,9 +4821,16 @@ getAppleRuntimeUnrollPreferences(Loop *L, ScalarEvolution &SE,
const SCEV *PtrSCEV = SE.getSCEV(Ptr);
if (SE.isLoopInvariant(PtrSCEV, L))
continue;
- if (isa<LoadInst>(&I))
- LoadedValues.insert(&I);
- else
+ if (isa<LoadInst>(&I)) {
+ LoadedValuesPlus.insert(&I);
+ // Included 1st users of loaded values
+ for (auto *U : I.users()) {
+ auto *Inst = dyn_cast<Instruction>(U);
+ if (!Inst || Inst->getParent() != BB)
+ continue;
+ LoadedValuesPlus.insert(U);
+ }
+ } else
Stores.push_back(cast<StoreInst>(&I));
}
}
@@ -4846,8 +4853,8 @@ getAppleRuntimeUnrollPreferences(Loop *L, ScalarEvolution &SE,
UC++;
}
- if (BestUC == 1 || none_of(Stores, [&LoadedValues](StoreInst *SI) {
- return LoadedValues.contains(SI->getOperand(0));
+ if (BestUC == 1 || none_of(Stores, [&LoadedValuesPlus](StoreInst *SI) {
+ return LoadedValuesPlus.contains(SI->getOperand(0));
}))
return;
>From f20211fb4c15a4e5a47e531bbb776beb605ae104 Mon Sep 17 00:00:00 2001
From: Ahmad Yasin <ahmad.yasin at apple.com>
Date: Wed, 23 Jul 2025 12:51:07 +0300
Subject: [PATCH 3/3] a test for the load-op-store scenario permitted by this
PR
---
.../LoopUnroll/AArch64/apple-unrolling.ll | 120 ++++++++++++++++++
1 file changed, 120 insertions(+)
diff --git a/llvm/test/Transforms/LoopUnroll/AArch64/apple-unrolling.ll b/llvm/test/Transforms/LoopUnroll/AArch64/apple-unrolling.ll
index 1a091e847ca34..5304b43601910 100644
--- a/llvm/test/Transforms/LoopUnroll/AArch64/apple-unrolling.ll
+++ b/llvm/test/Transforms/LoopUnroll/AArch64/apple-unrolling.ll
@@ -165,6 +165,126 @@ exit:
ret void
}
+define void @load_op_store_loop(ptr %src, ptr %dst, i64 %N, i64 %scale, float %k) {
+; APPLE-LABEL: define void @load_op_store_loop(
+; APPLE-SAME: ptr [[SRC:%.*]], ptr [[DST:%.*]], i64 [[N:%.*]], i64 [[SCALE:%.*]], float [[K:%.*]]) #[[ATTR0]] {
+; APPLE-NEXT: [[ENTRY:.*]]:
+; APPLE-NEXT: [[TMP0:%.*]] = add i64 [[N]], -1
+; APPLE-NEXT: [[XTRAITER:%.*]] = and i64 [[N]], 1
+; APPLE-NEXT: [[TMP1:%.*]] = icmp ult i64 [[TMP0]], 1
+; APPLE-NEXT: br i1 [[TMP1]], label %[[EXIT_UNR_LCSSA:.*]], label %[[ENTRY_NEW:.*]]
+; APPLE: [[ENTRY_NEW]]:
+; APPLE-NEXT: [[UNROLL_ITER:%.*]] = sub i64 [[N]], [[XTRAITER]]
+; APPLE-NEXT: br label %[[LOOP:.*]]
+; APPLE: [[LOOP]]:
+; APPLE-NEXT: [[IV:%.*]] = phi i64 [ 0, %[[ENTRY_NEW]] ], [ [[IV_NEXT_1:%.*]], %[[LOOP]] ]
+; APPLE-NEXT: [[NITER:%.*]] = phi i64 [ 0, %[[ENTRY_NEW]] ], [ [[NITER_NEXT_1:%.*]], %[[LOOP]] ]
+; APPLE-NEXT: [[SCALED_IV:%.*]] = mul nuw nsw i64 [[IV]], [[SCALE]]
+; APPLE-NEXT: [[GEP_SRC:%.*]] = getelementptr inbounds float, ptr [[SRC]], i64 [[SCALED_IV]]
+; APPLE-NEXT: [[L:%.*]] = load float, ptr [[GEP_SRC]], align 4
+; APPLE-NEXT: [[O:%.*]] = fadd float [[L]], [[K]]
+; APPLE-NEXT: [[GEP_DST:%.*]] = getelementptr inbounds float, ptr [[DST]], i64 [[IV]]
+; APPLE-NEXT: store float [[O]], ptr [[GEP_DST]], align 4
+; APPLE-NEXT: [[IV_NEXT:%.*]] = add nuw nsw i64 [[IV]], 1
+; APPLE-NEXT: [[SCALED_IV_1:%.*]] = mul nuw nsw i64 [[IV_NEXT]], [[SCALE]]
+; APPLE-NEXT: [[GEP_SRC_1:%.*]] = getelementptr inbounds float, ptr [[SRC]], i64 [[SCALED_IV_1]]
+; APPLE-NEXT: [[L_1:%.*]] = load float, ptr [[GEP_SRC_1]], align 4
+; APPLE-NEXT: [[O_1:%.*]] = fadd float [[L_1]], [[K]]
+; APPLE-NEXT: [[GEP_DST_1:%.*]] = getelementptr inbounds float, ptr [[DST]], i64 [[IV_NEXT]]
+; APPLE-NEXT: store float [[O_1]], ptr [[GEP_DST_1]], align 4
+; APPLE-NEXT: [[IV_NEXT_1]] = add nuw nsw i64 [[IV]], 2
+; APPLE-NEXT: [[NITER_NEXT_1]] = add i64 [[NITER]], 2
+; APPLE-NEXT: [[NITER_NCMP_1:%.*]] = icmp eq i64 [[NITER_NEXT_1]], [[UNROLL_ITER]]
+; APPLE-NEXT: br i1 [[NITER_NCMP_1]], label %[[EXIT_UNR_LCSSA_LOOPEXIT:.*]], label %[[LOOP]]
+; APPLE: [[EXIT_UNR_LCSSA_LOOPEXIT]]:
+; APPLE-NEXT: [[IV_UNR_PH:%.*]] = phi i64 [ [[IV_NEXT_1]], %[[LOOP]] ]
+; APPLE-NEXT: br label %[[EXIT_UNR_LCSSA]]
+; APPLE: [[EXIT_UNR_LCSSA]]:
+; APPLE-NEXT: [[IV_UNR:%.*]] = phi i64 [ 0, %[[ENTRY]] ], [ [[IV_UNR_PH]], %[[EXIT_UNR_LCSSA_LOOPEXIT]] ]
+; APPLE-NEXT: [[LCMP_MOD:%.*]] = icmp ne i64 [[XTRAITER]], 0
+; APPLE-NEXT: br i1 [[LCMP_MOD]], label %[[LOOP_EPIL_PREHEADER:.*]], label %[[EXIT:.*]]
+; APPLE: [[LOOP_EPIL_PREHEADER]]:
+; APPLE-NEXT: br label %[[LOOP_EPIL:.*]]
+; APPLE: [[LOOP_EPIL]]:
+; APPLE-NEXT: [[SCALED_IV_EPIL:%.*]] = mul nuw nsw i64 [[IV_UNR]], [[SCALE]]
+; APPLE-NEXT: [[GEP_SRC_EPIL:%.*]] = getelementptr inbounds float, ptr [[SRC]], i64 [[SCALED_IV_EPIL]]
+; APPLE-NEXT: [[L_EPIL:%.*]] = load float, ptr [[GEP_SRC_EPIL]], align 4
+; APPLE-NEXT: [[O_EPIL:%.*]] = fadd float [[L_EPIL]], [[K]]
+; APPLE-NEXT: [[GEP_DST_EPIL:%.*]] = getelementptr inbounds float, ptr [[DST]], i64 [[IV_UNR]]
+; APPLE-NEXT: store float [[O_EPIL]], ptr [[GEP_DST_EPIL]], align 4
+; APPLE-NEXT: br label %[[EXIT]]
+; APPLE: [[EXIT]]:
+; APPLE-NEXT: ret void
+;
+; OTHER-LABEL: define void @load_op_store_loop(
+; OTHER-SAME: ptr [[SRC:%.*]], ptr [[DST:%.*]], i64 [[N:%.*]], i64 [[SCALE:%.*]], float [[K:%.*]]) #[[ATTR0]] {
+; OTHER-NEXT: [[ENTRY:.*]]:
+; OTHER-NEXT: [[TMP0:%.*]] = add i64 [[N]], -1
+; OTHER-NEXT: [[XTRAITER:%.*]] = and i64 [[N]], 1
+; OTHER-NEXT: [[TMP1:%.*]] = icmp ult i64 [[TMP0]], 1
+; OTHER-NEXT: br i1 [[TMP1]], label %[[EXIT_UNR_LCSSA:.*]], label %[[ENTRY_NEW:.*]]
+; OTHER: [[ENTRY_NEW]]:
+; OTHER-NEXT: [[UNROLL_ITER:%.*]] = sub i64 [[N]], [[XTRAITER]]
+; OTHER-NEXT: br label %[[LOOP:.*]]
+; OTHER: [[LOOP]]:
+; OTHER-NEXT: [[IV:%.*]] = phi i64 [ 0, %[[ENTRY_NEW]] ], [ [[IV_NEXT_1:%.*]], %[[LOOP]] ]
+; OTHER-NEXT: [[NITER:%.*]] = phi i64 [ 0, %[[ENTRY_NEW]] ], [ [[NITER_NEXT_1:%.*]], %[[LOOP]] ]
+; OTHER-NEXT: [[SCALED_IV:%.*]] = mul nuw nsw i64 [[IV]], [[SCALE]]
+; OTHER-NEXT: [[GEP_SRC:%.*]] = getelementptr inbounds float, ptr [[SRC]], i64 [[SCALED_IV]]
+; OTHER-NEXT: [[L:%.*]] = load float, ptr [[GEP_SRC]], align 4
+; OTHER-NEXT: [[O:%.*]] = fadd float [[L]], [[K]]
+; OTHER-NEXT: [[GEP_DST:%.*]] = getelementptr inbounds float, ptr [[DST]], i64 [[IV]]
+; OTHER-NEXT: store float [[O]], ptr [[GEP_DST]], align 4
+; OTHER-NEXT: [[IV_NEXT:%.*]] = add nuw nsw i64 [[IV]], 1
+; OTHER-NEXT: [[SCALED_IV_1:%.*]] = mul nuw nsw i64 [[IV_NEXT]], [[SCALE]]
+; OTHER-NEXT: [[GEP_SRC_1:%.*]] = getelementptr inbounds float, ptr [[SRC]], i64 [[SCALED_IV_1]]
+; OTHER-NEXT: [[L_1:%.*]] = load float, ptr [[GEP_SRC_1]], align 4
+; OTHER-NEXT: [[O_1:%.*]] = fadd float [[L_1]], [[K]]
+; OTHER-NEXT: [[GEP_DST_1:%.*]] = getelementptr inbounds float, ptr [[DST]], i64 [[IV_NEXT]]
+; OTHER-NEXT: store float [[O_1]], ptr [[GEP_DST_1]], align 4
+; OTHER-NEXT: [[IV_NEXT_1]] = add nuw nsw i64 [[IV]], 2
+; OTHER-NEXT: [[NITER_NEXT_1]] = add i64 [[NITER]], 2
+; OTHER-NEXT: [[NITER_NCMP_1:%.*]] = icmp eq i64 [[NITER_NEXT_1]], [[UNROLL_ITER]]
+; OTHER-NEXT: br i1 [[NITER_NCMP_1]], label %[[EXIT_UNR_LCSSA_LOOPEXIT:.*]], label %[[LOOP]]
+; OTHER: [[EXIT_UNR_LCSSA_LOOPEXIT]]:
+; OTHER-NEXT: [[IV_UNR_PH:%.*]] = phi i64 [ [[IV_NEXT_1]], %[[LOOP]] ]
+; OTHER-NEXT: br label %[[EXIT_UNR_LCSSA]]
+; OTHER: [[EXIT_UNR_LCSSA]]:
+; OTHER-NEXT: [[IV_UNR:%.*]] = phi i64 [ 0, %[[ENTRY]] ], [ [[IV_UNR_PH]], %[[EXIT_UNR_LCSSA_LOOPEXIT]] ]
+; OTHER-NEXT: [[LCMP_MOD:%.*]] = icmp ne i64 [[XTRAITER]], 0
+; OTHER-NEXT: br i1 [[LCMP_MOD]], label %[[LOOP_EPIL_PREHEADER:.*]], label %[[EXIT:.*]]
+; OTHER: [[LOOP_EPIL_PREHEADER]]:
+; OTHER-NEXT: br label %[[LOOP_EPIL:.*]]
+; OTHER: [[LOOP_EPIL]]:
+; OTHER-NEXT: [[SCALED_IV_EPIL:%.*]] = mul nuw nsw i64 [[IV_UNR]], [[SCALE]]
+; OTHER-NEXT: [[GEP_SRC_EPIL:%.*]] = getelementptr inbounds float, ptr [[SRC]], i64 [[SCALED_IV_EPIL]]
+; OTHER-NEXT: [[L_EPIL:%.*]] = load float, ptr [[GEP_SRC_EPIL]], align 4
+; OTHER-NEXT: [[O_EPIL:%.*]] = fadd float [[L_EPIL]], [[K]]
+; OTHER-NEXT: [[GEP_DST_EPIL:%.*]] = getelementptr inbounds float, ptr [[DST]], i64 [[IV_UNR]]
+; OTHER-NEXT: store float [[O_EPIL]], ptr [[GEP_DST_EPIL]], align 4
+; OTHER-NEXT: br label %[[EXIT]]
+; OTHER: [[EXIT]]:
+; OTHER-NEXT: ret void
+;
+entry:
+ br label %loop
+
+loop:
+ %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
+ %scaled.iv = mul nuw nsw i64 %iv, %scale
+ %gep.src = getelementptr inbounds float, ptr %src, i64 %scaled.iv
+ %l = load float, ptr %gep.src, align 4
+ %o = fadd float %l, %k
+ %gep.dst = getelementptr inbounds float, ptr %dst, i64 %iv
+ store float %o, ptr %gep.dst, align 4
+ %iv.next = add nuw nsw i64 %iv, 1
+ %ec = icmp eq i64 %iv.next, %N
+ br i1 %ec, label %exit, label %loop
+
+exit:
+ ret void
+}
+
@A = external constant [9 x i8], align 1
@B = external constant [8 x i32], align 4
@C = external constant [8 x i32], align 4
More information about the llvm-commits
mailing list