[llvm] LSV: forbid load-cycles when vectorizing; fix bug (PR #104815)

Ramkumar Ramachandra via llvm-commits llvm-commits at lists.llvm.org
Tue Aug 20 12:04:07 PDT 2024


https://github.com/artagnon updated https://github.com/llvm/llvm-project/pull/104815

>From 6bfb08b3559bfbfce6643e588bcd6a832aa0c96d Mon Sep 17 00:00:00 2001
From: Ramkumar Ramachandra <ramkumar.ramachandra at codasip.com>
Date: Mon, 19 Aug 2024 16:53:31 +0100
Subject: [PATCH 1/3] LSV: forbid load-cycles when vectorizing; fix bug

Forbid zero SCEV diff in getConstantOffset, hence forbidding cycles
which would crash LoadStoreVectorizer when vectorizing.

Fixes #37865.
---
 .../Vectorize/LoadStoreVectorizer.cpp           |  6 ++++++
 .../LoadStoreVectorizer/AArch64/pr37865.ll      | 17 +++++++++++++----
 2 files changed, 19 insertions(+), 4 deletions(-)

diff --git a/llvm/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp b/llvm/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp
index c35ea431296b70..385349b338af93 100644
--- a/llvm/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp
@@ -1499,6 +1499,12 @@ std::optional<APInt> Vectorizer::getConstantOffset(Value *PtrA, Value *PtrB,
 
   // Try to compute B - A.
   const SCEV *DistScev = SE.getMinusSCEV(SE.getSCEV(PtrB), SE.getSCEV(PtrA));
+  if (DistScev->isZero()) {
+    // A load in the chain is dependent on another load in the chain, and
+    // attempting to vectorize this chain would create a cycle.
+    LLVM_DEBUG(dbgs() << "LSV: SCEV diff is zero; not vectorizing\n");
+    return std::nullopt;
+  }
   if (DistScev != SE.getCouldNotCompute()) {
     LLVM_DEBUG(dbgs() << "LSV: SCEV PtrB - PtrA =" << *DistScev << "\n");
     ConstantRange DistRange = SE.getSignedRange(DistScev);
diff --git a/llvm/test/Transforms/LoadStoreVectorizer/AArch64/pr37865.ll b/llvm/test/Transforms/LoadStoreVectorizer/AArch64/pr37865.ll
index 833e70814c2917..bf5d8a555bd908 100644
--- a/llvm/test/Transforms/LoadStoreVectorizer/AArch64/pr37865.ll
+++ b/llvm/test/Transforms/LoadStoreVectorizer/AArch64/pr37865.ll
@@ -1,9 +1,18 @@
-; REQUIRES: asserts
-; RUN: not --crash opt -mtriple=aarch64 -passes=load-store-vectorizer \
-; RUN:   -disable-output %s 2>&1 | FileCheck %s
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -mtriple=aarch64 -passes=load-store-vectorizer -S %s | FileCheck %s
 
 define i32 @load_cycle(ptr %x) {
-; CHECK: Unexpected cycle while re-ordering instructions
+; CHECK-LABEL: define i32 @load_cycle(
+; CHECK-SAME: ptr [[X:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[GEP_X_1:%.*]] = getelementptr inbounds [2 x i32], ptr [[X]], i32 0, i32 1
+; CHECK-NEXT:    [[LOAD_X_1:%.*]] = load i32, ptr [[GEP_X_1]], align 4
+; CHECK-NEXT:    [[REM:%.*]] = urem i32 [[LOAD_X_1]], 1
+; CHECK-NEXT:    [[GEP_X_2:%.*]] = getelementptr inbounds [2 x i32], ptr [[X]], i32 [[REM]], i32 0
+; CHECK-NEXT:    [[LOAD_X_2:%.*]] = load i32, ptr [[GEP_X_2]], align 4
+; CHECK-NEXT:    [[RET:%.*]] = add i32 [[LOAD_X_2]], [[LOAD_X_1]]
+; CHECK-NEXT:    ret i32 [[RET]]
+;
 entry:
   %gep.x.1 = getelementptr inbounds [2 x i32], ptr %x, i32 0, i32 1
   %load.x.1 = load i32, ptr %gep.x.1

>From 4465fef283cc626ec26a8352cd06d450ae5f5659 Mon Sep 17 00:00:00 2001
From: Ramkumar Ramachandra <ramkumar.ramachandra at codasip.com>
Date: Mon, 19 Aug 2024 18:11:43 +0100
Subject: [PATCH 2/3] LSV: address review

---
 llvm/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp       | 1 +
 llvm/test/Transforms/LoadStoreVectorizer/AArch64/pr37865.ll | 4 ++++
 2 files changed, 5 insertions(+)

diff --git a/llvm/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp b/llvm/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp
index 385349b338af93..187a6d3641dead 100644
--- a/llvm/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp
@@ -1505,6 +1505,7 @@ std::optional<APInt> Vectorizer::getConstantOffset(Value *PtrA, Value *PtrB,
     LLVM_DEBUG(dbgs() << "LSV: SCEV diff is zero; not vectorizing\n");
     return std::nullopt;
   }
+
   if (DistScev != SE.getCouldNotCompute()) {
     LLVM_DEBUG(dbgs() << "LSV: SCEV PtrB - PtrA =" << *DistScev << "\n");
     ConstantRange DistRange = SE.getSignedRange(DistScev);
diff --git a/llvm/test/Transforms/LoadStoreVectorizer/AArch64/pr37865.ll b/llvm/test/Transforms/LoadStoreVectorizer/AArch64/pr37865.ll
index bf5d8a555bd908..d657ea03cff82e 100644
--- a/llvm/test/Transforms/LoadStoreVectorizer/AArch64/pr37865.ll
+++ b/llvm/test/Transforms/LoadStoreVectorizer/AArch64/pr37865.ll
@@ -1,6 +1,10 @@
 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
 ; RUN: opt -mtriple=aarch64 -passes=load-store-vectorizer -S %s | FileCheck %s
 
+; LSV was attempting to vectorize this earlier, but crashed while re-ordering
+; instructions due to the load-load cycle. Now, the chain is no longer a
+; candidate for vectorization.
+
 define i32 @load_cycle(ptr %x) {
 ; CHECK-LABEL: define i32 @load_cycle(
 ; CHECK-SAME: ptr [[X:%.*]]) {

>From cb2ab8578ebe44bf6157ceb64605aee0c8155cee Mon Sep 17 00:00:00 2001
From: Ramkumar Ramachandra <ramkumar.ramachandra at codasip.com>
Date: Tue, 20 Aug 2024 19:36:49 +0100
Subject: [PATCH 3/3] LSV: redo patch

---
 .../Vectorize/LoadStoreVectorizer.cpp         | 34 +++++++---
 .../LoadStoreVectorizer/AArch64/pr37865.ll    | 62 ++++++++++++++++++-
 2 files changed, 85 insertions(+), 11 deletions(-)

diff --git a/llvm/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp b/llvm/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp
index 187a6d3641dead..888bedf700d7fc 100644
--- a/llvm/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp
@@ -996,10 +996,32 @@ bool Vectorizer::isSafeToMove(
   LLVM_DEBUG(dbgs() << "LSV: isSafeToMove(" << *ChainElem << " -> "
                     << *ChainBegin << ")\n");
 
-  assert(isa<LoadInst>(ChainElem) == IsLoadChain);
+  assert(isa<LoadInst>(ChainElem) == IsLoadChain &&
+         isa<LoadInst>(ChainBegin) == IsLoadChain);
+
   if (ChainElem == ChainBegin)
     return true;
 
+  if constexpr (IsLoadChain) {
+    // If ChainElem depends on ChainBegin, they're not safe to reorder.
+    SmallVector<Instruction *, 8> Worklist;
+    Worklist.emplace_back(ChainElem);
+    while (!Worklist.empty()) {
+      Instruction *I = Worklist.pop_back_val();
+      for (Use &O : I->operands()) {
+        if (isa<PHINode>(O))
+          continue;
+        if (auto *J = dyn_cast<Instruction>(O)) {
+          if (J == ChainBegin) {
+            LLVM_DEBUG(dbgs() << "LSV: dependent loads; not safe to reorder\n");
+            return false;
+          }
+          Worklist.emplace_back(J);
+        }
+      }
+    }
+  }
+
   // Invariant loads can always be reordered; by definition they are not
   // clobbered by stores.
   if (isInvariantLoad(ChainElem))
@@ -1028,7 +1050,8 @@ bool Vectorizer::isSafeToMove(
     if (!I->mayReadOrWriteMemory())
       continue;
 
-    // Loads can be reordered with other loads.
+    // Loads can be reordered with other loads. We've already checked that
+    // they're independent.
     if (IsLoadChain && isa<LoadInst>(I))
       continue;
 
@@ -1499,13 +1522,6 @@ std::optional<APInt> Vectorizer::getConstantOffset(Value *PtrA, Value *PtrB,
 
   // Try to compute B - A.
   const SCEV *DistScev = SE.getMinusSCEV(SE.getSCEV(PtrB), SE.getSCEV(PtrA));
-  if (DistScev->isZero()) {
-    // A load in the chain is dependent on another load in the chain, and
-    // attempting to vectorize this chain would create a cycle.
-    LLVM_DEBUG(dbgs() << "LSV: SCEV diff is zero; not vectorizing\n");
-    return std::nullopt;
-  }
-
   if (DistScev != SE.getCouldNotCompute()) {
     LLVM_DEBUG(dbgs() << "LSV: SCEV PtrB - PtrA =" << *DistScev << "\n");
     ConstantRange DistRange = SE.getSignedRange(DistScev);
diff --git a/llvm/test/Transforms/LoadStoreVectorizer/AArch64/pr37865.ll b/llvm/test/Transforms/LoadStoreVectorizer/AArch64/pr37865.ll
index d657ea03cff82e..0beca8c15305f6 100644
--- a/llvm/test/Transforms/LoadStoreVectorizer/AArch64/pr37865.ll
+++ b/llvm/test/Transforms/LoadStoreVectorizer/AArch64/pr37865.ll
@@ -2,8 +2,8 @@
 ; RUN: opt -mtriple=aarch64 -passes=load-store-vectorizer -S %s | FileCheck %s
 
 ; LSV was attempting to vectorize this earlier, but crashed while re-ordering
-; instructions due to the load-load cycle. Now, the chain is no longer a
-; candidate for vectorization.
+; instructions due to the load-load cycle. Now, the candidate loads are no
+; longer considered safe for reordering.
 
 define i32 @load_cycle(ptr %x) {
 ; CHECK-LABEL: define i32 @load_cycle(
@@ -26,3 +26,61 @@ entry:
   %ret = add i32 %load.x.2, %load.x.1
   ret i32 %ret
 }
+
+define i32 @load_cycle2(ptr %x, i32 %y) {
+; CHECK-LABEL: define i32 @load_cycle2(
+; CHECK-SAME: ptr [[X:%.*]], i32 [[Y:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[GEP_X_1:%.*]] = getelementptr inbounds [2 x i32], ptr [[X]], i32 [[Y]], i32 1
+; CHECK-NEXT:    [[LOAD_X_1:%.*]] = load i32, ptr [[GEP_X_1]], align 4
+; CHECK-NEXT:    [[MUL:%.*]] = mul i32 [[LOAD_X_1]], 2
+; CHECK-NEXT:    [[ADD:%.*]] = add i32 [[Y]], [[MUL]]
+; CHECK-NEXT:    [[SUB_1:%.*]] = sub i32 [[ADD]], [[LOAD_X_1]]
+; CHECK-NEXT:    [[SUB_2:%.*]] = sub i32 [[SUB_1]], [[LOAD_X_1]]
+; CHECK-NEXT:    [[GEP_X_2:%.*]] = getelementptr inbounds [2 x i32], ptr [[X]], i32 [[SUB_2]], i32 0
+; CHECK-NEXT:    [[LOAD_X_2:%.*]] = load i32, ptr [[GEP_X_2]], align 4
+; CHECK-NEXT:    [[RET:%.*]] = add i32 [[LOAD_X_2]], [[LOAD_X_1]]
+; CHECK-NEXT:    ret i32 [[RET]]
+;
+entry:
+  %gep.x.1 = getelementptr inbounds [2 x i32], ptr %x, i32 %y, i32 1
+  %load.x.1 = load i32, ptr %gep.x.1
+  %mul = mul i32 %load.x.1, 2
+  %add = add i32 %y, %mul
+  %sub.1 = sub i32 %add, %load.x.1
+  %sub.2 = sub i32 %sub.1, %load.x.1
+  %gep.x.2 = getelementptr inbounds [2 x i32], ptr %x, i32 %sub.2, i32 0
+  %load.x.2 = load i32, ptr %gep.x.2
+  %ret = add i32 %load.x.2, %load.x.1
+  ret i32 %ret
+}
+
+ at global.1 = global i32 0
+ at global.2 = global [1 x [3 x i32]] zeroinitializer
+
+define i16 @load_cycle3() {
+; CHECK-LABEL: define i16 @load_cycle3() {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[LOAD_1:%.*]] = load i32, ptr @global.1, align 4
+; CHECK-NEXT:    [[UREM_1:%.*]] = urem i32 [[LOAD_1]], 1
+; CHECK-NEXT:    [[GEP_1:%.*]] = getelementptr inbounds [1 x [3 x i32]], ptr @global.2, i32 0, i32 [[UREM_1]]
+; CHECK-NEXT:    [[GEP_2:%.*]] = getelementptr inbounds [3 x i32], ptr [[GEP_1]], i32 0, i32 2
+; CHECK-NEXT:    [[LOAD_2:%.*]] = load i32, ptr [[GEP_2]], align 4
+; CHECK-NEXT:    [[UREM_2:%.*]] = urem i32 [[LOAD_2]], 1
+; CHECK-NEXT:    [[GEP_3:%.*]] = getelementptr inbounds [1 x [3 x i32]], ptr @global.2, i32 0, i32 [[UREM_2]]
+; CHECK-NEXT:    [[GEP_4:%.*]] = getelementptr inbounds [3 x i32], ptr [[GEP_3]], i32 0, i32 1
+; CHECK-NEXT:    [[LOAD_3:%.*]] = load i32, ptr [[GEP_4]], align 4
+; CHECK-NEXT:    ret i16 0
+;
+entry:
+  %load.1 = load i32, ptr @global.1
+  %urem.1 = urem i32 %load.1, 1
+  %gep.1 = getelementptr inbounds [1 x [3 x i32]], ptr @global.2, i32 0, i32 %urem.1
+  %gep.2 = getelementptr inbounds [3 x i32], ptr %gep.1, i32 0, i32 2
+  %load.2 = load i32, ptr %gep.2
+  %urem.2 = urem i32 %load.2, 1
+  %gep.3 = getelementptr inbounds [1 x [3 x i32]], ptr @global.2, i32 0, i32 %urem.2
+  %gep.4 = getelementptr inbounds [3 x i32], ptr %gep.3, i32 0, i32 1
+  %load.3 = load i32, ptr %gep.4
+  ret i16 0
+}



More information about the llvm-commits mailing list