[SandboxVec][BottomUpVec] Add -sbvec-stop-at flag for debugging

[vporpo]

When debugging miscompiles we need a way to force-stop the vectorizer early. This helps figure out which invocation is generating incorrect code.

[llvmbot]

@llvm/pr-subscribers-vectorizers

@llvm/pr-subscribers-llvm-transforms

Author: vporpo (vporpo)

Changes

When debugging miscompiles we need a way to force-stop the vectorizer early. This helps figure out which invocation is generating incorrect code.

---

Full diff: https://github.com/llvm/llvm-project/pull/129097.diff

3 Files Affected:

• (modified) llvm/include/llvm/Transforms/Vectorize/SandboxVectorizer/Passes/BottomUpVec.h (+3)
• (modified) llvm/lib/Transforms/Vectorize/SandboxVectorizer/Passes/BottomUpVec.cpp (+10)
• (added) llvm/test/Transforms/SandboxVectorizer/stop_at.ll (+61)

diff --git a/llvm/include/llvm/Transforms/Vectorize/SandboxVectorizer/Passes/BottomUpVec.h b/llvm/include/llvm/Transforms/Vectorize/SandboxVectorizer/Passes/BottomUpVec.h
index b28e9948d6f55..2359d54213ef6 100644
--- a/llvm/include/llvm/Transforms/Vectorize/SandboxVectorizer/Passes/BottomUpVec.h
+++ b/llvm/include/llvm/Transforms/Vectorize/SandboxVectorizer/Passes/BottomUpVec.h
@@ -39,6 +39,9 @@ class BottomUpVec final : public RegionPass {
   DenseSet<Instruction *> DeadInstrCandidates;
   /// Maps scalars to vectors.
   std::unique_ptr<InstrMaps> IMaps;
+  /// Counter used for force-stopping the vectorizer after this many
+  /// invocations. Used for debugging miscompiles.
+  unsigned long BottomUpInvocationCnt = 0;
 
   /// Creates and returns a vector instruction that replaces the instructions in
   /// \p Bndl. \p Operands are the already vectorized operands.
diff --git a/llvm/lib/Transforms/Vectorize/SandboxVectorizer/Passes/BottomUpVec.cpp b/llvm/lib/Transforms/Vectorize/SandboxVectorizer/Passes/BottomUpVec.cpp
index 14438181f2602..16eff012c187a 100644
--- a/llvm/lib/Transforms/Vectorize/SandboxVectorizer/Passes/BottomUpVec.cpp
+++ b/llvm/lib/Transforms/Vectorize/SandboxVectorizer/Passes/BottomUpVec.cpp
@@ -25,6 +25,13 @@ static cl::opt<bool>
                           "emit new instructions (*very* expensive)."));
 #endif // NDEBUG
 
+static constexpr const unsigned long StopAtDisabled =
+    std::numeric_limits<unsigned long>::max();
+static cl::opt<unsigned long>
+    StopAt("sbvec-stop-at", cl::init(StopAtDisabled), cl::Hidden,
+           cl::desc("Vectorize if the invocation count is < than this. 0 "
+                    "disables vectorization."));
+
 namespace sandboxir {
 
 static SmallVector<Value *, 4> getOperand(ArrayRef<Value *> Bndl,
@@ -456,6 +463,9 @@ Value *BottomUpVec::emitVectors() {
 
 bool BottomUpVec::tryVectorize(ArrayRef<Value *> Bndl) {
   Change = false;
+  if (LLVM_UNLIKELY(BottomUpInvocationCnt++ >= StopAt &&
+                    StopAt != StopAtDisabled))
+    return false;
   DeadInstrCandidates.clear();
   Legality->clear();
   Actions.clear();
diff --git a/llvm/test/Transforms/SandboxVectorizer/stop_at.ll b/llvm/test/Transforms/SandboxVectorizer/stop_at.ll
new file mode 100644
index 0000000000000..cf276a6e39e37
--- /dev/null
+++ b/llvm/test/Transforms/SandboxVectorizer/stop_at.ll
@@ -0,0 +1,61 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -passes=sandbox-vectorizer -sbvec-vec-reg-bits=1024 -sbvec-allow-non-pow2 -sbvec-passes="bottom-up-vec<>" --sbvec-stop-at=0 %s -S | FileCheck %s --check-prefix=STOPAT0
+; RUN: opt -passes=sandbox-vectorizer -sbvec-vec-reg-bits=1024 -sbvec-allow-non-pow2 -sbvec-passes="bottom-up-vec<>" --sbvec-stop-at=1 %s -S | FileCheck %s --check-prefix=STOPAT1
+; RUN: opt -passes=sandbox-vectorizer -sbvec-vec-reg-bits=1024 -sbvec-allow-non-pow2 -sbvec-passes="bottom-up-vec<>" --sbvec-stop-at=2 %s -S | FileCheck %s --check-prefix=STOPAT2
+
+define void @widen(ptr %ptrA, ptr %ptrB) {
+; STOPAT0-LABEL: define void @widen(
+; STOPAT0-SAME: ptr [[PTRA:%.*]], ptr [[PTRB:%.*]]) {
+; STOPAT0-NEXT:    [[PTRA0:%.*]] = getelementptr float, ptr [[PTRA]], i32 0
+; STOPAT0-NEXT:    [[PTRA1:%.*]] = getelementptr float, ptr [[PTRA]], i32 1
+; STOPAT0-NEXT:    [[LDA0:%.*]] = load float, ptr [[PTRA0]], align 4
+; STOPAT0-NEXT:    [[LDA1:%.*]] = load float, ptr [[PTRA1]], align 4
+; STOPAT0-NEXT:    store float [[LDA0]], ptr [[PTRA0]], align 4
+; STOPAT0-NEXT:    store float [[LDA1]], ptr [[PTRA1]], align 4
+; STOPAT0-NEXT:    [[PTRB0:%.*]] = getelementptr float, ptr [[PTRB]], i32 0
+; STOPAT0-NEXT:    [[PTRB1:%.*]] = getelementptr float, ptr [[PTRB]], i32 1
+; STOPAT0-NEXT:    [[LDB0:%.*]] = load float, ptr [[PTRB0]], align 4
+; STOPAT0-NEXT:    [[LDB1:%.*]] = load float, ptr [[PTRB1]], align 4
+; STOPAT0-NEXT:    store float [[LDB0]], ptr [[PTRB0]], align 4
+; STOPAT0-NEXT:    store float [[LDB1]], ptr [[PTRB1]], align 4
+; STOPAT0-NEXT:    ret void
+;
+; STOPAT1-LABEL: define void @widen(
+; STOPAT1-SAME: ptr [[PTRA:%.*]], ptr [[PTRB:%.*]]) {
+; STOPAT1-NEXT:    [[PTRA0:%.*]] = getelementptr float, ptr [[PTRA]], i32 0
+; STOPAT1-NEXT:    [[VECL:%.*]] = load <2 x float>, ptr [[PTRA0]], align 4
+; STOPAT1-NEXT:    store <2 x float> [[VECL]], ptr [[PTRA0]], align 4
+; STOPAT1-NEXT:    [[PTRB0:%.*]] = getelementptr float, ptr [[PTRB]], i32 0
+; STOPAT1-NEXT:    [[PTRB1:%.*]] = getelementptr float, ptr [[PTRB]], i32 1
+; STOPAT1-NEXT:    [[LDB0:%.*]] = load float, ptr [[PTRB0]], align 4
+; STOPAT1-NEXT:    [[LDB1:%.*]] = load float, ptr [[PTRB1]], align 4
+; STOPAT1-NEXT:    store float [[LDB0]], ptr [[PTRB0]], align 4
+; STOPAT1-NEXT:    store float [[LDB1]], ptr [[PTRB1]], align 4
+; STOPAT1-NEXT:    ret void
+;
+; STOPAT2-LABEL: define void @widen(
+; STOPAT2-SAME: ptr [[PTRA:%.*]], ptr [[PTRB:%.*]]) {
+; STOPAT2-NEXT:    [[PTRA0:%.*]] = getelementptr float, ptr [[PTRA]], i32 0
+; STOPAT2-NEXT:    [[VECL:%.*]] = load <2 x float>, ptr [[PTRA0]], align 4
+; STOPAT2-NEXT:    store <2 x float> [[VECL]], ptr [[PTRA0]], align 4
+; STOPAT2-NEXT:    [[PTRB0:%.*]] = getelementptr float, ptr [[PTRB]], i32 0
+; STOPAT2-NEXT:    [[VECL1:%.*]] = load <2 x float>, ptr [[PTRB0]], align 4
+; STOPAT2-NEXT:    store <2 x float> [[VECL1]], ptr [[PTRB0]], align 4
+; STOPAT2-NEXT:    ret void
+;
+  %ptrA0 = getelementptr float, ptr %ptrA, i32 0
+  %ptrA1 = getelementptr float, ptr %ptrA, i32 1
+  %ldA0 = load float, ptr %ptrA0
+  %ldA1 = load float, ptr %ptrA1
+  store float %ldA0, ptr %ptrA0
+  store float %ldA1, ptr %ptrA1
+
+  %ptrB0 = getelementptr float, ptr %ptrB, i32 0
+  %ptrB1 = getelementptr float, ptr %ptrB, i32 1
+  %ldB0 = load float, ptr %ptrB0
+  %ldB1 = load float, ptr %ptrB1
+  store float %ldB0, ptr %ptrB0
+  store float %ldB1, ptr %ptrB1
+
+  ret void
+}