[SandboxVec][BottomUpVec] Add cost estimation and tr-accept-or-revert pass

The TransactionAcceptOrRevert pass is the final pass in the Sandbox Vectorizer's
default pass pipeline. It's job is to check the cost before/after vectorization
and accept or revert the IR to its original state.

Since we are now starting the transaction in BottomUpVec, tests that run a
custom pipeline need to accept the transaction. This is done with the help
of the TransactionAlwaysAccept pass (tr-accept).

Author: Vasileios Porpodas

llvm/include/llvm/SandboxIR/Tracker.h

@@ -473,6 +473,8 @@ class Tracker {
 
   ~Tracker();
   Context &getContext() const { return Ctx; }
+  /// \Returns true if there are no changes tracked.
+  bool empty() const { return Changes.empty(); }
   /// Record \p Change and take ownership. This is the main function used to
   /// track Sandbox IR changes.
   void track(std::unique_ptr<IRChangeBase> &&Change) {

llvm/include/llvm/Transforms/Vectorize/SandboxVectorizer/Passes/TransactionAcceptOrRevert.h

@@ -0,0 +1,30 @@
+//===- TransactionAcceptOrRevert.h ------------------------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This is a region pass that checks the region cost before/after vectorization
+// and accepts the state of Sandbox IR if the cost is better, or otherwise
+// reverts it.
+//
+
+#ifndef LLVM_TRANSFORMS_VECTORIZE_SANDBOXVECTORIZER_PASSES_TRANSACTIONACCEPTORREVERT_H
+#define LLVM_TRANSFORMS_VECTORIZE_SANDBOXVECTORIZER_PASSES_TRANSACTIONACCEPTORREVERT_H
+
+#include "llvm/SandboxIR/Pass.h"
+#include "llvm/SandboxIR/Region.h"
+
+namespace llvm::sandboxir {
+
+class TransactionAcceptOrRevert : public RegionPass {
+public:
+  TransactionAcceptOrRevert() : RegionPass("tr-accept-or-revert") {}
+  bool runOnRegion(Region &Rgn, const Analyses &A) final;
+};
+
+} // namespace llvm::sandboxir
+
+#endif // LLVM_TRANSFORMS_VECTORIZE_SANDBOXVECTORIZER_PASSES_TRANSACTIONACCEPTORREVERT_H

llvm/include/llvm/Transforms/Vectorize/SandboxVectorizer/Passes/TransactionAlwaysAccept.h

@@ -0,0 +1,34 @@
+//===- TransactionAlwaysAccept.h --------------------------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This is a region pass that always accepts the transaction without checking
+// its cost. This is mainly used as a final pass in lit tests.
+//
+
+#ifndef LLVM_TRANSFORMS_VECTORIZE_SANDBOXVECTORIZER_PASSES_TRANSACTIONALWAYSACCEPT_H
+#define LLVM_TRANSFORMS_VECTORIZE_SANDBOXVECTORIZER_PASSES_TRANSACTIONALWAYSACCEPT_H
+
+#include "llvm/SandboxIR/Pass.h"
+#include "llvm/SandboxIR/Region.h"
+
+namespace llvm::sandboxir {
+
+class TransactionAlwaysAccept : public RegionPass {
+public:
+  TransactionAlwaysAccept() : RegionPass("tr-accept") {}
+  bool runOnRegion(Region &Rgn, const Analyses &A) final {
+    auto &Tracker = Rgn.getContext().getTracker();
+    bool HasChanges = !Tracker.empty();
+    Tracker.accept();
+    return HasChanges;
+  }
+};
+
+} // namespace llvm::sandboxir
+
+#endif // LLVM_TRANSFORMS_VECTORIZE_SANDBOXVECTORIZER_PASSES_TRANSACTIONALWAYSACCEPT_H

llvm/lib/Transforms/Vectorize/CMakeLists.txt

@@ -9,6 +9,7 @@ add_llvm_component_library(LLVMVectorize
   SandboxVectorizer/Legality.cpp
   SandboxVectorizer/Passes/BottomUpVec.cpp
   SandboxVectorizer/Passes/RegionsFromMetadata.cpp
+  SandboxVectorizer/Passes/TransactionAcceptOrRevert.cpp
   SandboxVectorizer/SandboxVectorizer.cpp
   SandboxVectorizer/SandboxVectorizerPassBuilder.cpp
   SandboxVectorizer/Scheduler.cpp

llvm/lib/Transforms/Vectorize/SandboxVectorizer/Passes/BottomUpVec.cpp

@@ -12,6 +12,7 @@
 #include "llvm/SandboxIR/Function.h"
 #include "llvm/SandboxIR/Instruction.h"
 #include "llvm/SandboxIR/Module.h"
+#include "llvm/SandboxIR/Region.h"
 #include "llvm/SandboxIR/Utils.h"
 #include "llvm/Transforms/Vectorize/SandboxVectorizer/SandboxVectorizerPassBuilder.h"
 #include "llvm/Transforms/Vectorize/SandboxVectorizer/SeedCollector.h"
@@ -448,13 +449,24 @@ bool BottomUpVec::runOnFunction(Function &F, const Analyses &A) {
 
           assert(SeedSlice.size() >= 2 && "Should have been rejected!");
 
-          // TODO: If vectorization succeeds, run the RegionPassManager on the
-          // resulting region.
-
           // TODO: Refactor to remove the unnecessary copy to SeedSliceVals.
           SmallVector<Value *> SeedSliceVals(SeedSlice.begin(),
                                              SeedSlice.end());
-          Change |= tryVectorize(SeedSliceVals);
+          // Create an empty region. Instructions get added to the region
+          // automatically by the callbacks.
+          auto &Ctx = F.getContext();
+          Region Rgn(Ctx, A.getTTI());
+          // Save the state of the IR before we make any changes. The
+          // transaction gets accepted/reverted by the tr-accept-or-revert pass.
+          Ctx.save();
+          // Try to vectorize starting from the seed slice. The returned value
+          // is true if we found vectorizable code and generated some vector
+          // code for it. It does not mean that the code is profitable.
+          bool VecSuccess = tryVectorize(SeedSliceVals);
+          if (VecSuccess)
+            // WARNING: All passes should return false, except those that
+            // accept/revert the state.
+            Change |= RPM.runOnRegion(Rgn, A);
         }
       }
     }

llvm/lib/Transforms/Vectorize/SandboxVectorizer/Passes/PassRegistry.def

@@ -19,6 +19,8 @@
 
 REGION_PASS("null", ::llvm::sandboxir::NullPass)
 REGION_PASS("print-instruction-count", ::llvm::sandboxir::PrintInstructionCount)
+REGION_PASS("tr-accept", ::llvm::sandboxir::TransactionAlwaysAccept)
+REGION_PASS("tr-accept-or-revert", ::llvm::sandboxir::TransactionAcceptOrRevert)
 
 #undef REGION_PASS
 

llvm/lib/Transforms/Vectorize/SandboxVectorizer/Passes/TransactionAcceptOrRevert.cpp

@@ -0,0 +1,37 @@
+//===- TransactionAcceptOrRevert.cpp - Check cost and accept/revert region ===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Transforms/Vectorize/SandboxVectorizer/Passes/TransactionAcceptOrRevert.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/InstructionCost.h"
+
+namespace llvm {
+
+static cl::opt<bool> CostThreshold("sbvec-cost-threshold", cl::init(0),
+                                   cl::Hidden,
+                                   cl::desc("Vectorization cost threshold."));
+
+namespace sandboxir {
+
+bool TransactionAcceptOrRevert::runOnRegion(Region &Rgn, const Analyses &A) {
+  const auto &SB = Rgn.getScoreboard();
+  InstructionCost CostAfterMinusBefore = SB.getAfterCost() - SB.getBeforeCost();
+  // TODO: Print costs / write to remarks.
+  auto &Tracker = Rgn.getContext().getTracker();
+  if (CostAfterMinusBefore < -CostThreshold) {
+    bool HasChanges = !Tracker.empty();
+    Tracker.accept();
+    return HasChanges;
+  }
+  // Revert the IR.
+  Rgn.getContext().getTracker().revert();
+  return false;
+}
+
+} // namespace sandboxir
+} // namespace llvm

llvm/lib/Transforms/Vectorize/SandboxVectorizer/SandboxVectorizer.cpp

@@ -31,9 +31,10 @@ static cl::opt<std::string> UserDefinedPassPipeline(
 
 SandboxVectorizerPass::SandboxVectorizerPass() : FPM("fpm") {
   if (UserDefinedPassPipeline == DefaultPipelineMagicStr) {
-    // TODO: Add region passes to the default pipeline.
+    // TODO: Add passes to the default pipeline. It currently contains:
+    //       - the bottom-up-vectorizer pass
     FPM.setPassPipeline(
-        "bottom-up-vec<>",
+        "bottom-up-vec<tr-accept-or-revert>",
         sandboxir::SandboxVectorizerPassBuilder::createFunctionPass);
   } else {
     // Create the user-defined pipeline.

llvm/lib/Transforms/Vectorize/SandboxVectorizer/SandboxVectorizerPassBuilder.cpp

@@ -4,6 +4,8 @@
 #include "llvm/Transforms/Vectorize/SandboxVectorizer/Passes/NullPass.h"
 #include "llvm/Transforms/Vectorize/SandboxVectorizer/Passes/PrintInstructionCount.h"
 #include "llvm/Transforms/Vectorize/SandboxVectorizer/Passes/RegionsFromMetadata.h"
+#include "llvm/Transforms/Vectorize/SandboxVectorizer/Passes/TransactionAcceptOrRevert.h"
+#include "llvm/Transforms/Vectorize/SandboxVectorizer/Passes/TransactionAlwaysAccept.h"
 
 namespace llvm::sandboxir {
 

llvm/test/Transforms/SandboxVectorizer/X86/simple_cost_test.ll

@@ -0,0 +1,55 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -passes=sandbox-vectorizer -mtriple=x86_64-- -mattr=+sse4.1 %s -S | FileCheck %s
+
+define void @simple_cost_test(ptr %ptr) {
+; CHECK-LABEL: define void @simple_cost_test(
+; CHECK-SAME: ptr [[PTR:%.*]]) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:    [[PTR0:%.*]] = getelementptr double, ptr [[PTR]], i32 0
+; CHECK-NEXT:    [[VECL:%.*]] = load <2 x double>, ptr [[PTR0]], align 8
+; CHECK-NEXT:    store <2 x double> [[VECL]], ptr [[PTR0]], align 8
+; CHECK-NEXT:    ret void
+;
+  %ptr0 = getelementptr double, ptr %ptr, i32 0
+  %ptr1 = getelementptr double, ptr %ptr, i32 1
+  %ld0 = load double, ptr %ptr0
+  %ld1 = load double, ptr %ptr1
+  store double %ld0, ptr %ptr0
+  store double %ld1, ptr %ptr1
+  ret void
+}
+
+define void @pack_cost_test_(ptr %ptr) {
+; CHECK-LABEL: define void @pack_cost_test_(
+; CHECK-SAME: ptr [[PTR:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:    [[PTR0:%.*]] = getelementptr float, ptr [[PTR]], i32 0
+; CHECK-NEXT:    [[PTR1:%.*]] = getelementptr float, ptr [[PTR]], i32 1
+; CHECK-NEXT:    [[LD0:%.*]] = load float, ptr [[PTR0]], align 4
+; CHECK-NEXT:    [[LD1:%.*]] = load float, ptr [[PTR1]], align 4
+; CHECK-NEXT:    [[PACK4:%.*]] = insertelement <4 x float> poison, float [[LD0]], i32 0
+; CHECK-NEXT:    [[PACK5:%.*]] = insertelement <4 x float> [[PACK4]], float [[LD1]], i32 1
+; CHECK-NEXT:    [[PACK6:%.*]] = insertelement <4 x float> [[PACK5]], float [[LD0]], i32 2
+; CHECK-NEXT:    [[PACK7:%.*]] = insertelement <4 x float> [[PACK6]], float [[LD1]], i32 3
+; CHECK-NEXT:    [[PACK:%.*]] = insertelement <4 x float> poison, float [[LD0]], i32 0
+; CHECK-NEXT:    [[PACK1:%.*]] = insertelement <4 x float> [[PACK]], float [[LD1]], i32 1
+; CHECK-NEXT:    [[PACK2:%.*]] = insertelement <4 x float> [[PACK1]], float [[LD0]], i32 2
+; CHECK-NEXT:    [[PACK3:%.*]] = insertelement <4 x float> [[PACK2]], float [[LD1]], i32 3
+; CHECK-NEXT:    [[VEC:%.*]] = fmul <4 x float> [[PACK3]], [[PACK7]]
+; CHECK-NEXT:    store <4 x float> [[VEC]], ptr [[PTR0]], align 4
+; CHECK-NEXT:    ret void
+;
+  %ptr0 = getelementptr float, ptr %ptr, i32 0
+  %ptr1 = getelementptr float, ptr %ptr, i32 1
+  %ptr2 = getelementptr float, ptr %ptr, i32 2
+  %ptr3 = getelementptr float, ptr %ptr, i32 3
+  %ld0 = load float, ptr %ptr0
+  %ld1 = load float, ptr %ptr1
+  %mul0 = fmul float %ld0, %ld0
+  %mul1 = fmul float %ld1, %ld1
+  %mul2 = fmul float %ld0, %ld0
+  %mul3 = fmul float %ld1, %ld1
+  store float %mul0, ptr %ptr0
+  store float %mul1, ptr %ptr1
+  store float %mul2, ptr %ptr2
+  store float %mul3, ptr %ptr3
+  ret void
+}