Add #pragma vectorize enable/disable to LLVM

The intended behaviour is to force vectorization on the presence
of the flag (either turn on or off), and to continue the behaviour
as expected in its absence. Tests were added to make sure the all
cases are covered in opt. No tests were added in other tools with
the assumption that they should use the PassManagerBuilder in the
same way.

This patch also removes the outdated -late-vectorize flag, which was
on by default and not helping much.

The pragma metadata is being attached to the same place as other loop
metadata, but nothing forbids one from attaching it to a function
(to enable #pragma optimize) or basic blocks (to hint the basic-block
vectorizers), etc. The logic should be the same all around.

Patches to Clang to produce the metadata will be produced after the
initial implementation is agreed upon and committed. Patches to other
vectorizers (such as SLP and BB) will be added once we're happy with
the pass manager changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@196537 91177308-0d34-0410-b5e6-96231b3b80d8

include/llvm/Transforms/IPO/PassManagerBuilder.h

@@ -111,7 +111,6 @@ public:
   bool BBVectorize;
   bool SLPVectorize;
   bool LoopVectorize;
-  bool LateVectorize;
   bool RerollLoops;
 
 private:

include/llvm/Transforms/Vectorize.h

@@ -114,7 +114,8 @@ createBBVectorizePass(const VectorizeConfig &C = VectorizeConfig());
 //
 // LoopVectorize - Create a loop vectorization pass.
 //
-Pass *createLoopVectorizePass(bool NoUnrolling = false);
+Pass *createLoopVectorizePass(bool NoUnrolling = false,
+                              bool AlwaysVectorize = true);
 
 //===----------------------------------------------------------------------===//
 //

lib/Transforms/IPO/PassManagerBuilder.cpp

@@ -32,11 +32,6 @@ static cl::opt<bool>
 RunLoopVectorization("vectorize-loops", cl::Hidden,
                      cl::desc("Run the Loop vectorization passes"));
 
-static cl::opt<bool>
-LateVectorization("late-vectorize", cl::init(true), cl::Hidden,
-                  cl::desc("Run the vectorization pasess late in the pass "
-                           "pipeline (after the inliner)"));
-
 static cl::opt<bool>
 RunSLPVectorization("vectorize-slp", cl::Hidden,
                     cl::desc("Run the SLP vectorization passes"));
@@ -68,7 +63,6 @@ PassManagerBuilder::PassManagerBuilder() {
     BBVectorize = RunBBVectorization;
     SLPVectorize = RunSLPVectorization;
     LoopVectorize = RunLoopVectorization;
-    LateVectorize = LateVectorization;
     RerollLoops = RunLoopRerolling;
 }
 
@@ -200,9 +194,6 @@ void PassManagerBuilder::populateModulePassManager(PassManagerBase &MPM) {
   MPM.add(createLoopIdiomPass());             // Recognize idioms like memset.
   MPM.add(createLoopDeletionPass());          // Delete dead loops
 
-  if (!LateVectorize && LoopVectorize)
-      MPM.add(createLoopVectorizePass(DisableUnrollLoops));
-
   if (!DisableUnrollLoops)
     MPM.add(createLoopUnrollPass());          // Unroll small loops
   addExtensionsToPM(EP_LoopOptimizerEnd, MPM);
@@ -243,21 +234,18 @@ void PassManagerBuilder::populateModulePassManager(PassManagerBase &MPM) {
   MPM.add(createCFGSimplificationPass()); // Merge & remove BBs
   MPM.add(createInstructionCombiningPass());  // Clean up after everything.
 
-  // As an experimental mode, run any vectorization passes in a separate
+  // FIXME: This is a HACK! The inliner pass above implicitly creates a CGSCC
-  // pipeline from the CGSCC pass manager that runs iteratively with the
+  // pass manager that we are specifically trying to avoid. To prevent this
-  // inliner.
+  // we must insert a no-op module pass to reset the pass manager.
-  if (LateVectorize && LoopVectorize) {
+  MPM.add(createBarrierNoopPass());
-    // FIXME: This is a HACK! The inliner pass above implicitly creates a CGSCC
+  MPM.add(createLoopVectorizePass(DisableUnrollLoops, LoopVectorize));
-    // pass manager that we are specifically trying to avoid. To prevent this
+  // FIXME: Because of #pragma vectorize enable, the passes below are always
-    // we must insert a no-op module pass to reset the pass manager.
+  // inserted in the pipeline, even when the vectorizer doesn't run (ex. when
-    MPM.add(createBarrierNoopPass());
+  // on -O1 and no #pragma is found). Would be good to have these two passes
-
+  // as function calls, so that we can only pass them when the vectorizer
-    // Add the various vectorization passes and relevant cleanup passes for
+  // changed the code.
-    // them since we are no longer in the middle of the main scalar pipeline.
+  MPM.add(createInstructionCombiningPass());
-    MPM.add(createLoopVectorizePass(DisableUnrollLoops));
+  MPM.add(createCFGSimplificationPass());
-    MPM.add(createInstructionCombiningPass());
-    MPM.add(createCFGSimplificationPass());
-  }
 
   if (!DisableUnitAtATime) {
     // FIXME: We shouldn't bother with this anymore.

lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -763,10 +763,13 @@ struct LoopVectorizeHints {
   unsigned Width;
   /// Vectorization unroll factor.
   unsigned Unroll;
+  /// Vectorization forced (-1 not selected, 0 force disabled, 1 force enabled)
+  int Force;
 
   LoopVectorizeHints(const Loop *L, bool DisableUnrolling)
   : Width(VectorizationFactor)
   , Unroll(DisableUnrolling ? 1 : VectorizationUnroll)
+  , Force(-1)
   , LoopID(L->getLoopID()) {
     getHints(L);
     // The command line options override any loop metadata except for when
@@ -877,6 +880,11 @@ private:
         Unroll = Val;
       else
         DEBUG(dbgs() << "LV: ignoring invalid unroll hint metadata\n");
+    } else if (Hint == "enable") {
+      if (C->getBitWidth() == 1)
+        Force = Val;
+      else
+        DEBUG(dbgs() << "LV: ignoring invalid enable hint metadata\n");
     } else {
       DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint << '\n');
     }
@@ -888,8 +896,10 @@ struct LoopVectorize : public LoopPass {
   /// Pass identification, replacement for typeid
   static char ID;
 
-  explicit LoopVectorize(bool NoUnrolling = false)
+  explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
-    : LoopPass(ID), DisableUnrolling(NoUnrolling) {
+    : LoopPass(ID),
+      DisableUnrolling(NoUnrolling),
+      AlwaysVectorize(AlwaysVectorize) {
     initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
   }
 
@@ -900,6 +910,7 @@ struct LoopVectorize : public LoopPass {
   DominatorTree *DT;
   TargetLibraryInfo *TLI;
   bool DisableUnrolling;
+  bool AlwaysVectorize;
 
   virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
     // We only vectorize innermost loops.
@@ -919,7 +930,7 @@ struct LoopVectorize : public LoopPass {
       return false;
 
     if (DL == NULL) {
-      DEBUG(dbgs() << "LV: Not vectorizing because of missing data layout\n");
+      DEBUG(dbgs() << "LV: Not vectorizing: Missing data layout\n");
       return false;
     }
 
@@ -928,15 +939,25 @@ struct LoopVectorize : public LoopPass {
 
     LoopVectorizeHints Hints(L, DisableUnrolling);
 
+    if (Hints.Force == 0) {
+      DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
+      return false;
+    }
+
+    if (!AlwaysVectorize && Hints.Force != 1) {
+      DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
+      return false;
+    }
+
     if (Hints.Width == 1 && Hints.Unroll == 1) {
-      DEBUG(dbgs() << "LV: Not vectorizing.\n");
+      DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
       return false;
     }
 
     // Check if it is legal to vectorize the loop.
     LoopVectorizationLegality LVL(L, SE, DL, DT, TLI);
     if (!LVL.canVectorize()) {
-      DEBUG(dbgs() << "LV: Not vectorizing.\n");
+      DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
       return false;
     }
 
@@ -949,7 +970,8 @@ struct LoopVectorize : public LoopPass {
     Attribute::AttrKind SzAttr = Attribute::OptimizeForSize;
     Attribute::AttrKind FlAttr = Attribute::NoImplicitFloat;
     unsigned FnIndex = AttributeSet::FunctionIndex;
-    bool OptForSize = F->getAttributes().hasAttribute(FnIndex, SzAttr);
+    bool OptForSize = Hints.Force != 1 &&
+                      F->getAttributes().hasAttribute(FnIndex, SzAttr);
     bool NoFloat = F->getAttributes().hasAttribute(FnIndex, FlAttr);
 
     if (NoFloat) {
@@ -973,6 +995,7 @@ struct LoopVectorize : public LoopPass {
       DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
       if (UF == 1)
         return false;
+      DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
       // We decided not to vectorize, but we may want to unroll.
       InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
       Unroller.vectorize(&LVL);
@@ -5016,8 +5039,8 @@ INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
 
 namespace llvm {
-  Pass *createLoopVectorizePass(bool NoUnrolling) {
+  Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
-    return new LoopVectorize(NoUnrolling);
+    return new LoopVectorize(NoUnrolling, AlwaysVectorize);
   }
 }
 

test/Transforms/LoopVectorize/metadata-enable.ll

@@ -0,0 +1,175 @@
+; RUN: opt < %s -O1 -S | FileCheck %s --check-prefix=O1
+; RUN: opt < %s -O2 -S | FileCheck %s --check-prefix=O2
+; RUN: opt < %s -O3 -S | FileCheck %s --check-prefix=O3
+; RUN: opt < %s -Os -S | FileCheck %s --check-prefix=Os
+; RUN: opt < %s -Oz -S | FileCheck %s --check-prefix=Oz
+; RUN: opt < %s -O1 -vectorize-loops -S | FileCheck %s --check-prefix=O1VEC
+; RUN: opt < %s -Oz -vectorize-loops -S | FileCheck %s --check-prefix=OzVEC
+; RUN: opt < %s -O1 -loop-vectorize -S | FileCheck %s --check-prefix=O1VEC2
+; RUN: opt < %s -Oz -loop-vectorize -S | FileCheck %s --check-prefix=OzVEC2
+; RUN: opt < %s -O3 -disable-loop-vectorization -S | FileCheck %s --check-prefix=O3DIS
+
+; This file tests the llvm.vectorizer.pragma forcing vectorization even when
+; optimization levels are too low, or when vectorization is disabled.
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+; O1-LABEL: @enabled(
+; O1: store <4 x i32>
+; O1: ret i32
+; O2-LABEL: @enabled(
+; O2: store <4 x i32>
+; O2: ret i32
+; O3-LABEL: @enabled(
+; O3: store <4 x i32>
+; O3: ret i32
+; Pragma always wins!
+; O3DIS-LABEL: @enabled(
+; O3DIS: store <4 x i32>
+; O3DIS: ret i32
+; Os-LABEL: @enabled(
+; Os: store <4 x i32>
+; Os: ret i32
+; Oz-LABEL: @enabled(
+; Oz: store <4 x i32>
+; Oz: ret i32
+; O1VEC-LABEL: @enabled(
+; O1VEC: store <4 x i32>
+; O1VEC: ret i32
+; OzVEC-LABEL: @enabled(
+; OzVEC: store <4 x i32>
+; OzVEC: ret i32
+; O1VEC2-LABEL: @enabled(
+; O1VEC2: store <4 x i32>
+; O1VEC2: ret i32
+; OzVEC2-LABEL: @enabled(
+; OzVEC2: store <4 x i32>
+; OzVEC2: ret i32
+
+define i32 @enabled(i32* noalias nocapture %a, i32* noalias nocapture readonly %b, i32 %N) {
+entry:
+  br label %for.body
+
+for.body:                                         ; preds = %for.body, %entry
+  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
+  %arrayidx = getelementptr inbounds i32* %b, i64 %indvars.iv
+  %0 = load i32* %arrayidx, align 4
+  %add = add nsw i32 %0, %N
+  %arrayidx2 = getelementptr inbounds i32* %a, i64 %indvars.iv
+  store i32 %add, i32* %arrayidx2, align 4
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  %exitcond = icmp eq i64 %indvars.iv.next, 32
+  br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
+
+for.end:                                          ; preds = %for.body
+  %1 = load i32* %a, align 4
+  ret i32 %1
+}
+
+; O1-LABEL: @nopragma(
+; O1-NOT: store <4 x i32>
+; O1: ret i32
+; O2-LABEL: @nopragma(
+; O2: store <4 x i32>
+; O2: ret i32
+; O3-LABEL: @nopragma(
+; O3: store <4 x i32>
+; O3: ret i32
+; O3DIS-LABEL: @nopragma(
+; O3DIS-NOT: store <4 x i32>
+; O3DIS: ret i32
+; Os-LABEL: @nopragma(
+; Os: store <4 x i32>
+; Os: ret i32
+; Oz-LABEL: @nopragma(
+; Oz-NOT: store <4 x i32>
+; Oz: ret i32
+; O1VEC-LABEL: @nopragma(
+; O1VEC: store <4 x i32>
+; O1VEC: ret i32
+; OzVEC-LABEL: @nopragma(
+; OzVEC: store <4 x i32>
+; OzVEC: ret i32
+; O1VEC2-LABEL: @nopragma(
+; O1VEC2: store <4 x i32>
+; O1VEC2: ret i32
+; OzVEC2-LABEL: @nopragma(
+; OzVEC2: store <4 x i32>
+; OzVEC2: ret i32
+
+define i32 @nopragma(i32* noalias nocapture %a, i32* noalias nocapture readonly %b, i32 %N) {
+entry:
+  br label %for.body
+
+for.body:                                         ; preds = %for.body, %entry
+  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
+  %arrayidx = getelementptr inbounds i32* %b, i64 %indvars.iv
+  %0 = load i32* %arrayidx, align 4
+  %add = add nsw i32 %0, %N
+  %arrayidx2 = getelementptr inbounds i32* %a, i64 %indvars.iv
+  store i32 %add, i32* %arrayidx2, align 4
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  %exitcond = icmp eq i64 %indvars.iv.next, 32
+  br i1 %exitcond, label %for.end, label %for.body
+
+for.end:                                          ; preds = %for.body
+  %1 = load i32* %a, align 4
+  ret i32 %1
+}
+
+; O1-LABEL: @disabled(
+; O1-NOT: store <4 x i32>
+; O1: ret i32
+; O2-LABEL: @disabled(
+; O2-NOT: store <4 x i32>
+; O2: ret i32
+; O3-LABEL: @disabled(
+; O3-NOT: store <4 x i32>
+; O3: ret i32
+; O3DIS-LABEL: @disabled(
+; O3DIS-NOT: store <4 x i32>
+; O3DIS: ret i32
+; Os-LABEL: @disabled(
+; Os-NOT: store <4 x i32>
+; Os: ret i32
+; Oz-LABEL: @disabled(
+; Oz-NOT: store <4 x i32>
+; Oz: ret i32
+; O1VEC-LABEL: @disabled(
+; O1VEC-NOT: store <4 x i32>
+; O1VEC: ret i32
+; OzVEC-LABEL: @disabled(
+; OzVEC-NOT: store <4 x i32>
+; OzVEC: ret i32
+; O1VEC2-LABEL: @disabled(
+; O1VEC2-NOT: store <4 x i32>
+; O1VEC2: ret i32
+; OzVEC2-LABEL: @disabled(
+; OzVEC2-NOT: store <4 x i32>
+; OzVEC2: ret i32
+
+define i32 @disabled(i32* noalias nocapture %a, i32* noalias nocapture readonly %b, i32 %N) {
+entry:
+  br label %for.body
+
+for.body:                                         ; preds = %for.body, %entry
+  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
+  %arrayidx = getelementptr inbounds i32* %b, i64 %indvars.iv
+  %0 = load i32* %arrayidx, align 4
+  %add = add nsw i32 %0, %N
+  %arrayidx2 = getelementptr inbounds i32* %a, i64 %indvars.iv
+  store i32 %add, i32* %arrayidx2, align 4
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  %exitcond = icmp eq i64 %indvars.iv.next, 32
+  br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !2
+
+for.end:                                          ; preds = %for.body
+  %1 = load i32* %a, align 4
+  ret i32 %1
+}
+
+!0 = metadata !{metadata !0, metadata !1}
+!1 = metadata !{metadata !"llvm.vectorizer.enable", i1 1}
+!2 = metadata !{metadata !2, metadata !3}
+!3 = metadata !{metadata !"llvm.vectorizer.enable", i1 0}

tools/opt/opt.cpp

@@ -471,8 +471,14 @@ static void AddOptimizationPasses(PassManagerBase &MPM,FunctionPassManager &FPM,
   Builder.DisableUnrollLoops = (DisableLoopUnrolling.getNumOccurrences() > 0) ?
                                DisableLoopUnrolling : OptLevel == 0;
 
-  Builder.LoopVectorize =
+  // This is final, unless there is a #pragma vectorize enable
-      DisableLoopVectorization ? false : OptLevel > 1 && SizeLevel < 2;
+  if (DisableLoopVectorization)
+    Builder.LoopVectorize = false;
+  // If option wasn't forced via cmd line (-vectorize-loops, -loop-vectorize)
+  else if (!Builder.LoopVectorize)
+    Builder.LoopVectorize = OptLevel > 1 && SizeLevel < 2;
+
+  // When #pragma vectorize is on for SLP, do the same as above
   Builder.SLPVectorize =
       DisableSLPVectorization ? false : OptLevel > 1 && SizeLevel < 2;