LoopVectorizer: Perform redundancy elimination on induction variables

When the loop vectorizer was part of the SCC inliner pass manager gvn would
run after the loop vectorizer followed by instcombine. This way redundancy
(multiple uses) were removed and instcombine could perform scalarization on the
induction variables. Having moved the loop vectorizer to later we no longer run
any form of redundancy elimination before we perform instcombine. This caused
vectorized induction variables to survive that did not before.

On a recent iMac this helps linpack back from 6000Mflops to 7000Mflops.

This should also help lpbench and paq8p.

I ran a Release (without Asserts) build over the test-suite and did not see any
negative impact on compile time.

radar://15339680

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@193891 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/Vectorize/LoopVectorize.cpp b/lib/Transforms/Vectorize/LoopVectorize.cpp
index ee94173eb18ab70fba33273f420bffa27c613d9a..7f77784d28208ae6130ed33f652a1e0d37241d34 100644 (file)
--- a/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -2272,8 +2272,41 @@ InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
     (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, ReducedPartRdx);
     (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
   }// end of for each redux variable.
- 
+
   fixLCSSAPHIs();
+
+  // Perform simple cse.
+  SmallPtrSet<Instruction*, 16> Visited;
+  SmallVector<Instruction*, 16> ToRemove;
+  for (BasicBlock::iterator I = LoopVectorBody->begin(),
+       E = LoopVectorBody->end(); I != E; ++I) {
+      Instruction *In = I;
+
+      if (!isa<InsertElementInst>(In) && !isa<ExtractElementInst>(In) &&
+          !isa<ShuffleVectorInst>(In) && !isa<GetElementPtrInst>(In))
+        continue;
+
+      // Check if we can replace this instruction with any of the
+      // visited instructions.
+      for (SmallPtrSet<Instruction*, 16>::iterator v = Visited.begin(),
+           ve = Visited.end(); v != ve; ++v) {
+        if (In->isIdenticalTo(*v)) {
+          In->replaceAllUsesWith(*v);
+          ToRemove.push_back(In);
+          In = 0;
+          break;
+        }
+      }
+      if (In)
+        Visited.insert(In);
+
+  }
+  // Erase all of the instructions that we RAUWed.
+  for (SmallVectorImpl<Instruction *>::iterator v = ToRemove.begin(),
+       ve = ToRemove.end(); v != ve; ++v) {
+    assert((*v)->getNumUses() == 0 && "Can't remove instructions with uses");
+    (*v)->eraseFromParent();
+  }
 }
 
 void InnerLoopVectorizer::fixLCSSAPHIs() {

diff --git a/test/Transforms/LoopVectorize/global_alias.ll b/test/Transforms/LoopVectorize/global_alias.ll
index 4fd4c989de43b782dfcc8e2aad1cbc04cdff687d..0118fb47412a41817cf5cdc72b442cca40be26d8 100644 (file)
--- a/test/Transforms/LoopVectorize/global_alias.ll
+++ b/test/Transforms/LoopVectorize/global_alias.ll
@@ -336,9 +336,8 @@ for.end:                                          ; preds = %for.cond
 ;   return Foo.A[a];
 ; }
 ; CHECK-LABEL: define i32 @noAlias07(
-; CHECK: sub nsw <4 x i32>
+; CHECK: store <4 x i32>
 ; CHECK: ret
-
 define i32 @noAlias07(i32 %a) #0 {
 entry:
   %a.addr = alloca i32, align 4
@@ -552,7 +551,7 @@ for.end:                                          ; preds = %for.cond
 ;   return Bar.A[N][a];
 ; }
 ; CHECK-LABEL: define i32 @noAlias11(
-; CHECK: sub nsw <4 x i32>
+; CHECK: store <4 x i32>
 ; CHECK: ret
 
 define i32 @noAlias11(i32 %a) #0 {
@@ -612,7 +611,7 @@ for.end:                                          ; preds = %for.cond
 ;   return Bar.A[N][a];
 ; }
 ; CHECK-LABEL: define i32 @noAlias12(
-; CHECK: sub nsw <4 x i32>
+; CHECK: store <4 x i32>
 ; CHECK: ret
 
 define i32 @noAlias12(i32 %a) #0 {

diff --git a/test/Transforms/LoopVectorize/induction.ll b/test/Transforms/LoopVectorize/induction.ll
index d4cc50e98626a4e534058c383b1af076a08d2ffd..2471c52ac24617162a230cc86859e5147cfd35ba 100644 (file)
--- a/test/Transforms/LoopVectorize/induction.ll
+++ b/test/Transforms/LoopVectorize/induction.ll
@@ -28,3 +28,41 @@ for.end:
   ret void
 }
 
+; RUN: opt < %s -loop-vectorize -force-vector-unroll=1 -force-vector-width=2 -instcombine -S | FileCheck %s --check-prefix=IND
+
+; Make sure we remove unneeded vectorization of induction variables.
+; In order for instcombine to cleanup the vectorized induction variables that we
+; create in the loop vectorizer we need to perform some form of redundancy
+; elimination to get rid of multiple uses.
+
+; IND-LABEL: scalar_use
+
+; IND:     br label %vector.body
+; IND:     vector.body:
+;   Vectorized induction variable.
+; IND-NOT:  insertelement <2 x i64>
+; IND-NOT:  shufflevector <2 x i64>
+; IND:     br {{.*}}, label %vector.body
+
+define void @scalar_use(float* %a, float %b, i64 %offset, i64 %offset2, i64 %n) {
+entry:
+  br label %for.body
+
+for.body:
+  %iv = phi i64 [ 0, %entry ], [ %iv.next, %for.body ]
+  %ind.sum = add i64 %iv, %offset
+  %arr.idx = getelementptr inbounds float* %a, i64 %ind.sum
+  %l1 = load float* %arr.idx, align 4
+  %ind.sum2 = add i64 %iv, %offset2
+  %arr.idx2 = getelementptr inbounds float* %a, i64 %ind.sum2
+  %l2 = load float* %arr.idx2, align 4
+  %m = fmul fast float %b, %l2
+  %ad = fadd fast float %l1, %m
+  store float %ad, float* %arr.idx, align 4
+  %iv.next = add nuw nsw i64 %iv, 1
+  %exitcond = icmp eq i64 %iv.next, %n
+  br i1 %exitcond, label %loopexit, label %for.body
+
+loopexit:
+  ret void
+}

diff --git a/test/Transforms/LoopVectorize/induction_plus.ll b/test/Transforms/LoopVectorize/induction_plus.ll
index 6141c39462a6a4759de97abbc0cfab070429a5a1..9c8201ab7805374e386665df0d1e6d46f07900c2 100644 (file)
--- a/test/Transforms/LoopVectorize/induction_plus.ll
+++ b/test/Transforms/LoopVectorize/induction_plus.ll
@@ -6,8 +6,8 @@ target triple = "x86_64-apple-macosx10.8.0"
 @array = common global [1024 x i32] zeroinitializer, align 16
 
 ;CHECK-LABEL: @array_at_plus_one(
-;CHECK: trunc i64
 ;CHECK: add i64 %index, 12
+;CHECK: trunc i64
 ;CHECK: ret i32
 define i32 @array_at_plus_one(i32 %n) nounwind uwtable ssp {
   %1 = icmp sgt i32 %n, 0