From d5b92c389133c5d587e4094af553ec345ed40045 Mon Sep 17 00:00:00 2001
From: Nadav Rotem <nrotem@apple.com>
Date: Sat, 5 Jan 2013 01:15:47 +0000
Subject: [PATCH] iLoopVectorize: Non commutative operators can be used as
 reduction variables as long as the reduction chain is used in the LHS.

PR14803.



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@171583 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Transforms/Vectorize/LoopVectorize.cpp    | 17 ++++++++++----
 test/Transforms/LoopVectorize/gcc-examples.ll |  2 +-
 test/Transforms/LoopVectorize/reduction.ll    | 32 +++++++++++++++++++++++++--
 3 files changed, 44 insertions(+), 7 deletions(-)

diff --git a/lib/Transforms/Vectorize/LoopVectorize.cpp b/lib/Transforms/Vectorize/LoopVectorize.cpp
index 5e2d7971f84..af2e8464d01 100644
--- a/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -1912,10 +1912,6 @@ bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
     if (Iter->use_empty())
       return false;
 
-    // Any reduction instr must be of one of the allowed kinds.
-    if (!isReductionInstr(Iter, Kind))
-      return false;
-
     // Did we find a user inside this loop already ?
     bool FoundInBlockUser = false;
     // Did we reach the initial PHI node already ?
@@ -1953,6 +1949,16 @@ bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
       if (FoundInBlockUser)
         return false;
       FoundInBlockUser = true;
+
+      // Any reduction instr must be of one of the allowed kinds.
+      if (!isReductionInstr(U, Kind))
+        return false;
+
+      // Reductions of instructions such as Div, and Sub is only
+      // possible if the LHS is the reduction variable.
+      if (!U->isCommutative() && U->getOperand(0) != Iter)
+        return false;
+
       Iter = U;
     }
 
@@ -1985,8 +1991,11 @@ LoopVectorizationLegality::isReductionInstr(Instruction *I,
   case Instruction::PHI:
     // possibly.
     return true;
+  case Instruction::Sub:
   case Instruction::Add:
     return Kind == IntegerAdd;
+  case Instruction::SDiv:
+  case Instruction::UDiv:
   case Instruction::Mul:
     return Kind == IntegerMult;
   case Instruction::And:
diff --git a/test/Transforms/LoopVectorize/gcc-examples.ll b/test/Transforms/LoopVectorize/gcc-examples.ll
index de3f59f41af..b8b125f3ee7 100644
--- a/test/Transforms/LoopVectorize/gcc-examples.ll
+++ b/test/Transforms/LoopVectorize/gcc-examples.ll
@@ -241,7 +241,7 @@ define void @example8(i32 %x) nounwind uwtable ssp {
 }
 
 ;CHECK: @example9
-;CHECK-NOT: phi <4 x i32>
+;CHECK: phi <4 x i32>
 ;CHECK: ret i32
 define i32 @example9() nounwind uwtable readonly ssp {
   br label %1
diff --git a/test/Transforms/LoopVectorize/reduction.ll b/test/Transforms/LoopVectorize/reduction.ll
index fbce072b8b1..6b29e2537bd 100644
--- a/test/Transforms/LoopVectorize/reduction.ll
+++ b/test/Transforms/LoopVectorize/reduction.ll
@@ -296,11 +296,12 @@ for.end:                                          ; preds = %for.body, %entry
   ret i32 %x.0.lcssa
 }
 
-;CHECK: @reduction_sub_lhs
+; In this code the subtracted variable is on the RHS and this is not an induction variable.
+;CHECK: @reduction_sub_rhs
 ;CHECK-NOT: phi <4 x i32>
 ;CHECK-NOT: sub nsw <4 x i32>
 ;CHECK: ret i32
-define i32 @reduction_sub_lhs(i32 %n, i32* noalias nocapture %A) nounwind uwtable readonly {
+define i32 @reduction_sub_rhs(i32 %n, i32* noalias nocapture %A) nounwind uwtable readonly {
 entry:
   %cmp4 = icmp sgt i32 %n, 0
   br i1 %cmp4, label %for.body, label %for.end
@@ -320,3 +321,30 @@ for.end:                                          ; preds = %for.body, %entry
   %x.0.lcssa = phi i32 [ 0, %entry ], [ %sub, %for.body ]
   ret i32 %x.0.lcssa
 }
+
+
+; In this test the reduction variable is on the LHS and we can vectorize it.
+;CHECK: @reduction_sub_lhs
+;CHECK: phi <4 x i32>
+;CHECK: sub nsw <4 x i32>
+;CHECK: ret i32
+define i32 @reduction_sub_lhs(i32 %n, i32* noalias nocapture %A) nounwind uwtable readonly {
+entry:
+  %cmp4 = icmp sgt i32 %n, 0
+  br i1 %cmp4, label %for.body, label %for.end
+
+for.body:                                         ; preds = %entry, %for.body
+  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
+  %x.05 = phi i32 [ %sub, %for.body ], [ 0, %entry ]
+  %arrayidx = getelementptr inbounds i32* %A, i64 %indvars.iv
+  %0 = load i32* %arrayidx, align 4
+  %sub = sub nsw i32 %x.05, %0
+  %indvars.iv.next = add i64 %indvars.iv, 1
+  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+  %exitcond = icmp eq i32 %lftr.wideiv, %n
+  br i1 %exitcond, label %for.end, label %for.body
+
+for.end:                                          ; preds = %for.body, %entry
+  %x.0.lcssa = phi i32 [ 0, %entry ], [ %sub, %for.body ]
+  ret i32 %x.0.lcssa
+}
-- 
2.11.0