[DependenceAnalysis] Extend unifySubscriptType for handling coupled subscript groups.

Summary:
In continuation to an earlier commit to DependenceAnalysis.cpp by jingyue (r222100), the type for all subscripts in a coupled group need to be the same since constraints from one subscript may be propagated to another during testing. During testing, new SCEVs may be created and the operands for these need to be the same.
This patch extends unifySubscriptType() to work on lists of subscript pairs, ensuring a common extended type for all of them.

Test Plan:
Added a test case to NonCanonicalizedSubscript.ll which causes dependence analysis to crash without this fix.

All regression tests pass.

Reviewers: spop, sebpop, jingyue

Reviewed By: jingyue

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D9698

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@238573 91177308-0d34-0410-b5e6-96231b3b80d8

include/llvm/Analysis/DependenceAnalysis.h

@@ -41,6 +41,7 @@
 #define LLVM_ANALYSIS_DEPENDENCEANALYSIS_H
 
 #include "llvm/ADT/SmallBitVector.h"
+#include "llvm/ADT/ArrayRef.h"
 #include "llvm/IR/Instructions.h"
 #include "llvm/Pass.h"
 
@@ -520,11 +521,11 @@ namespace llvm {
     /// in LoopNest.
     bool isLoopInvariant(const SCEV *Expression, const Loop *LoopNest) const;
 
-    /// Makes sure both subscripts (i.e. Pair->Src and Pair->Dst) share the same
+    /// Makes sure all subscript pairs share the same integer type by 
-    /// integer type by sign-extending one of them when necessary.
+    /// sign-extending as necessary.
     /// Sign-extending a subscript is safe because getelementptr assumes the
-    /// array subscripts are signed.
+    /// array subscripts are signed. 
-    void unifySubscriptType(Subscript *Pair);
+    void unifySubscriptType(ArrayRef<Subscript *> Pairs);
 
     /// removeMatchingExtensions - Examines a subscript pair.
     /// If the source and destination are identically sign (or zero)

lib/Analysis/DependenceAnalysis.cpp

@@ -779,23 +779,56 @@ void DependenceAnalysis::collectCommonLoops(const SCEV *Expression,
   }
 }
 
-void DependenceAnalysis::unifySubscriptType(Subscript *Pair) {
+void DependenceAnalysis::unifySubscriptType(ArrayRef<Subscript *> Pairs) {
-  const SCEV *Src = Pair->Src;
+
-  const SCEV *Dst = Pair->Dst;
+  unsigned widestWidthSeen = 0;
-  IntegerType *SrcTy = dyn_cast<IntegerType>(Src->getType());
+  Type *widestType;
-  IntegerType *DstTy = dyn_cast<IntegerType>(Dst->getType());
+
-  if (SrcTy == nullptr || DstTy == nullptr) {
+  // Go through each pair and find the widest bit to which we need
-    assert(SrcTy == DstTy && "This function only unify integer types and "
+  // to extend all of them.
-                             "expect Src and Dst share the same type "
+  for (unsigned i = 0; i < Pairs.size(); i++) {
-                             "otherwise.");
+    const SCEV *Src = Pairs[i]->Src;
-    return;
+    const SCEV *Dst = Pairs[i]->Dst;
+    IntegerType *SrcTy = dyn_cast<IntegerType>(Src->getType());
+    IntegerType *DstTy = dyn_cast<IntegerType>(Dst->getType());
+    if (SrcTy == nullptr || DstTy == nullptr) {
+      assert(SrcTy == DstTy && "This function only unify integer types and "
+             "expect Src and Dst share the same type "
+             "otherwise.");
+      continue;
+    }
+    if (SrcTy->getBitWidth() > widestWidthSeen) {
+      widestWidthSeen = SrcTy->getBitWidth();
+      widestType = SrcTy;
+    }
+    if (DstTy->getBitWidth() > widestWidthSeen) {
+      widestWidthSeen = DstTy->getBitWidth();
+      widestType = DstTy;
+    }
   }
-  if (SrcTy->getBitWidth() > DstTy->getBitWidth()) {
+
-    // Sign-extend Dst to typeof(Src) if typeof(Src) is wider than typeof(Dst).
+
-    Pair->Dst = SE->getSignExtendExpr(Dst, SrcTy);
+  assert(widestWidthSeen > 0);
-  } else if (SrcTy->getBitWidth() < DstTy->getBitWidth()) {
+
-    // Sign-extend Src to typeof(Dst) if typeof(Dst) is wider than typeof(Src).
+  // Now extend each pair to the widest seen.
-    Pair->Src = SE->getSignExtendExpr(Src, DstTy);
+  for (unsigned i = 0; i < Pairs.size(); i++) {
+    const SCEV *Src = Pairs[i]->Src;
+    const SCEV *Dst = Pairs[i]->Dst;
+    IntegerType *SrcTy = dyn_cast<IntegerType>(Src->getType());
+    IntegerType *DstTy = dyn_cast<IntegerType>(Dst->getType());
+    if (SrcTy == nullptr || DstTy == nullptr) {
+      assert(SrcTy == DstTy && "This function only unify integer types and "
+             "expect Src and Dst share the same type "
+             "otherwise.");
+      continue;
+    }
+    if (SrcTy->getBitWidth() < widestWidthSeen)
+      // Sign-extend Src to widestType
+      Pairs[i]->Src = SE->getSignExtendExpr(Src, widestType);
+    if (DstTy->getBitWidth() < widestWidthSeen) {
+      // Sign-extend Dst to widestType
+      Pairs[i]->Dst = SE->getSignExtendExpr(Dst, widestType);
+    }
   }
 }
 
@@ -2937,7 +2970,7 @@ const SCEV *DependenceAnalysis::getUpperBound(BoundInfo *Bound) const {
 // return the coefficient (the step)
 // corresponding to the specified loop.
 // If there isn't one, return 0.
-// For example, given a*i + b*j + c*k, zeroing the coefficient
+// For example, given a*i + b*j + c*k, finding the coefficient
 // corresponding to the j loop would yield b.
 const SCEV *DependenceAnalysis::findCoefficient(const SCEV *Expr,
                                                 const Loop *TargetLoop)  const {
@@ -3574,13 +3607,16 @@ DependenceAnalysis::depends(Instruction *Src, Instruction *Dst,
       SmallBitVector Sivs(Pairs);
       SmallBitVector Mivs(Pairs);
       SmallBitVector ConstrainedLevels(MaxLevels + 1);
+      SmallVector<Subscript *, 4> PairsInGroup;
       for (int SJ = Group.find_first(); SJ >= 0; SJ = Group.find_next(SJ)) {
         DEBUG(dbgs() << SJ << " ");
         if (Pair[SJ].Classification == Subscript::SIV)
           Sivs.set(SJ);
         else
           Mivs.set(SJ);
+        PairsInGroup.push_back(&Pair[SJ]);
       }
+      unifySubscriptType(PairsInGroup);
       DEBUG(dbgs() << "}\n");
       while (Sivs.any()) {
         bool Changed = false;

test/Analysis/DependenceAnalysis/NonCanonicalizedSubscript.ll

@@ -38,3 +38,73 @@ for.body:
 for.end:
   ret void
 }
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+;  unsigned i, j;
+;  for (i = 1; i < SIZE; i++) {
+;    for (j = i; j < SIZE; j++) {
+;      a[i][j] = a[i+1][j-1] + 2;
+;    }
+;  }
+;  Extends the previous example to coupled MIV subscripts.
+
+
+@a = global [10004 x [10004 x i32]] zeroinitializer, align 16
+
+; Function Attrs: nounwind uwtable
+define void @coupled_miv_type_mismatch(i32 %n) #0 {
+; CHECK-LABEL: 'Dependence Analysis' for function 'coupled_miv_type_mismatch'
+; DELIN-LABEL: 'Dependence Analysis' for function 'coupled_miv_type_mismatch'
+entry:
+  br label %for.cond
+
+; CHECK: da analyze - input [0 *]!
+; CHECK: da analyze - anti [1 *]!
+; CHECK: da analyze - none!
+; DELIN: da analyze - input [0 *]!
+; DELIN: da analyze - anti [1 *]!
+; DELIN: da analyze - none!
+for.cond:                                         ; preds = %for.inc11, %entry
+  %indvars.iv11 = phi i64 [ %indvars.iv.next12, %for.inc11 ], [ 1, %entry ]
+  %exitcond14 = icmp ne i64 %indvars.iv11, 10000
+  br i1 %exitcond14, label %for.cond1.preheader, label %for.end13
+
+for.cond1.preheader:                              ; preds = %for.cond
+  %0 = trunc i64 %indvars.iv11 to i32
+  br label %for.cond1
+
+for.cond1:                                        ; preds = %for.cond1.preheader, %for.body3
+  %indvars.iv8 = phi i64 [ %indvars.iv11, %for.cond1.preheader ], [ %indvars.iv.next9, %for.body3 ]
+  %j.0 = phi i32 [ %inc, %for.body3 ], [ %0, %for.cond1.preheader ]
+  %lftr.wideiv = trunc i64 %indvars.iv8 to i32
+  %exitcond = icmp ne i32 %lftr.wideiv, 10000
+  br i1 %exitcond, label %for.body3, label %for.inc11
+
+for.body3:                                        ; preds = %for.cond1
+  %sub = add nsw i32 %j.0, -1
+  %idxprom = zext i32 %sub to i64
+  %1 = add nuw nsw i64 %indvars.iv11, 1
+  %arrayidx5 = getelementptr inbounds [10004 x [10004 x i32]], [10004 x [10004 x i32]]* @a, i64 0, i64 %1, i64 %idxprom
+  %2 = load i32, i32* %arrayidx5, align 4
+  %add6 = add nsw i32 %2, 2
+  %arrayidx10 = getelementptr inbounds [10004 x [10004 x i32]], [10004 x [10004 x i32]]* @a, i64 0, i64 %indvars.iv11, i64 %indvars.iv8
+  store i32 %add6, i32* %arrayidx10, align 4
+  %indvars.iv.next9 = add nuw nsw i64 %indvars.iv8, 1
+  %inc = add nuw nsw i32 %j.0, 1
+  br label %for.cond1
+
+for.inc11:                                        ; preds = %for.cond1
+  %indvars.iv.next12 = add nuw nsw i64 %indvars.iv11, 1
+  br label %for.cond
+
+for.end13:                                        ; preds = %for.cond
+  ret void
+}
+
+attributes #0 = { nounwind uwtable "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+sse,+sse2" "unsafe-fp-math"="false" "use-soft-float"="false" }
+
+!llvm.ident = !{!0}
+
+!0 = !{!"clang version 3.7.0 (https://vaivaswatha@bitbucket.org/compilertree/amd_clang.git 93a05fb75ee3411d24e8b2b184fc766a5318403e) (https://vaivaswatha@bitbucket.org/compilertree/amd_llvm.git 166d93d26efc912b517739f64d054a435e8e95cd)"}