Move folding of icmp with zero after checking for min/max idioms.

[ArturGainullin]

Summary

The following transformation for cmp instruction:

icmp smin(x, PositiveValue), 0 -> icmp x, 0

should only be done after checking for min/max to prevent infinite
looping caused by a reverse canonicalization. That is why this
transformation was moved to place after the mentioned check.

[ArturGainullin]

Fixed comment.

[spatel]

LGTM.

Do you have commit access?

[ArturGainullin]

Sanjay, no, I don't have. I have asked my colleague to commit. Thanks!

llvm/trunk/lib/Transforms/InstCombine/InstCombineCompares.cpp

static Instruction *processUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B,
   // The inner add was the result of the narrow add, zero extended to the
   // wider type.  Replace it with the result computed by the intrinsic.
   IC.replaceInstUsesWith(*OrigAdd, ZExt);
 
   // The original icmp gets replaced with the overflow value.
   return ExtractValueInst::Create(Call, 1, "sadd.overflow");
 }
 
+// Handle (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0)
+Instruction *InstCombiner::foldICmpWithZero(ICmpInst &Cmp) {
+  CmpInst::Predicate Pred = Cmp.getPredicate();
+  Value *X = Cmp.getOperand(0);
+
+  if (match(Cmp.getOperand(1), m_Zero()) && Pred == ICmpInst::ICMP_SGT) {
+    Value *A, *B;
+    SelectPatternResult SPR = matchSelectPattern(X, A, B);
+    if (SPR.Flavor == SPF_SMIN) {
+      if (isKnownPositive(A, DL, 0, &AC, &Cmp, &DT))
+        return new ICmpInst(Pred, B, Cmp.getOperand(1));
+      if (isKnownPositive(B, DL, 0, &AC, &Cmp, &DT))
+        return new ICmpInst(Pred, A, Cmp.getOperand(1));
+    }
+  }
+  return nullptr;
+}
+
 // Fold icmp Pred X, C.
 Instruction *InstCombiner::foldICmpWithConstant(ICmpInst &Cmp) {
   CmpInst::Predicate Pred = Cmp.getPredicate();
   Value *X = Cmp.getOperand(0);
 
   const APInt *C;
   if (!match(Cmp.getOperand(1), m_APInt(C)))
     return nullptr;
// if (sum+128 >u 255)  ...  -> llvm.sadd.with.overflow.i8
     ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI
     if (Pred == ICmpInst::ICMP_UGT &&
         match(X, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2))))
       if (Instruction *Res = processUGT_ADDCST_ADD(
               Cmp, A, B, CI2, cast<ConstantInt>(Cmp.getOperand(1)), *this))
         return Res;
   }
 
-  // (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0)
-  if (C->isNullValue() && Pred == ICmpInst::ICMP_SGT) {
-    SelectPatternResult SPR = matchSelectPattern(X, A, B);
-    if (SPR.Flavor == SPF_SMIN) {
-      if (isKnownPositive(A, DL, 0, &AC, &Cmp, &DT))
-        return new ICmpInst(Pred, B, Cmp.getOperand(1));
-      if (isKnownPositive(B, DL, 0, &AC, &Cmp, &DT))
-        return new ICmpInst(Pred, A, Cmp.getOperand(1));
-    }
-  }
-
   // FIXME: Use m_APInt to allow folds for splat constants.
   ConstantInt *CI = dyn_cast<ConstantInt>(Cmp.getOperand(1));
   if (!CI)
     return nullptr;
 
   // Canonicalize icmp instructions based on dominating conditions.
   BasicBlock *Parent = Cmp.getParent();
   BasicBlock *Dom = Parent->getSinglePredecessor();
Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
   // operands has at least one user besides the compare (the select),
   // which would often largely negate the benefit of folding anyway.
   if (I.hasOneUse())
     if (SelectInst *SI = dyn_cast<SelectInst>(*I.user_begin()))
       if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
           (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
         return nullptr;
 
+  // Do this after checking for min/max to prevent infinite looping.
+  if (Instruction *Res = foldICmpWithZero(I))
+    return Res;
+
   // FIXME: We only do this after checking for min/max to prevent infinite
   // looping caused by a reverse canonicalization of these patterns for min/max.
   // FIXME: The organization of folds is a mess. These would naturally go into
   // canonicalizeCmpWithConstant(), but we can't move all of the above folds
   // down here after the min/max restriction.
   ICmpInst::Predicate Pred = I.getPredicate();
   const APInt *C;
   if (match(Op1, m_APInt(C))) {

llvm/trunk/lib/Transforms/InstCombine/InstCombineInternal.h

private:
   Instruction *foldICmpWithCastAndCast(ICmpInst &ICI);
 
   Instruction *foldICmpUsingKnownBits(ICmpInst &Cmp);
   Instruction *foldICmpWithConstant(ICmpInst &Cmp);
   Instruction *foldICmpInstWithConstant(ICmpInst &Cmp);
   Instruction *foldICmpInstWithConstantNotInt(ICmpInst &Cmp);
   Instruction *foldICmpBinOp(ICmpInst &Cmp);
   Instruction *foldICmpEquality(ICmpInst &Cmp);
+  Instruction *foldICmpWithZero(ICmpInst &Cmp);
 
   Instruction *foldICmpSelectConstant(ICmpInst &Cmp, SelectInst *Select,
                                       ConstantInt *C);
   Instruction *foldICmpTruncConstant(ICmpInst &Cmp, TruncInst *Trunc,
                                      const APInt &C);
   Instruction *foldICmpAndConstant(ICmpInst &Cmp, BinaryOperator *And,
                                    const APInt &C);
   Instruction *foldICmpXorConstant(ICmpInst &Cmp, BinaryOperator *Xor,

llvm/trunk/test/Transforms/InstCombine/minmax-fold.ll

 
 ; Check that clamp is recognized and there is no infinite
 ; loop because of reverse cmp transformation:
 ; (icmp sgt smin(PositiveA, B) 0) -> (icmp sgt B 0)
 define i32 @clamp_check_for_no_infinite_loop1(i32 %i) {
 ; CHECK-LABEL: @clamp_check_for_no_infinite_loop1(
 ; CHECK-NEXT:    [[CMP1:%.*]] = icmp slt i32 [[I:%.*]], 255
 ; CHECK-NEXT:    [[SEL1:%.*]] = select i1 [[CMP1]], i32 [[I]], i32 255
-; CHECK-NEXT:    [[TMP1:%.*]] = icmp sgt i32 [[I]], 0
+; CHECK-NEXT:    [[TMP1:%.*]] = icmp sgt i32 [[SEL1]], 0
 ; CHECK-NEXT:    [[RES:%.*]] = select i1 [[TMP1]], i32 [[SEL1]], i32 0
 ; CHECK-NEXT:    ret i32 [[RES]]
 ;
   %cmp1 = icmp slt i32 %i, 255
   %sel1 = select i1 %cmp1, i32 %i, i32 255
   %cmp2 = icmp slt i32 %i, 0
   %res = select i1 %cmp2, i32 0, i32 %sel1
   ret i32 %res