[InstCombine] Fix PR37526: MinMax patterns produce an infinite loop.

[ABataev]

Summary

This patch fixes PR37526 by simplifying the newly generated LoadInst
instructions. If the pointer address is a bitcast from the pointer to
the NewType, we can just remove this extra bitcast instead of creating
the new one. This fixes the PR37526 + may speed up the whole compilation
process.

[xbolva00]

lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp

467

NewPtr = Op0 ?

[ABataev]

Improved code.

[spatel]

lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp

462–467

Use a matcher to simplify this?

Value *NewPtr;
if (!(match(Ptr, m_BitCast(m_Value(NewPtr))) &&
      NewPtr->getType()->getPointerElementType() == NewTy &&
      NewPtr->getType()->getPointerAddressSpace() == AS))
  NewPtr = IC.Builder.CreateBitCast(Ptr, NewTy->getPointerTo(AS));

test/Transforms/InstCombine/PR37526.ll

4–6

We shouldn't need a struct or loop to repro - something like this should be ok?

define void @PR37526(i32* %pz, i32* %px, i32* %py) {
  %t1 = bitcast i32* %pz to i64*
  %t2 = load i32, i32* %py
  %t3 = load i32, i32* %px
  %cmp = icmp slt i32 %t2, %t3
  %select = select i1 %cmp, i32* %px, i32* %py
  %bc = bitcast i32* %select to i64*
  %r = load i64, i64* %bc
  store i64 %r, i64* %t1
  ret void
}

[ABataev]

Address comments from Sanjay Patel

[xbolva00]

lg

lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp

 /// that pointer type, load it, etc.
 ///
 /// Note that this will create all of the instructions with whatever insert
 /// point the \c InstCombiner currently is using.
 static LoadInst *combineLoadToNewType(InstCombiner &IC, LoadInst &LI, Type *NewTy,
                                       const Twine &Suffix = "") {
   assert((!LI.isAtomic() || isSupportedAtomicType(NewTy)) &&
          "can't fold an atomic load to requested type");
-  
   Value *Ptr = LI.getPointerOperand();
   unsigned AS = LI.getPointerAddressSpace();
   SmallVector<std::pair<unsigned, MDNode *>, 8> MD;
   LI.getAllMetadata(MD);
 
+  Value *NewPtr = nullptr;
+  if (auto *BC = dyn_cast<BitCastInst>(Ptr)) {
+    Value *Op0 = BC->getOperand(0);
+    if (Op0->getType()->getPointerElementType() == NewTy &&
+        Op0->getType()->getPointerAddressSpace() == AS)
+      NewPtr = Op0;
+  }
+  if (!NewPtr)
+    NewPtr = IC.Builder.CreateBitCast(Ptr, NewTy->getPointerTo(AS));
+
   LoadInst *NewLoad = IC.Builder.CreateAlignedLoad(
-      IC.Builder.CreateBitCast(Ptr, NewTy->getPointerTo(AS)),
-      LI.getAlignment(), LI.isVolatile(), LI.getName() + Suffix);
+      NewPtr, LI.getAlignment(), LI.isVolatile(), LI.getName() + Suffix);
   NewLoad->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
   MDBuilder MDB(NewLoad->getContext());
   for (const auto &MDPair : MD) {
     unsigned ID = MDPair.first;
     MDNode *N = MDPair.second;
     // Note, essentially every kind of metadata should be preserved here! This
     // routine is supposed to clone a load instruction changing *only its type*.
     // The only metadata it makes sense to drop is metadata which is invalidated

test/Transforms/InstCombine/PR37526.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -instcombine -S < %s | FileCheck %s
+
+%struct.x5 = type { i32 (...)**, i32, i32 }
+
+define void @repro(%struct.x5*, %struct.x5* %x12, %struct.x5* %x13) {
+; CHECK-LABEL: @repro(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[X37_I_I:%.*]] = getelementptr inbounds [[STRUCT_X5:%.*]], %struct.x5* [[X12:%.*]], i64 0, i32 1
+; CHECK-NEXT:    [[X37_I:%.*]] = getelementptr inbounds [[STRUCT_X5]], %struct.x5* [[X13:%.*]], i64 0, i32 1
+; CHECK-NEXT:    br label [[FOR_COND:%.*]]
+; CHECK:       for.cond:
+; CHECK-NEXT:    [[X14_0:%.*]] = phi %struct.x5* [ undef, [[ENTRY:%.*]] ], [ [[INCDEC_PTR:%.*]], [[FOR_COND]] ]
+; CHECK-NEXT:    [[INCDEC_PTR]] = getelementptr inbounds [[STRUCT_X5]], %struct.x5* [[X14_0]], i64 -1
+; CHECK-NEXT:    [[TMP1:%.*]] = load i32, i32* [[X37_I_I]], align 4
+; CHECK-NEXT:    [[X37_I2_I:%.*]] = getelementptr inbounds [[STRUCT_X5]], %struct.x5* [[X14_0]], i64 -1, i32 1
+; CHECK-NEXT:    [[TMP2:%.*]] = load i32, i32* [[X37_I2_I]], align 4
+; CHECK-NEXT:    [[CMP_I:%.*]] = icmp slt i32 [[TMP1]], [[TMP2]]
+; CHECK-NEXT:    [[SPEC_SELECT:%.*]] = select i1 [[CMP_I]], i32* [[X37_I2_I]], i32* [[X37_I_I]]
+; CHECK-NEXT:    [[STOREMERGE1:%.*]] = load i32, i32* [[SPEC_SELECT]], align 8
+; CHECK-NEXT:    store i32 [[STOREMERGE1]], i32* [[X37_I]], align 8
+; CHECK-NEXT:    br label [[FOR_COND]]
+;
+entry:
+  %x37.i.i = getelementptr inbounds %struct.x5, %struct.x5* %x12, i64 0, i32 1
+  %x37.i = getelementptr inbounds %struct.x5, %struct.x5* %x13, i64 0, i32 1
+  %1 = bitcast i32* %x37.i to i64*
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.cond, %entry
+  %x14.0 = phi %struct.x5* [ undef, %entry ], [ %incdec.ptr, %for.cond ]
+  %incdec.ptr = getelementptr inbounds %struct.x5, %struct.x5* %x14.0, i64 -1
+  %2 = load i32, i32* %x37.i.i, align 4
+  %x37.i2.i = getelementptr inbounds %struct.x5, %struct.x5* %x14.0, i64 -1, i32 1
+  %3 = load i32, i32* %x37.i2.i, align 4
+  %cmp.i = icmp slt i32 %2, %3
+  %spec.select = select i1 %cmp.i, i32* %x37.i2.i, i32* %x37.i.i
+  %storemerge.in = bitcast i32* %spec.select to i64*
+  %storemerge = load i64, i64* %storemerge.in, align 8
+  store i64 %storemerge, i64* %1, align 8
+  br label %for.cond
+}
+