[TTI] Refine the cost of EXT in getUserCost()

[haicheng]

Summary

Now, getUserCost() only checks the src and dst types of EXT to decide it is free or not. This change first checks the types, then calls isExtFreeImpl(), and check if EXT can form ExtLoad at last. Currently, only AArch64 has customized implementation of isExtFreeImpl() to check if EXT can be folded into its use. This patch just moves code around and calls existing APIs.

[junbuml]

This looks good to me. It will be good if you can share any performance or code size impact with this change?

test/Transforms/Inline/AArch64/ext.ll

3

You should remove -mcpu=kryo.

[haicheng]

In D34458#792223, @junbuml wrote:

This looks good to me. It will be good if you can share any performance or code size impact with this change?

Here is the performance and code size change of SPEC20xx on AArch64.

Benchmark	Performance (+ is better)	Code Size (+ is larger)
spec2000/gap	+1.15%	0%
spec2006/h264ref	-2.55%	0%
spec2017/leela	+5.22%	0%
spec2017/perlbench	-1.66%	0%
spec2017/x264	+1.70%	0%
spec2017/xz	-1.29%	0%

The regression in spec2006/h264ref is caused by a fully unrolled loop. The loop has an indrect call inside it. If transforming the indirect call to direct call, the performance can be improved by +10%, and then unrolling this loop can get additional +3% improvement. We are working on it.

[haicheng]

Update the test case.

[qcolombet]

include/llvm/Target/TargetLowering.h

2015

I found the name not descriptive of what it is checking.
At the very least, add a comment on what this method is doing.

[haicheng]

Add a comment and rename some argument names. Thank you, Quentin.

[haicheng]

Edit the comment.

[haicheng]

Kindly ping.

[efriedma]

The regression in spec2006/h264ref is caused by a fully unrolled loop. The loop has an indrect call inside it. If transforming the indirect call to direct call, the performance can be improved by +10%, and then unrolling this loop can get additional +3% improvement. We are working on it.

You're planning a change to the unroller? Or something else?

---

It looks like this has some impact on other targets; do we have any test coverage?

include/llvm/Analysis/TargetTransformInfoImpl.h

736

Do you need to pass through Arguments here? It looks like getExtCost depends on the operands.

[haicheng]

Add tests for other targets. Pass through operands argument.

[haicheng]

In D34458#806904, @efriedma wrote:

The regression in spec2006/h264ref is caused by a fully unrolled loop. The loop has an indrect call inside it. If transforming the indirect call to direct call, the performance can be improved by +10%, and then unrolling this loop can get additional +3% improvement. We are working on it.

You're planning a change to the unroller? Or something else?

We plan to work on transforming the indirect call to direct calls, not on the unroller.

---

It looks like this has some impact on other targets; do we have any test coverage?

I added two more tests for other targets.

Thank you,

Haicheng

[efriedma]

It looks like AArch64TargetLowering::isExtFreeImpl has some other special-cases which might be relevant here; could you add a test for shl(sext(x))?

[haicheng]

Add two test cases for shl(ext(x)).

[efriedma]

LGTM with one minor comment.

test/Transforms/Inline/AArch64/ext.ll

257

"zext"? Also, this isn't really testing what you want to test because isZExtFree is true for 32->64.

[haicheng]

Remove an inappropriate test case.

[delena]

llvm/trunk/include/llvm/Analysis/TargetTransformInfoImpl.h

735 ↗

(On Diff #106747)

This change affects not only on ARM targets. You changed cost calculation on X86 and we measure significant performance degradation after this patch. I assume that the cost of Ext instruction should be provided by target.

[delena]

I assume this patch should be reverted and reviewed again.

[eastig]

In D34458#815909, @delena wrote:

I assume this patch should be reverted and reviewed again.

Hi Elena,

Could you please share how significant the regressions are? What benchmarks? Any improvements? Any analysis of the regressions? What are differences in IR?

I have examples when regressions happened not because of the original changes but because other passes had bugs or did not support new cases introduced by the changes.

Thanks,
Evgeny

[delena]

llvm/trunk/include/llvm/CodeGen/BasicTTIImpl.h

167 ↗

(On Diff #106747)

At this point we should call to getOperationCost() and ask the target about the cost of Ext for the given type.

[haicheng]

llvm/trunk/include/llvm/CodeGen/BasicTTIImpl.h

167 ↗

(On Diff #106747)

Sorry about the performance regression.

I think isExtFree(I) above checks isZExtFree() with its types. The part that can affect x86 is isExtLoad(). Do you want me to add a target hook to choose use refined ext cost?

[delena]

llvm/trunk/include/llvm/CodeGen/BasicTTIImpl.h

167 ↗

(On Diff #106747)

First, I should understand why if the Src is not ExtLoad you return TargetTransformInfo::TCC_Basic.
Why you are not asking the target about the actual cost of the Ext?

[haicheng]

llvm/trunk/include/llvm/CodeGen/BasicTTIImpl.h

167 ↗

(On Diff #106747)

I think that getTLI()->isExtFree(I) above asks the actual cost of the Ext which is equivalent to calling getOperationCost() since they both check target specific isZExtFree() with type.

[delena]

llvm/trunk/include/llvm/CodeGen/BasicTTIImpl.h

167 ↗

(On Diff #106747)

getTLI()->isExtFree(I) is a boolean function that returns true or false. But if the Ext is not free you should provide an actual cost and not TCC_Basic.

[haicheng]

llvm/trunk/include/llvm/CodeGen/BasicTTIImpl.h

167 ↗

(On Diff #106747)

Do you want me to change this line to

return getOperationCost(Operator::getOpcode(I), I->getType(), I->getOperand(0)->getType());

which returns TCC_Basic, TCC_Free, or TCC_Expensive?

[delena]

llvm/trunk/include/llvm/CodeGen/BasicTTIImpl.h

167 ↗

(On Diff #106747)

I'll try to make this change and measure performance..

[haicheng]

llvm/trunk/include/llvm/Analysis/TargetTransformInfoImpl.h

124 ↗

(On Diff #106747)

Do you want to make the same change here?

[delena]

I dug into the performance degradation and found that isExtLoad() reduces cost of a loop, as a result the loop is being unrolled. Unrolling this loop causes performance degradation.
But the load + zext cost is correct and it is 1 Op. I'll work more with our architects in order to understand what parameter is incorrect on X86.

[haicheng]

In D34458#818835, @delena wrote:

I dug into the performance degradation and found that isExtLoad() reduces cost of a loop, as a result the loop is being unrolled. Unrolling this loop causes performance degradation.
But the load + zext cost is correct and it is 1 Op. I'll work more with our architects in order to understand what parameter is incorrect on X86.

Thank you very much for clearing things up. Please let me know if you want anything from me.

Haicheng

include/llvm/Analysis/TargetTransformInfo.h

public:
   /// \brief Estimate the cost of a GEP operation when lowered.
   ///
   /// The contract for this function is the same as \c getOperationCost except
   /// that it supports an interface that provides extra information specific to
   /// the GEP operation.
   int getGEPCost(Type *PointeeType, const Value *Ptr,
                  ArrayRef<const Value *> Operands) const;
 
+  /// \brief Estimate the cost of a EXT operation when lowered.
+  ///
+  /// The contract for this function is the same as \c getOperationCost except
+  /// that it supports an interface that provides extra information specific to
+  /// the EXT operation.
+  int getExtCost(const Instruction *I, const Value *Src) const;
+
   /// \brief Estimate the cost of a function call when lowered.
   ///
   /// The contract for this is the same as \c getOperationCost except that it
   /// supports an interface that provides extra information specific to call
   /// instructions.
   ///
   /// This is the most basic query for estimating call cost: it only knows the
   /// function type and (potentially) the number of arguments at the call site.

 
 class TargetTransformInfo::Concept {
 public:
   virtual ~Concept() = 0;
   virtual const DataLayout &getDataLayout() const = 0;
   virtual int getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) = 0;
   virtual int getGEPCost(Type *PointeeType, const Value *Ptr,
                          ArrayRef<const Value *> Operands) = 0;
+  virtual int getExtCost(const Instruction *I, const Value *Src) = 0;
   virtual int getCallCost(FunctionType *FTy, int NumArgs) = 0;
   virtual int getCallCost(const Function *F, int NumArgs) = 0;
   virtual int getCallCost(const Function *F,
                           ArrayRef<const Value *> Arguments) = 0;
   virtual unsigned getInliningThresholdMultiplier() = 0;
   virtual int getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
                                ArrayRef<Type *> ParamTys) = 0;
   virtual int getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
public:
 
   int getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) override {
     return Impl.getOperationCost(Opcode, Ty, OpTy);
   }
   int getGEPCost(Type *PointeeType, const Value *Ptr,
                  ArrayRef<const Value *> Operands) override {
     return Impl.getGEPCost(PointeeType, Ptr, Operands);
   }
+  int getExtCost(const Instruction *I, const Value *Src) override {
+    return Impl.getExtCost(I, Src);
+  }
   int getCallCost(FunctionType *FTy, int NumArgs) override {
     return Impl.getCallCost(FTy, NumArgs);
   }
   int getCallCost(const Function *F, int NumArgs) override {
     return Impl.getCallCost(F, NumArgs);
   }
   int getCallCost(const Function *F,
                   ArrayRef<const Value *> Arguments) override {

include/llvm/Analysis/TargetTransformInfoImpl.h

public:
   }
 
   unsigned getEstimatedNumberOfCaseClusters(const SwitchInst &SI,
                                             unsigned &JTSize) {
     JTSize = 0;
     return SI.getNumCases();
   }
 
+  int getExtCost(const Instruction *I, const Value *Src) {
+    return TTI::TCC_Basic;
+  }
+
   unsigned getCallCost(FunctionType *FTy, int NumArgs) {
     assert(FTy && "FunctionType must be provided to this routine.");
 
     // The target-independent implementation just measures the size of the
     // function by approximating that each argument will take on average one
     // instruction to prepare.
 
     if (NumArgs < 0)
unsigned getUserCost(const User *U, ArrayRef<const Value *> Operands) {
     }
 
     if (const CastInst *CI = dyn_cast<CastInst>(U)) {
       // Result of a cmp instruction is often extended (to be used by other
       // cmp instructions, logical or return instructions). These are usually
       // nop on most sane targets.
       if (isa<CmpInst>(CI->getOperand(0)))
         return TTI::TCC_Free;
+      if (isa<SExtInst>(CI) || isa<ZExtInst>(CI) || isa<FPExtInst>(CI))
+        return static_cast<T *>(this)->getExtCost(CI, Operands.back());
     }
 
     return static_cast<T *>(this)->getOperationCost(
         Operator::getOpcode(U), U->getType(),
         U->getNumOperands() == 1 ? U->getOperand(0)->getType() : nullptr);
   }
 };
 }
 
 #endif

include/llvm/CodeGen/BasicTTIImpl.h

bool isTypeLegal(Type *Ty) {
     return getTLI()->isTypeLegal(VT);
   }
 
   int getGEPCost(Type *PointeeType, const Value *Ptr,
                  ArrayRef<const Value *> Operands) {
     return BaseT::getGEPCost(PointeeType, Ptr, Operands);
   }
 
+  int getExtCost(const Instruction *I, const Value *Src) {
+    if (getTLI()->isExtFree(I))
+      return TargetTransformInfo::TCC_Free;
+
+    if (isa<ZExtInst>(I) || isa<SExtInst>(I))
+      if (const LoadInst *LI = dyn_cast<LoadInst>(Src))
+        if (getTLI()->isExtLoad(LI, I, DL))
+          return TargetTransformInfo::TCC_Free;
+
+    return TargetTransformInfo::TCC_Basic;
+  }
+
   unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
                             ArrayRef<const Value *> Arguments) {
     return BaseT::getIntrinsicCost(IID, RetTy, Arguments);
   }
 
   unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
                             ArrayRef<Type *> ParamTys) {
     if (IID == Intrinsic::cttz) {

include/llvm/Target/TargetLowering.h

bool isExtFree(const Instruction *I) const {
     case Instruction::SExt:
       break;
     default:
       llvm_unreachable("Instruction is not an extension");
     }
     return isExtFreeImpl(I);
   }
 
+  /// Return true if \p Load and \p Ext can form an ExtLoad.
+  /// For example, in AArch64
+  ///   %L = load i8, i8* %ptr
+  ///   %E = zext i8 %L to i32
+  /// can be lowered into one load instruction
+  ///   ldrb w0, [x0]
+  bool isExtLoad(const LoadInst *Load, const Instruction *Ext,
+                 const DataLayout &DL) const {
+    EVT VT = getValueType(DL, Ext->getType());
+    EVT LoadVT = getValueType(DL, Load->getType());
+
+    // If the load has other users and the truncate is not free, the ext
+    // probably isn't free.
+    if (!Load->hasOneUse() && (isTypeLegal(LoadVT) || !isTypeLegal(VT)) &&
+        !isTruncateFree(Ext->getType(), Load->getType()))
+      return false;
+
+    // Check whether the target supports casts folded into loads.
+    unsigned LType;
+    if (isa<ZExtInst>(Ext))
+      LType = ISD::ZEXTLOAD;
+    else {
+      assert(isa<SExtInst>(Ext) && "Unexpected ext type!");
+      LType = ISD::SEXTLOAD;
+    }
+
+    return isLoadExtLegal(LType, VT, LoadVT);
+  }
+
   /// Return true if any actual instruction that defines a value of type FromTy
   /// implicitly zero-extends the value to ToTy in the result register.
   ///
   /// The function should return true when it is likely that the truncate can
   /// be freely folded with an instruction defining a value of FromTy. If
   /// the defining instruction is unknown (because you're looking at a
   /// function argument, PHI, etc.) then the target may require an
   /// explicit truncate, which is not necessarily free, but this function

lib/Analysis/TargetTransformInfo.cpp

unsigned TargetTransformInfo::getInliningThresholdMultiplier() const {
   return TTIImpl->getInliningThresholdMultiplier();
 }
 
 int TargetTransformInfo::getGEPCost(Type *PointeeType, const Value *Ptr,
                                     ArrayRef<const Value *> Operands) const {
   return TTIImpl->getGEPCost(PointeeType, Ptr, Operands);
 }
 
+int TargetTransformInfo::getExtCost(const Instruction *I,
+                                    const Value *Src) const {
+  return TTIImpl->getExtCost(I, Src);
+}
+
 int TargetTransformInfo::getIntrinsicCost(
     Intrinsic::ID IID, Type *RetTy, ArrayRef<const Value *> Arguments) const {
   int Cost = TTIImpl->getIntrinsicCost(IID, RetTy, Arguments);
   assert(Cost >= 0 && "TTI should not produce negative costs!");
   return Cost;
 }
 
 unsigned

lib/CodeGen/CodeGenPrepare.cpp

if (!LI)
     return false;
 
   // If they're already in the same block, there's nothing to do.
   // Make the cheap checks first if we did not promote.
   // If we promoted, we need to check if it is indeed profitable.
   if (!HasPromoted && LI->getParent() == Inst->getParent())
     return false;
 
-  EVT VT = TLI->getValueType(*DL, Inst->getType());
-  EVT LoadVT = TLI->getValueType(*DL, LI->getType());
-
-  // If the load has other users and the truncate is not free, this probably
-  // isn't worthwhile.
-  if (!LI->hasOneUse() && (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) &&
-      !TLI->isTruncateFree(Inst->getType(), LI->getType()))
-    return false;
-
-  // Check whether the target supports casts folded into loads.
-  unsigned LType;
-  if (isa<ZExtInst>(Inst))
-    LType = ISD::ZEXTLOAD;
-  else {
-    assert(isa<SExtInst>(Inst) && "Unexpected ext type!");
-    LType = ISD::SEXTLOAD;
-  }
-
-  return TLI->isLoadExtLegal(LType, VT, LoadVT);
+  return TLI->isExtLoad(LI, Inst, *DL);
 }
 
 /// Move a zext or sext fed by a load into the same basic block as the load,
 /// unless conditions are unfavorable. This allows SelectionDAG to fold the
 /// extend into the load.
 ///
 /// E.g.,
 /// \code

test/Transforms/Inline/AArch64/ext.ll

+; REQUIRES: asserts
+; RUN: opt -inline -mtriple=aarch64--linux-gnu -S -debug-only=inline-cost < %s 2>&1 | FileCheck %s
+
+target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
+target triple = "aarch64--linux-gnu"
+
+define i32 @outer1(i32* %ptr, i32 %i) {
+  %C = call i32 @inner1(i32* %ptr, i32 %i)
+  ret i32 %C
+}
+
+; sext can be folded into gep.
+; CHECK: Analyzing call of inner1
+; CHECK: NumInstructionsSimplified: 3
+; CHECK: NumInstructions: 4
+define i32 @inner1(i32* %ptr, i32 %i) {
+  %E = sext i32 %i to i64
+  %G = getelementptr inbounds i32, i32* %ptr, i64 %E
+  %L = load i32, i32* %G
+  ret i32 %L
+}
+
+define i32 @outer2(i32* %ptr, i32 %i) {
+  %C = call i32 @inner2(i32* %ptr, i32 %i)
+  ret i32 %C
+}
+
+; zext from i32 to i64 is free.
+; CHECK: Analyzing call of inner2
+; CHECK: NumInstructionsSimplified: 3
+; CHECK: NumInstructions: 4
+define i32 @inner2(i32* %ptr, i32 %i) {
+  %E = zext i32 %i to i64
+  %G = getelementptr inbounds i32, i32* %ptr, i64 %E
+  %L = load i32, i32* %G
+  ret i32 %L
+}
+
+define i32 @outer3(i32* %ptr, i16 %i) {
+  %C = call i32 @inner3(i32* %ptr, i16 %i)
+  ret i32 %C
+}
+
+; zext can be folded into gep.
+; CHECK: Analyzing call of inner3
+; CHECK: NumInstructionsSimplified: 3
+; CHECK: NumInstructions: 4
+define i32 @inner3(i32* %ptr, i16 %i) {
+  %E = zext i16 %i to i64
+  %G = getelementptr inbounds i32, i32* %ptr, i64 %E
+  %L = load i32, i32* %G
+  ret i32 %L
+}
+
+define i16 @outer4(i8* %ptr) {
+  %C = call i16 @inner4(i8* %ptr)
+  ret i16 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner4
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i16 @inner4(i8* %ptr) {
+  %L = load i8, i8* %ptr
+  %E = zext i8 %L to i16
+  ret i16 %E
+}
+
+define i16 @outer5(i8* %ptr) {
+  %C = call i16 @inner5(i8* %ptr)
+  ret i16 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner5
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i16 @inner5(i8* %ptr) {
+  %L = load i8, i8* %ptr
+  %E = sext i8 %L to i16
+  ret i16 %E
+}
+
+define i32 @outer6(i8* %ptr) {
+  %C = call i32 @inner6(i8* %ptr)
+  ret i32 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner6
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i32 @inner6(i8* %ptr) {
+  %L = load i8, i8* %ptr
+  %E = zext i8 %L to i32
+  ret i32 %E
+}
+
+define i32 @outer7(i8* %ptr) {
+  %C = call i32 @inner7(i8* %ptr)
+  ret i32 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner7
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i32 @inner7(i8* %ptr) {
+  %L = load i8, i8* %ptr
+  %E = sext i8 %L to i32
+  ret i32 %E
+}
+
+define i32 @outer8(i16* %ptr) {
+  %C = call i32 @inner8(i16* %ptr)
+  ret i32 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner8
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i32 @inner8(i16* %ptr) {
+  %L = load i16, i16* %ptr
+  %E = zext i16 %L to i32
+  ret i32 %E
+}
+
+define i32 @outer9(i16* %ptr) {
+  %C = call i32 @inner9(i16* %ptr)
+  ret i32 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner9
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i32 @inner9(i16* %ptr) {
+  %L = load i16, i16* %ptr
+  %E = sext i16 %L to i32
+  ret i32 %E
+}
+
+define i64 @outer10(i8* %ptr) {
+  %C = call i64 @inner10(i8* %ptr)
+  ret i64 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner10
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i64 @inner10(i8* %ptr) {
+  %L = load i8, i8* %ptr
+  %E = zext i8 %L to i64
+  ret i64 %E
+}
+
+define i64 @outer11(i8* %ptr) {
+  %C = call i64 @inner11(i8* %ptr)
+  ret i64 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner11
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i64 @inner11(i8* %ptr) {
+  %L = load i8, i8* %ptr
+  %E = sext i8 %L to i64
+  ret i64 %E
+}
+
+define i64 @outer12(i16* %ptr) {
+  %C = call i64 @inner12(i16* %ptr)
+  ret i64 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner12
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i64 @inner12(i16* %ptr) {
+  %L = load i16, i16* %ptr
+  %E = zext i16 %L to i64
+  ret i64 %E
+}
+
+define i64 @outer13(i16* %ptr) {
+  %C = call i64 @inner13(i16* %ptr)
+  ret i64 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner13
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i64 @inner13(i16* %ptr) {
+  %L = load i16, i16* %ptr
+  %E = sext i16 %L to i64
+  ret i64 %E
+}
+
+define i64 @outer14(i32* %ptr) {
+  %C = call i64 @inner14(i32* %ptr)
+  ret i64 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner14
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i64 @inner14(i32* %ptr) {
+  %L = load i32, i32* %ptr
+  %E = zext i32 %L to i64
+  ret i64 %E
+}
+
+define i64 @outer15(i32* %ptr) {
+  %C = call i64 @inner15(i32* %ptr)
+  ret i64 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner15
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i64 @inner15(i32* %ptr) {
+  %L = load i32, i32* %ptr
+  %E = sext i32 %L to i64
+  ret i64 %E
+}

test/Transforms/Inline/PowerPC/ext.ll

+; REQUIRES: asserts
+; RUN: opt -inline -S -debug-only=inline-cost < %s 2>&1 | FileCheck %s
+
+target datalayout = "E-m:e-i64:64-n32:64"
+target triple = "powerpc64le-ibm-linux-gnu"
+
+define i16 @outer1(i8* %ptr) {
+  %C = call i16 @inner1(i8* %ptr)
+  ret i16 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner1
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i16 @inner1(i8* %ptr) {
+  %L = load i8, i8* %ptr
+  %E = zext i8 %L to i16
+  ret i16 %E
+}
+
+define i32 @outer2(i8* %ptr) {
+  %C = call i32 @inner2(i8* %ptr)
+  ret i32 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner2
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i32 @inner2(i8* %ptr) {
+  %L = load i8, i8* %ptr
+  %E = zext i8 %L to i32
+  ret i32 %E
+}
+
+define i32 @outer3(i16* %ptr) {
+  %C = call i32 @inner3(i16* %ptr)
+  ret i32 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner3
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i32 @inner3(i16* %ptr) {
+  %L = load i16, i16* %ptr
+  %E = zext i16 %L to i32
+  ret i32 %E
+}
+
+define i32 @outer4(i16* %ptr) {
+  %C = call i32 @inner4(i16* %ptr)
+  ret i32 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner4
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i32 @inner4(i16* %ptr) {
+  %L = load i16, i16* %ptr
+  %E = sext i16 %L to i32
+  ret i32 %E
+}
+
+define i64 @outer5(i8* %ptr) {
+  %C = call i64 @inner5(i8* %ptr)
+  ret i64 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner5
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i64 @inner5(i8* %ptr) {
+  %L = load i8, i8* %ptr
+  %E = zext i8 %L to i64
+  ret i64 %E
+}
+
+define i64 @outer6(i16* %ptr) {
+  %C = call i64 @inner6(i16* %ptr)
+  ret i64 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner6
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i64 @inner6(i16* %ptr) {
+  %L = load i16, i16* %ptr
+  %E = zext i16 %L to i64
+  ret i64 %E
+}
+
+define i64 @outer7(i16* %ptr) {
+  %C = call i64 @inner7(i16* %ptr)
+  ret i64 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner7
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i64 @inner7(i16* %ptr) {
+  %L = load i16, i16* %ptr
+  %E = sext i16 %L to i64
+  ret i64 %E
+}
+
+define i64 @outer8(i32* %ptr) {
+  %C = call i64 @inner8(i32* %ptr)
+  ret i64 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner8
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i64 @inner8(i32* %ptr) {
+  %L = load i32, i32* %ptr
+  %E = zext i32 %L to i64
+  ret i64 %E
+}
+
+define i64 @outer9(i32* %ptr) {
+  %C = call i64 @inner9(i32* %ptr)
+  ret i64 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner9
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i64 @inner9(i32* %ptr) {
+  %L = load i32, i32* %ptr
+  %E = sext i32 %L to i64
+  ret i64 %E
+}

test/Transforms/Inline/PowerPC/lit.local.cfg

+if not 'PowerPC' in config.root.targets:
+    config.unsupported = True
+

test/Transforms/Inline/X86/ext.ll

+; REQUIRES: asserts
+; RUN: opt -inline -mtriple=x86_64-unknown-unknown -S -debug-only=inline-cost < %s 2>&1 | FileCheck %s
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-unknown"
+
+define i32 @outer1(i32* %ptr, i32 %i) {
+  %C = call i32 @inner1(i32* %ptr, i32 %i)
+  ret i32 %C
+}
+
+; zext from i32 to i64 is free.
+; CHECK: Analyzing call of inner1
+; CHECK: NumInstructionsSimplified: 3
+; CHECK: NumInstructions: 4
+define i32 @inner1(i32* %ptr, i32 %i) {
+  %E = zext i32 %i to i64
+  %G = getelementptr inbounds i32, i32* %ptr, i64 %E
+  %L = load i32, i32* %G
+  ret i32 %L
+}
+
+define i16 @outer2(i8* %ptr) {
+  %C = call i16 @inner2(i8* %ptr)
+  ret i16 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner2
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i16 @inner2(i8* %ptr) {
+  %L = load i8, i8* %ptr
+  %E = zext i8 %L to i16
+  ret i16 %E
+}
+
+define i16 @outer3(i8* %ptr) {
+  %C = call i16 @inner3(i8* %ptr)
+  ret i16 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner3
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i16 @inner3(i8* %ptr) {
+  %L = load i8, i8* %ptr
+  %E = sext i8 %L to i16
+  ret i16 %E
+}
+
+define i32 @outer4(i8* %ptr) {
+  %C = call i32 @inner4(i8* %ptr)
+  ret i32 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner4
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i32 @inner4(i8* %ptr) {
+  %L = load i8, i8* %ptr
+  %E = zext i8 %L to i32
+  ret i32 %E
+}
+
+define i32 @outer5(i8* %ptr) {
+  %C = call i32 @inner5(i8* %ptr)
+  ret i32 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner5
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i32 @inner5(i8* %ptr) {
+  %L = load i8, i8* %ptr
+  %E = sext i8 %L to i32
+  ret i32 %E
+}
+
+define i32 @outer6(i16* %ptr) {
+  %C = call i32 @inner6(i16* %ptr)
+  ret i32 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner6
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i32 @inner6(i16* %ptr) {
+  %L = load i16, i16* %ptr
+  %E = zext i16 %L to i32
+  ret i32 %E
+}
+
+define i32 @outer7(i16* %ptr) {
+  %C = call i32 @inner7(i16* %ptr)
+  ret i32 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner7
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i32 @inner7(i16* %ptr) {
+  %L = load i16, i16* %ptr
+  %E = sext i16 %L to i32
+  ret i32 %E
+}
+
+define i64 @outer8(i8* %ptr) {
+  %C = call i64 @inner8(i8* %ptr)
+  ret i64 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner8
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i64 @inner8(i8* %ptr) {
+  %L = load i8, i8* %ptr
+  %E = zext i8 %L to i64
+  ret i64 %E
+}
+
+define i64 @outer9(i8* %ptr) {
+  %C = call i64 @inner9(i8* %ptr)
+  ret i64 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner9
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i64 @inner9(i8* %ptr) {
+  %L = load i8, i8* %ptr
+  %E = sext i8 %L to i64
+  ret i64 %E
+}
+
+define i64 @outer10(i16* %ptr) {
+  %C = call i64 @inner10(i16* %ptr)
+  ret i64 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner10
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i64 @inner10(i16* %ptr) {
+  %L = load i16, i16* %ptr
+  %E = zext i16 %L to i64
+  ret i64 %E
+}
+
+define i64 @outer11(i16* %ptr) {
+  %C = call i64 @inner11(i16* %ptr)
+  ret i64 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner11
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i64 @inner11(i16* %ptr) {
+  %L = load i16, i16* %ptr
+  %E = sext i16 %L to i64
+  ret i64 %E
+}
+
+define i64 @outer12(i32* %ptr) {
+  %C = call i64 @inner12(i32* %ptr)
+  ret i64 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner12
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i64 @inner12(i32* %ptr) {
+  %L = load i32, i32* %ptr
+  %E = zext i32 %L to i64
+  ret i64 %E
+}
+
+define i64 @outer13(i32* %ptr) {
+  %C = call i64 @inner13(i32* %ptr)
+  ret i64 %C
+}
+
+; It is an ExtLoad.
+; CHECK: Analyzing call of inner13
+; CHECK: NumInstructionsSimplified: 2
+; CHECK: NumInstructions: 3
+define i64 @inner13(i32* %ptr) {
+  %L = load i32, i32* %ptr
+  %E = sext i32 %L to i64
+  ret i64 %E
+}