[InlineCost] Find more free binary operations

[haicheng]

Summary

Currently, inline cost model considers a binary operator as free only if both its operands are constants. Some simple cases are missing such as a + 0, a - a, etc. This patch modifies visitBinaryOperator() to call SimplifyBinOp() without going through simplifyInstruction() to get rid of the constant restriction. Thus, visitAnd() and visitOr() are not needed.

Here is the impact on code size and performance (spec20xx)

Benchmark	Code Size (%)	Perf (%)
	(+ is bigger)	(+ is faster)
spec2000/gcc	0.02	2.26
spec2006/gcc	0.04	2.18
spec2017/imagick	0.03	1.02
spec2006/perlbench	0.00	0.60
spec2000/crafty	0.43	0.58
spec2017/parest	0.01	0.16
spec2017/gcc	-0.01	0.14
spec2006/xalancbmk	0.00	0.12
spec2006/dealII	0.04	0.10
spec2017/blender	0.00	-0.04
spec2017/povray	-1.46	-0.12
spec2017/xalancbmk	0.61	-0.14
spec2006/povray	-1.35	-0.15

[hfinkel]

LGTM

It looks like we could do a similar thing in visitCmpInst too (and visitCallSite, however, we'd also need to fix the FIXME about interface inefficiency in CallAnalyzer::simplifyCallSite first?).

lib/Analysis/InlineCost.cpp

class CallAnalyzer : public InstVisitor<CallAnalyzer, bool> {
   bool visitPHI(PHINode &I);
   bool visitGetElementPtr(GetElementPtrInst &I);
   bool visitBitCast(BitCastInst &I);
   bool visitPtrToInt(PtrToIntInst &I);
   bool visitIntToPtr(IntToPtrInst &I);
   bool visitCastInst(CastInst &I);
   bool visitUnaryInstruction(UnaryInstruction &I);
   bool visitCmpInst(CmpInst &I);
-  bool visitAnd(BinaryOperator &I);
-  bool visitOr(BinaryOperator &I);
   bool visitSub(BinaryOperator &I);
   bool visitBinaryOperator(BinaryOperator &I);
   bool visitLoad(LoadInst &I);
   bool visitStore(StoreInst &I);
   bool visitExtractValue(ExtractValueInst &I);
   bool visitInsertValue(InsertValueInst &I);
   bool visitCallSite(CallSite CS);
   bool visitReturnInst(ReturnInst &RI);
if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) {
     }
 
     disableSROA(CostIt);
   }
 
   return false;
 }
 
-bool CallAnalyzer::visitOr(BinaryOperator &I) {
-  // This is necessary because the generic simplify instruction only works if
-  // both operands are constants.
-  for (unsigned i = 0; i < 2; ++i) {
-    if (ConstantInt *C = dyn_cast_or_null<ConstantInt>(
-            SimplifiedValues.lookup(I.getOperand(i))))
-      if (C->isAllOnesValue()) {
-        SimplifiedValues[&I] = C;
-        return true;
-      }
-  }
-  return Base::visitOr(I);
-}
-
-bool CallAnalyzer::visitAnd(BinaryOperator &I) {
-  // This is necessary because the generic simplify instruction only works if
-  // both operands are constants.
-  for (unsigned i = 0; i < 2; ++i) {
-    if (ConstantInt *C = dyn_cast_or_null<ConstantInt>(
-            SimplifiedValues.lookup(I.getOperand(i))))
-      if (C->isZero()) {
-        SimplifiedValues[&I] = C;
-        return true;
-      }
-  }
-  return Base::visitAnd(I);
-}
-
 bool CallAnalyzer::visitSub(BinaryOperator &I) {
   // Try to handle a special case: we can fold computing the difference of two
   // constant-related pointers.
   Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
   Value *LHSBase, *RHSBase;
   APInt LHSOffset, RHSOffset;
   std::tie(LHSBase, LHSOffset) = ConstantOffsetPtrs.lookup(LHS);
   if (LHSBase) {
bool CallAnalyzer::visitSub(BinaryOperator &I) {
 
   // Otherwise, fall back to the generic logic for simplifying and handling
   // instructions.
   return Base::visitSub(I);
 }
 
 bool CallAnalyzer::visitBinaryOperator(BinaryOperator &I) {
   Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
-  auto Evaluate = [&](SmallVectorImpl<Constant *> &COps) {
-    Value *SimpleV = nullptr;
-    if (auto FI = dyn_cast<FPMathOperator>(&I))
-      SimpleV = SimplifyFPBinOp(I.getOpcode(), COps[0], COps[1],
-                                FI->getFastMathFlags(), DL);
-    else
-      SimpleV = SimplifyBinOp(I.getOpcode(), COps[0], COps[1], DL);
-    return dyn_cast_or_null<Constant>(SimpleV);
-  };
+  Constant *CLHS = dyn_cast<Constant>(LHS);
+  if (!CLHS)
+    CLHS = SimplifiedValues.lookup(LHS);
+  Constant *CRHS = dyn_cast<Constant>(RHS);
+  if (!CRHS)
+    CRHS = SimplifiedValues.lookup(RHS);
+
+    SimpleV = SimplifyFPBinOp(I.getOpcode(), CLHS ? CLHS : LHS,
+                              CRHS ? CRHS : RHS, FI->getFastMathFlags(), DL);
+    SimpleV =
+        SimplifyBinOp(I.getOpcode(), CLHS ? CLHS : LHS, CRHS ? CRHS : RHS, DL);
+
+  if (Constant *C = dyn_cast_or_null<Constant>(SimpleV))
+    SimplifiedValues[&I] = C;
 
-  if (simplifyInstruction(I, Evaluate))
+  if (SimpleV)
     return true;
 
   // Disable any SROA on arguments to arbitrary, unsimplified binary operators.
   disableSROA(LHS);
   disableSROA(RHS);
 
   return false;
 }

test/Transforms/Inline/AArch64/binop.ll

+; RUN: opt -inline -mtriple=aarch64--linux-gnu -S -o - < %s -inline-threshold=0 | 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"
+
+declare void @pad()
+@glbl = external global i32
+
+define i32 @outer_add1(i32 %a) {
+; CHECK-LABEL: @outer_add1(
+; CHECK-NOT: call i32 @add
+  %C = call i32 @add(i32 %a, i32 0)
+  ret i32 %C
+}
+
+define i32 @outer_add2(i32 %a) {
+; CHECK-LABEL: @outer_add2(
+; CHECK-NOT: call i32 @add
+  %C = call i32 @add(i32 0, i32 %a)
+  ret i32 %C
+}
+
+define i32 @add(i32 %a, i32 %b) {
+  %add = add i32 %a, %b
+  call void @pad()
+  store i32 0, i32* @glbl
+  ret i32 %add
+}
+
+
+
+define i32 @outer_sub1(i32 %a) {
+; CHECK-LABEL: @outer_sub1(
+; CHECK-NOT: call i32 @sub1
+  %C = call i32 @sub1(i32 %a, i32 0)
+  ret i32 %C
+}
+
+define i32 @sub1(i32 %a, i32 %b) {
+  %sub = sub i32 %a, %b
+  call void @pad()
+  store i32 0, i32* @glbl
+  ret i32 %sub
+}
+
+
+define i32 @outer_sub2(i32 %a) {
+; CHECK-LABEL: @outer_sub2(
+; CHECK-NOT: call i32 @sub2
+  %C = call i32 @sub2(i32 %a)
+  ret i32 %C
+}
+
+define i32 @sub2(i32 %a) {
+  %sub = sub i32 %a, %a
+  call void @pad()
+  ret i32 %sub
+}
+
+
+
+define i32 @outer_mul1(i32 %a) {
+; CHECK-LABEL: @outer_mul1(
+; CHECK-NOT: call i32 @mul
+  %C = call i32 @mul(i32 %a, i32 0)
+  ret i32 %C
+}
+
+define i32 @outer_mul2(i32 %a) {
+; CHECK-LABEL: @outer_mul2(
+; CHECK-NOT: call i32 @mul
+  %C = call i32 @mul(i32 %a, i32 1)
+  ret i32 %C
+}
+
+define i32 @mul(i32 %a, i32 %b) {
+  %mul = mul i32 %a, %b
+  call void @pad()
+  store i32 0, i32* @glbl
+  ret i32 %mul
+}
+
+
+
+define i32 @outer_div1(i32 %a) {
+; CHECK-LABEL: @outer_div1(
+; CHECK-NOT: call i32 @div1
+  %C = call i32 @div1(i32 0, i32 %a)
+  ret i32 %C
+}
+
+define i32 @outer_div2(i32 %a) {
+; CHECK-LABEL: @outer_div2(
+; CHECK-NOT: call i32 @div1
+  %C = call i32 @div1(i32 %a, i32 1)
+  ret i32 %C
+}
+
+define i32 @div1(i32 %a, i32 %b) {
+  %div = sdiv i32 %a, %b
+  call void @pad()
+  store i32 0, i32* @glbl
+  ret i32 %div
+}
+
+
+define i32 @outer_div3(i32 %a) {
+; CHECK-LABEL: @outer_div3(
+; CHECK-NOT: call i32 @div
+  %C = call i32 @div2(i32 %a)
+  ret i32 %C
+}
+
+define i32 @div2(i32 %a) {
+  %div = sdiv i32 %a, %a
+  call void @pad()
+  ret i32 %div
+}
+
+
+
+define i32 @outer_rem1(i32 %a) {
+; CHECK-LABEL: @outer_rem1(
+; CHECK-NOT: call i32 @rem
+  %C = call i32 @rem1(i32 0, i32 %a)
+  ret i32 %C
+}
+
+define i32 @outer_rem2(i32 %a) {
+; CHECK-LABEL: @outer_rem2(
+; CHECK-NOT: call i32 @rem
+  %C = call i32 @rem1(i32 %a, i32 1)
+  ret i32 %C
+}
+
+define i32 @rem1(i32 %a, i32 %b) {
+  %rem = urem i32 %a, %b
+  call void @pad()
+  store i32 0, i32* @glbl
+  ret i32 %rem
+}
+
+
+define i32 @outer_rem3(i32 %a) {
+; CHECK-LABEL: @outer_rem3(
+; CHECK-NOT: call i32 @rem
+  %C = call i32 @rem2(i32 %a)
+  ret i32 %C
+}
+
+define i32 @rem2(i32 %a) {
+  %rem = urem i32 %a, %a
+  call void @pad()
+  ret i32 %rem
+}
+
+
+
+define i32 @outer_shl1(i32 %a) {
+; CHECK-LABEL: @outer_shl1(
+; CHECK-NOT: call i32 @shl
+  %C = call i32 @shl(i32 %a, i32 0)
+  ret i32 %C
+}
+
+define i32 @shl(i32 %a, i32 %b) {
+  %shl = shl i32 %a, %b
+  call void @pad()
+  store i32 0, i32* @glbl
+  ret i32 %shl
+}
+
+
+
+define i32 @outer_shr1(i32 %a) {
+; CHECK-LABEL: @outer_shr1(
+; CHECK-NOT: call i32 @shr
+  %C = call i32 @shr(i32 %a, i32 0)
+  ret i32 %C
+}
+
+define i32 @shr(i32 %a, i32 %b) {
+  %shr = ashr i32 %a, %b
+  call void @pad()
+  store i32 0, i32* @glbl
+  ret i32 %shr
+}
+
+
+
+define i1 @outer_and1(i1 %a) {
+; check-label: @outer_and1(
+; check-not: call i1 @and1
+  %c = call i1 @and1(i1 %a, i1 false)
+  ret i1 %c
+}
+
+define i1 @outer_and2(i1 %a) {
+; check-label: @outer_and2(
+; check-not: call i1 @and1
+  %c = call i1 @and1(i1 %a, i1 true)
+  ret i1 %c
+}
+
+define i1 @and1(i1 %a, i1 %b) {
+  %and = and i1 %a, %b
+  call void @pad()
+  store i32 0, i32* @glbl
+  ret i1 %and
+}
+
+
+define i1 @outer_and3(i1 %a) {
+; check-label: @outer_and3(
+; check-not: call i1 @and2
+  %c = call i1 @and2(i1 %a)
+  ret i1 %c
+}
+
+define i1 @and2(i1 %a) {
+  %and = and i1 %a, %a
+  call void @pad()
+  ret i1 %and
+}
+
+
+
+define i1 @outer_or1(i1 %a) {
+; check-label: @outer_or1(
+; check-not: call i1 @or1
+  %c = call i1 @or1(i1 %a, i1 false)
+  ret i1 %c
+}
+
+define i1 @outer_or2(i1 %a) {
+; check-label: @outer_or2(
+; check-not: call i1 @or1
+  %c = call i1 @or1(i1 %a, i1 true)
+  ret i1 %c
+}
+
+define i1 @or1(i1 %a, i1 %b) {
+  %or = or i1 %a, %b
+  call void @pad()
+  store i32 0, i32* @glbl
+  ret i1 %or
+}
+
+
+define i1 @outer_or3(i1 %a) {
+; check-label: @outer_or3(
+; check-not: call i1 @or2
+  %c = call i1 @or2(i1 %a)
+  ret i1 %c
+}
+
+define i1 @or2(i1 %a) {
+  %or = or i1 %a, %a
+  call void @pad()
+  ret i1 %or
+}
+
+
+
+define i1 @outer_xor1(i1 %a) {
+; check-label: @outer_xor1(
+; check-not: call i1 @xor
+  %c = call i1 @xor1(i1 %a, i1 false)
+  ret i1 %c
+}
+
+define i1 @xor1(i1 %a, i1 %b) {
+  %xor = xor i1 %a, %b
+  call void @pad()
+  store i32 0, i32* @glbl
+  ret i1 %xor
+}
+
+
+define i1 @outer_xor3(i1 %a) {
+; check-label: @outer_xor3(
+; check-not: call i1 @xor
+  %c = call i1 @xor2(i1 %a)
+  ret i1 %c
+}
+
+define i1 @xor2(i1 %a) {
+  %xor = xor i1 %a, %a
+  call void @pad()
+  ret i1 %xor
+}