[SDAG] Now that we have a way to communicate the exact bit on sdiv use it to simplify sdiv by a constant.

We had a hack in SDAGBuilder in place to work around this but now we
can avoid that. Call BuildExactSDIV from BuildSDIV so DAGCombiner can
perform this trick automatically.

The added check in DAGCombiner is necessary to prevent exact sdiv by pow2
from regressing as the target-specific pow2 lowering is not aware of
exact bits yet.

This is mostly covered by existing tests. One side effect is that we
get the better lowering for exact vector sdivs now too :)

llvm-svn: 240891

llvm/include/llvm/Target/TargetLowering.h

@@ -2718,8 +2718,6 @@ public:
   //===--------------------------------------------------------------------===//
   // Div utility functions
   //
-  SDValue BuildExactSDIV(SDValue Op1, SDValue Op2, SDLoc dl,
-                         SelectionDAG &DAG) const;
   SDValue BuildSDIV(SDNode *N, const APInt &Divisor, SelectionDAG &DAG,
                     bool IsAfterLegalization,
                     std::vector<SDNode *> *Created) const;

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

@@ -2178,7 +2178,11 @@ SDValue DAGCombiner::visitSDIV(SDNode *N) {
   }
 
   // fold (sdiv X, pow2) -> simple ops after legalize
+  // FIXME: We check for the exact bit here because the generic lowering gives
+  // better results in that case. The target-specific lowering should learn how
+  // to handle exact sdivs efficiently.
   if (N1C && !N1C->isNullValue() && !N1C->isOpaque() &&
+      !cast<BinaryWithFlagsSDNode>(N)->Flags.hasExact() &&
       (N1C->getAPIntValue().isPowerOf2() ||
        (-N1C->getAPIntValue()).isPowerOf2())) {
     // If dividing by powers of two is cheap, then don't perform the following

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

@@ -2238,17 +2238,11 @@ void SelectionDAGBuilder::visitSDiv(const User &I) {
   SDValue Op1 = getValue(I.getOperand(0));
   SDValue Op2 = getValue(I.getOperand(1));
 
-  // Turn exact SDivs into multiplications.
-  // FIXME: This should be in DAGCombiner, but it doesn't have access to the
-  // exact bit.
-  if (isa<BinaryOperator>(&I) && cast<BinaryOperator>(&I)->isExact() &&
-      !isa<ConstantSDNode>(Op1) &&
-      isa<ConstantSDNode>(Op2) && !cast<ConstantSDNode>(Op2)->isNullValue())
-    setValue(&I, DAG.getTargetLoweringInfo()
-                     .BuildExactSDIV(Op1, Op2, getCurSDLoc(), DAG));
-  else
-    setValue(&I, DAG.getNode(ISD::SDIV, getCurSDLoc(), Op1.getValueType(),
-                             Op1, Op2));
+  SDNodeFlags Flags;
+  Flags.setExact(isa<PossiblyExactOperator>(&I) &&
+                 cast<PossiblyExactOperator>(&I)->isExact());
+  setValue(&I, DAG.getNode(ISD::SDIV, getCurSDLoc(), Op1.getValueType(), Op1,
+                           Op2, &Flags));
 }
 
 void SelectionDAGBuilder::visitICmp(const User &I) {

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp

@@ -2674,10 +2674,9 @@ void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo,
 
 /// \brief Given an exact SDIV by a constant, create a multiplication
 /// with the multiplicative inverse of the constant.
-SDValue TargetLowering::BuildExactSDIV(SDValue Op1, SDValue Op2, SDLoc dl,
-                                       SelectionDAG &DAG) const {
-  ConstantSDNode *C = cast<ConstantSDNode>(Op2);
-  APInt d = C->getAPIntValue();
+static SDValue BuildExactSDIV(const TargetLowering &TLI, SDValue Op1, APInt d,
+                              SDLoc dl, SelectionDAG &DAG,
+                              std::vector<SDNode *> &Created) {
   assert(d != 0 && "Division by zero!");
 
   // Shift the value upfront if it is even, so the LSB is one.
@@ -2685,10 +2684,11 @@ SDValue TargetLowering::BuildExactSDIV(SDValue Op1, SDValue Op2, SDLoc dl,
   if (ShAmt) {
     // TODO: For UDIV use SRL instead of SRA.
     SDValue Amt =
-        DAG.getConstant(ShAmt, dl, getShiftAmountTy(Op1.getValueType()));
+        DAG.getConstant(ShAmt, dl, TLI.getShiftAmountTy(Op1.getValueType()));
     SDNodeFlags Flags;
     Flags.setExact(true);
     Op1 = DAG.getNode(ISD::SRA, dl, Op1.getValueType(), Op1, Amt, &Flags);
+    Created.push_back(Op1.getNode());
     d = d.ashr(ShAmt);
   }
 
@@ -2697,8 +2697,10 @@ SDValue TargetLowering::BuildExactSDIV(SDValue Op1, SDValue Op2, SDLoc dl,
   while ((t = d*xn) != 1)
     xn *= APInt(d.getBitWidth(), 2) - t;
 
-  Op2 = DAG.getConstant(xn, dl, Op1.getValueType());
-  return DAG.getNode(ISD::MUL, dl, Op1.getValueType(), Op1, Op2);
+  SDValue Op2 = DAG.getConstant(xn, dl, Op1.getValueType());
+  SDValue Mul = DAG.getNode(ISD::MUL, dl, Op1.getValueType(), Op1, Op2);
+  Created.push_back(Mul.getNode());
+  return Mul;
 }
 
 /// \brief Given an ISD::SDIV node expressing a divide by constant,
@@ -2718,6 +2720,10 @@ SDValue TargetLowering::BuildSDIV(SDNode *N, const APInt &Divisor,
   if (!isTypeLegal(VT))
     return SDValue();
 
+  // If the sdiv has an 'exact' bit we can use a simpler lowering.
+  if (cast<BinaryWithFlagsSDNode>(N)->Flags.hasExact())
+    return BuildExactSDIV(*this, N->getOperand(0), Divisor, dl, DAG, *Created);
+
   APInt::ms magics = Divisor.magic();
 
   // Multiply the numerator (operand 0) by the magic value

llvm/test/CodeGen/X86/sdiv-exact.ll

@@ -1,4 +1,4 @@
-; RUN: llc -march=x86 < %s | FileCheck %s
+; RUN: llc -march=x86 -mattr=+sse2 < %s | FileCheck %s
 
 define i32 @test1(i32 %x) {
   %div = sdiv exact i32 %x, 25
@@ -16,3 +16,14 @@ define i32 @test2(i32 %x) {
 ; CHECK-NEXT: imull	$-1431655765
 ; CHECK-NEXT: ret
 }
+
+define <4 x i32> @test3(<4 x i32> %x) {
+  %div = sdiv exact <4 x i32> %x, <i32 24, i32 24, i32 24, i32 24>
+  ret <4 x i32> %div
+; CHECK-LABEL: test3:
+; CHECK: psrad	$3,
+; CHECK: pmuludq
+; CHECK: pmuludq
+; CHECK-NOT: psrad
+; CHECK: ret
+}