[InstCombine] Optimise x / sqrt(y / z) with fast-math pattern.

Replace the pattern with
x * sqrt(z/y)

Author: zjaffal

llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp

@@ -1709,6 +1709,31 @@ static Instruction *foldFDivPowDivisor(BinaryOperator &I,
   return BinaryOperator::CreateFMulFMF(Op0, Pow, &I);
 }
 
+/// Convert div to mul if we have an sqrt divisor iff sqrt's operand is a fdiv
+/// instruction.
+static Instruction *foldFDivSqrtDivisor(BinaryOperator &I,
+                                        InstCombiner::BuilderTy &Builder) {
+  // X / sqrt(Y / Z) -->  X * sqrt(Z / Y)
+  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
+  auto *II = dyn_cast<IntrinsicInst>(Op1);
+  if (!II || II->getIntrinsicID() != Intrinsic::sqrt || !II->hasOneUse() ||
+      !I.hasAllowReassoc() || !I.hasAllowReciprocal())
+    return nullptr;
+
+  Value *Y, *Z;
+  auto *DivOp = dyn_cast<Instruction>(II->getOperand(0));
+  if (!DivOp || !DivOp->hasOneUse() || !DivOp->hasAllowReassoc() ||
+      !I.hasAllowReciprocal())
+    return nullptr;
+  if (match(DivOp, m_FDiv(m_Value(Y), m_Value(Z)))) {
+    Value *SwapDiv = Builder.CreateFDivFMF(Z, Y, DivOp);
+    Value *NewSqrt = Builder.CreateIntrinsic(II->getIntrinsicID(),
+                                             II->getType(), {SwapDiv}, II);
+    return BinaryOperator::CreateFMulFMF(Op0, NewSqrt, &I);
+  }
+  return nullptr;
+}
+
 Instruction *InstCombinerImpl::visitFDiv(BinaryOperator &I) {
   Module *M = I.getModule();
 
@@ -1816,6 +1841,9 @@ Instruction *InstCombinerImpl::visitFDiv(BinaryOperator &I) {
   if (Instruction *Mul = foldFDivPowDivisor(I, Builder))
     return Mul;
 
+  if (Instruction *Mul = foldFDivSqrtDivisor(I, Builder))
+    return Mul;
+
   // pow(X, Y) / X --> pow(X, Y-1)
   if (I.hasAllowReassoc() &&
       match(Op0, m_OneUse(m_Intrinsic<Intrinsic::pow>(m_Specific(Op1),

llvm/test/Transforms/InstCombine/fdiv-sqrt.ll

@@ -6,9 +6,9 @@ declare double @llvm.sqrt.f64(double)
 define double @sqrt_div_fast(double %x, double %y, double %z) {
 ; CHECK-LABEL: @sqrt_div_fast(
 ; CHECK-NEXT:  entry:
-; CHECK-NEXT:    [[DIV:%.*]] = fdiv fast double [[Y:%.*]], [[Z:%.*]]
-; CHECK-NEXT:    [[SQRT:%.*]] = call fast double @llvm.sqrt.f64(double [[DIV]])
-; CHECK-NEXT:    [[DIV1:%.*]] = fdiv fast double [[X:%.*]], [[SQRT]]
+; CHECK-NEXT:    [[TMP0:%.*]] = fdiv fast double [[Z:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[TMP1:%.*]] = call fast double @llvm.sqrt.f64(double [[TMP0]])
+; CHECK-NEXT:    [[DIV1:%.*]] = fmul fast double [[TMP1]], [[X:%.*]]
 ; CHECK-NEXT:    ret double [[DIV1]]
 ;
 entry:
@@ -36,9 +36,9 @@ entry:
 define double @sqrt_div_reassoc_arcp(double %x, double %y, double %z) {
 ; CHECK-LABEL: @sqrt_div_reassoc_arcp(
 ; CHECK-NEXT:  entry:
-; CHECK-NEXT:    [[DIV:%.*]] = fdiv reassoc arcp double [[Y:%.*]], [[Z:%.*]]
-; CHECK-NEXT:    [[SQRT:%.*]] = call reassoc arcp double @llvm.sqrt.f64(double [[DIV]])
-; CHECK-NEXT:    [[DIV1:%.*]] = fdiv reassoc arcp double [[X:%.*]], [[SQRT]]
+; CHECK-NEXT:    [[TMP0:%.*]] = fdiv reassoc arcp double [[Z:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[TMP1:%.*]] = call reassoc arcp double @llvm.sqrt.f64(double [[TMP0]])
+; CHECK-NEXT:    [[DIV1:%.*]] = fmul reassoc arcp double [[TMP1]], [[X:%.*]]
 ; CHECK-NEXT:    ret double [[DIV1]]
 ;
 entry:
@@ -66,9 +66,9 @@ entry:
 define double @sqrt_div_reassoc_missing2(double %x, double %y, double %z) {
 ; CHECK-LABEL: @sqrt_div_reassoc_missing2(
 ; CHECK-NEXT:  entry:
-; CHECK-NEXT:    [[DIV:%.*]] = fdiv reassoc arcp double [[Y:%.*]], [[Z:%.*]]
-; CHECK-NEXT:    [[SQRT:%.*]] = call arcp double @llvm.sqrt.f64(double [[DIV]])
-; CHECK-NEXT:    [[DIV1:%.*]] = fdiv reassoc arcp double [[X:%.*]], [[SQRT]]
+; CHECK-NEXT:    [[TMP0:%.*]] = fdiv reassoc arcp double [[Z:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[TMP1:%.*]] = call arcp double @llvm.sqrt.f64(double [[TMP0]])
+; CHECK-NEXT:    [[DIV1:%.*]] = fmul reassoc arcp double [[TMP1]], [[X:%.*]]
 ; CHECK-NEXT:    ret double [[DIV1]]
 ;
 entry:
@@ -96,9 +96,9 @@ entry:
 define double @sqrt_div_arcp_missing(double %x, double %y, double %z) {
 ; CHECK-LABEL: @sqrt_div_arcp_missing(
 ; CHECK-NEXT:  entry:
-; CHECK-NEXT:    [[DIV:%.*]] = fdiv reassoc double [[Y:%.*]], [[Z:%.*]]
-; CHECK-NEXT:    [[SQRT:%.*]] = call reassoc arcp double @llvm.sqrt.f64(double [[DIV]])
-; CHECK-NEXT:    [[DIV1:%.*]] = fdiv reassoc arcp double [[X:%.*]], [[SQRT]]
+; CHECK-NEXT:    [[TMP0:%.*]] = fdiv reassoc double [[Z:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[TMP1:%.*]] = call reassoc arcp double @llvm.sqrt.f64(double [[TMP0]])
+; CHECK-NEXT:    [[DIV1:%.*]] = fmul reassoc arcp double [[TMP1]], [[X:%.*]]
 ; CHECK-NEXT:    ret double [[DIV1]]
 ;
 entry:
@@ -111,9 +111,9 @@ entry:
 define double @sqrt_div_arcp_missing2(double %x, double %y, double %z) {
 ; CHECK-LABEL: @sqrt_div_arcp_missing2(
 ; CHECK-NEXT:  entry:
-; CHECK-NEXT:    [[DIV:%.*]] = fdiv reassoc arcp double [[Y:%.*]], [[Z:%.*]]
-; CHECK-NEXT:    [[SQRT:%.*]] = call reassoc double @llvm.sqrt.f64(double [[DIV]])
-; CHECK-NEXT:    [[DIV1:%.*]] = fdiv reassoc arcp double [[X:%.*]], [[SQRT]]
+; CHECK-NEXT:    [[TMP0:%.*]] = fdiv reassoc arcp double [[Z:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[TMP1:%.*]] = call reassoc double @llvm.sqrt.f64(double [[TMP0]])
+; CHECK-NEXT:    [[DIV1:%.*]] = fmul reassoc arcp double [[TMP1]], [[X:%.*]]
 ; CHECK-NEXT:    ret double [[DIV1]]
 ;
 entry: