[SimplifyLibCall] pow(x, 3.0) -> x*x*x

[mcrosier]

Summary

Please take a look.

Chad

[joerg]

Am I reading this correctly and this transform is done by default? If yes, this seems to be a bit problematic, since it replaces a potentially correctly rounded operation with something that does rounding twice?

[mcrosier]

@joerg: you're correct. This should be guarded by fast math.

lib/Transforms/Utils/SimplifyLibCalls.cpp

     return Op1;
   if (Op2C->isExactlyValue(2.0)) // pow(x, 2.0) -> x*x
     return B.CreateFMul(Op1, Op1, "pow2");
+  if (Op2C->isExactlyValue(3.0)) { // pow(x, 3.0) -> x*x*x
+    Value *Pow2 = B.CreateFMul(Op1, Op1, "pow3");
+    return B.CreateFMul(Pow2, Op1, "pow3");
+  }
   if (Op2C->isExactlyValue(-1.0)) // pow(x, -1.0) -> 1.0/x
     return B.CreateFDiv(ConstantFP::get(CI->getType(), 1.0), Op1, "powrecip");
   return nullptr;

test/Transforms/InstCombine/pow-1.ll

 ; CHECK-NO-EXP10: call double @pow
 }
 
+; Check pow(x, 3.0) -> x*x*x.
+
+define float @test_simplify20(float %x) {
+; CHECK-LABEL: @test_simplify20(
+  %retval = call float @powf(float %x, float 3.0)
+; CHECK-NEXT: [[SQUARE:%[a-z0-9]+]] = fmul float %x, %x
+; CHECK-NEXT: [[CUBE:%[a-z0-9]+]] = fmul float [[SQUARE]], %x
+  ret float %retval
+; CHECK-NEXT: ret float [[CUBE]]
+}
+
+define double @test_simplify21(double %x) {
+; CHECK-LABEL: @test_simplify21(
+  %retval = call double @pow(double %x, double 3.0)
+; CHECK-NEXT: [[SQUARE:%[a-z0-9]+]] = fmul double %x, %x
+; CHECK-NEXT: [[CUBE:%[a-z0-9]+]] = fmul double [[SQUARE]], %x
+  ret double %retval
+; CHECK-NEXT: ret double [[CUBE]]
+}
+
 ; CHECK: attributes [[NUW_RO]] = { nounwind readonly }