Index: llvm/trunk/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
===================================================================
--- llvm/trunk/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
+++ llvm/trunk/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
@@ -531,6 +531,11 @@
     }
   }
 
+  // sqrt(X) * sqrt(X) -> X
+  if (AllowReassociate && (Op0 == Op1))
+    if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op0))
+      if (II->getIntrinsicID() == Intrinsic::sqrt)
+        return ReplaceInstUsesWith(I, II->getOperand(0));
 
   // Under unsafe algebra do:
   // X * log2(0.5*Y) = X*log2(Y) - X
Index: llvm/trunk/test/Transforms/InstCombine/fmul.ll
===================================================================
--- llvm/trunk/test/Transforms/InstCombine/fmul.ll
+++ llvm/trunk/test/Transforms/InstCombine/fmul.ll
@@ -123,3 +123,32 @@
 ; CHECK-NOT: fadd float
 ; CHECK: fadd fast float
 }
+
+; PR21126: http://llvm.org/bugs/show_bug.cgi?id=21126
+; With unsafe/fast math, sqrt(X) * sqrt(X) is just X.
+declare double @llvm.sqrt.f64(double)
+
+define double @sqrt_squared1(double %f) #0 {
+  %sqrt = call double @llvm.sqrt.f64(double %f)
+  %mul = fmul fast double %sqrt, %sqrt
+  ret double %mul
+; CHECK-LABEL: @sqrt_squared1(
+; CHECK-NEXT: ret double %f
+}
+
+; With unsafe/fast math, sqrt(X) * sqrt(X) is just X, 
+; but make sure another use of the sqrt is intact.
+; Note that the remaining fmul is altered but is not 'fast'
+; itself because it was not marked 'fast' originally. 
+; Thus, we have an overall fast result, but no more indication of
+; 'fast'ness in the code.
+define double @sqrt_squared2(double %f) #0 {
+  %sqrt = call double @llvm.sqrt.f64(double %f)
+  %mul1 = fmul fast double %sqrt, %sqrt
+  %mul2 = fmul double %mul1, %sqrt
+  ret double %mul2
+; CHECK-LABEL: @sqrt_squared2(
+; CHECK-NEXT: %sqrt = call double @llvm.sqrt.f64(double %f)
+; CHECK-NEXT: %mul2 = fmul double %sqrt, %f
+; CHECK-NEXT: ret double %mul2
+}