Index: llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp
===================================================================
--- llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp
+++ llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp
@@ -620,10 +620,12 @@
     return nullptr;
 
   Type *Ty = Den->getType();
-  Function *Decl = Intrinsic::getDeclaration(Mod, Intrinsic::amdgcn_rcp, Ty);
   if (const ConstantFP *CLHS = dyn_cast<ConstantFP>(Num)) {
     if (AllowInaccurateRcp || RcpIsAccurate) {
       if (CLHS->isExactlyValue(1.0)) {
+        Function *Decl = Intrinsic::getDeclaration(
+          Mod, Intrinsic::amdgcn_rcp, Ty);
+
         // v_rcp_f32 and v_rsq_f32 do not support denormals, and according to
         // the CI documentation has a worst case error of 1 ulp.
         // OpenCL requires <= 2.5 ulp for 1.0 / x, so it should always be OK to
@@ -636,10 +638,13 @@
 
         // 1.0 / x -> rcp(x)
         return Builder.CreateCall(Decl, { Den });
-       }
+      }
 
        // Same as for 1.0, but expand the sign out of the constant.
-       if (CLHS->isExactlyValue(-1.0)) {
+      if (CLHS->isExactlyValue(-1.0)) {
+        Function *Decl = Intrinsic::getDeclaration(
+          Mod, Intrinsic::amdgcn_rcp, Ty);
+
          // -1.0 / x -> rcp (fneg x)
          Value *FNeg = Builder.CreateFNeg(Den);
          return Builder.CreateCall(Decl, { FNeg });
@@ -648,6 +653,9 @@
   }
 
   if (AllowInaccurateRcp) {
+    Function *Decl = Intrinsic::getDeclaration(
+      Mod, Intrinsic::amdgcn_rcp, Ty);
+
     // Turn into multiply by the reciprocal.
     // x / y -> x * (1.0 / y)
     Value *Recip = Builder.CreateCall(Decl, { Den });