Add nvvm sqrt

Author: wsmoses

enzyme/CMakeLists.txt

@@ -75,14 +75,6 @@ else()
 file(READ ${LLVM_IDIR}/llvm/Analysis/ScalarEvolutionExpander.h INPUT_TEXT)
 endif()
 
-find_package(MPI)
-if (${MPI_FOUND})
-    add_definitions(-DBUILDMPI)
-    include_directories(SYSTEM ${MPI_C_INCLUDE_PATH})
-else()
-    set(MPI_C_LIBRARIES "")
-endif()
-
 if (${LLVM_VERSION_MAJOR} LESS 12)
 string(REPLACE "#define LLVM_ANALYSIS_SCALAREVOLUTIONEXPANDER_H" "#define LLVM_ANALYSIS_SCALAREVOLUTIONEXPANDER_H\n#include \"SCEV/ScalarEvolution.h\""  INPUT_TEXT "${INPUT_TEXT}")
 string(REPLACE "LLVM_ANALYSIS" "FAKELLVM_ANALYSIS" INPUT_TEXT "${INPUT_TEXT}")

enzyme/Enzyme/AdjointGenerator.h

@@ -1956,6 +1956,7 @@ class AdjointGenerator
       case Intrinsic::nearbyint:
       case Intrinsic::round:
       case Intrinsic::sqrt:
+      case Intrinsic::nvvm_sqrt_rn_d:
       case Intrinsic::fma:
         return;
       default:
@@ -2071,6 +2072,7 @@ class AdjointGenerator
         return;
       }
 
+      case Intrinsic::nvvm_sqrt_rn_d:
       case Intrinsic::sqrt: {
         if (vdiff && !gutils->isConstantValue(orig_ops[0])) {
           SmallVector<Value *, 2> args = {

enzyme/Enzyme/Enzyme.cpp

@@ -147,7 +147,7 @@ class Enzyme : public ModulePass {
     bool AtomicAdd = Arch == Triple::nvptx || Arch == Triple::nvptx64 ||
                      Arch == Triple::amdgcn;
 
-    std::map<int, Type*> byVal;
+    std::map<int, Type *> byVal;
     for (unsigned i = 1; i < CI->getNumArgOperands(); ++i) {
       Value *res = CI->getArgOperand(i);
 
@@ -358,9 +358,11 @@ class Enzyme : public ModulePass {
         }
         res = Builder.CreateBitCast(res, PTy);
       }
+#if LLVM_VERSION_MAJOR >= 9
       if (CI->isByValArgument(i)) {
         byVal[args.size()] = CI->getParamByValType(i);
       }
+#endif
       args.push_back(res);
       if (ty == DIFFE_TYPE::DUP_ARG || ty == DIFFE_TYPE::DUP_NONEED) {
         ++i;
@@ -479,9 +481,13 @@ class Enzyme : public ModulePass {
     CallInst *diffret = cast<CallInst>(Builder.CreateCall(newFunc, args));
     diffret->setCallingConv(CI->getCallingConv());
     diffret->setDebugLoc(CI->getDebugLoc());
+#if LLVM_VERSION_MAJOR >= 9
     for (auto pair : byVal) {
-        diffret->addParamAttr(pair.first, Attribute::getWithByValType(diffret->getContext(), pair.second));
+      diffret->addParamAttr(
+          pair.first,
+          Attribute::getWithByValType(diffret->getContext(), pair.second));
     }
+#endif
 
     if (!diffret->getType()->isEmptyTy() && !diffret->getType()->isVoidTy()) {
       unsigned idxs[] = {0};

enzyme/test/Enzyme/ReverseMode/byval.ll

@@ -490,8 +490,11 @@ attributes #11 = { noreturn }
 !21 = !{!22, !22, i64 0}
 !22 = !{!"int", !5, i64 0}
 
+; CHECK: define dso_local double @_Z14dcar_erg_atpos3card(%class.car* nocapture readonly byval(%class.car) align 8 %car1, double %pos)
+; CHECK: call { double } @diffe_Z13car_erg_atpos3card(%class.car* nonnull byval(%class.car) %car1, double %pos, double 1.000000e+00)
+
 ; CHECK: define internal { double } @diffe_Z13car_erg_atpos3card(%class.car* byval(%class.car) align 8 %car1, double %pos, double %differeturn)
 ; CHECK-NEXT: entry:
 ; CHECK-NEXT:   %"car1'ipa" = alloca %class.car
 ; CHECK-NEXT:   %0 = bitcast %class.car* %"car1'ipa" to i8*
-; CHECK-NEXT:   call void @llvm.memset.p0i8.i64(i8* nonnull %0, i8 0, i64 24, i1 false)
+; CHECK-NEXT:   call void @llvm.memset.p0i8.i64(i8* nonnull %0, i8 0, i64 24, i1 false)