[flang][NFC] make IR generation for ieee_next deterministic

[jeanPerier]

C++ function call argument evaluation order is unspecified. When piping functions that generates IR, this creates indeterminism in the IR generated by flang (e.g., depend which C++ compiler compiled flang).

While it has interesting fuzzing property for flang (Initially, most of expression lowering used that pattern for binary operation where Fortran also does not specify any order, and we did found bugs exposed by some of the IR version and not the other), it is not ideal for lit tests (I found this because of a CI failure when not properly adding -DAG in test updates), and many people also really value bit to bit reproducibility from compilers.

[llvmbot]

@llvm/pr-subscribers-flang-fir-hlfir

Author: None (jeanPerier)

Changes

C++ function call argument evaluation order is unspecified. When piping functions that generates IR, this creates indeterminism in the IR generated by flang (e.g., depend which C++ compiler compiled flang).

While it has interesting fuzzing property for flang (Initially, most of expression lowering used that pattern for binary operation where Fortran also does not specify any order, and we did found bugs exposed by some of the IR version and not the other), it is not ideal for lit tests (I found this because of a CI failure when not properly adding -DAG in test updates), and many people also really value bit to bit reproducibility from compilers.

---

Full diff: https://github.com/llvm/llvm-project/pull/125055.diff

1 Files Affected:

• (modified) flang/lib/Optimizer/Builder/IntrinsicCall.cpp (+8-7)

diff --git a/flang/lib/Optimizer/Builder/IntrinsicCall.cpp b/flang/lib/Optimizer/Builder/IntrinsicCall.cpp
index 970fb2b1c2f701..a6a77dd58677b1 100644
--- a/flang/lib/Optimizer/Builder/IntrinsicCall.cpp
+++ b/flang/lib/Optimizer/Builder/IntrinsicCall.cpp
@@ -6366,10 +6366,12 @@ mlir::Value IntrinsicLibrary::genNearest(mlir::Type resultType,
   mlir::FloatType xType = mlir::dyn_cast<mlir::FloatType>(x.getType());
   const unsigned xBitWidth = xType.getWidth();
   mlir::Type i1Ty = builder.getI1Type();
-  if constexpr (proc == NearestProc::NextAfter)
+  if constexpr (proc == NearestProc::NextAfter) {
     // If isNan(Y), set X to a qNaN that will propagate to the resultIsX result.
-    x = builder.create<mlir::arith::SelectOp>(
-        loc, genIsFPClass(i1Ty, args[1], nanTest), genQNan(xType), x);
+    mlir::Value qNan = genQNan(xType);
+    mlir::Value isFPClass = genIsFPClass(i1Ty, args[1], nanTest);
+    x = builder.create<mlir::arith::SelectOp>(loc, isFPClass, qNan, x);
+  }
   mlir::Value resultIsX = genIsFPClass(i1Ty, x, nanTest);
   mlir::Type intType = builder.getIntegerType(xBitWidth);
   mlir::Value one = builder.createIntegerConstant(loc, intType, 1);
@@ -6489,12 +6491,11 @@ mlir::Value IntrinsicLibrary::genNearest(mlir::Type resultType,
   } else {
     // Kind 2, 3, 4, 8, 16. Increment or decrement X cast to integer.
     mlir::Value intX = builder.create<mlir::arith::BitcastOp>(loc, intType, x);
+    mlir::Value add = builder.create<mlir::arith::AddIOp>(loc, intX, one);
+    mlir::Value sub = builder.create<mlir::arith::SubIOp>(loc, intX, one);
     result = builder.create<mlir::arith::BitcastOp>(
         loc, resultType,
-        builder.create<mlir::arith::SelectOp>(
-            loc, magnitudeUp,
-            builder.create<mlir::arith::AddIOp>(loc, intX, one),
-            builder.create<mlir::arith::SubIOp>(loc, intX, one)));
+        builder.create<mlir::arith::SelectOp>(loc, magnitudeUp, add, sub));
     if constexpr (proc == NearestProc::Nearest ||
                   proc == NearestProc::NextAfter) {
       genRaiseExcept(_FORTRAN_RUNTIME_IEEE_OVERFLOW |