================
@@ -1250,6 +1253,392 @@ getSafeRepackAttrs(Fortran::lower::AbstractConverter
&converter) {
return attrs.empty() ? mlir::ArrayAttr{} : builder.getArrayAttr(attrs);
}
+//===----------------------------------------------------------------------===//
+// -finit-local= helpers
+//===----------------------------------------------------------------------===//
+
+/// Returns true when \p var is an automatic local variable eligible for
+/// -finit-local= initialization. Excluded: variables without a symbol,
+/// globals, dummy arguments, SAVE'd vars, ALLOCATABLE/POINTER, vars in
+/// an EQUIVALENCE set, vars with explicit or default initialization, and
+/// CUDA variables whose storage is not host-accessible (device, constant,
+/// shared, usedevice).
+static bool shouldInitLocal(const Fortran::lower::pft::Variable &var) {
+ if (!var.hasSymbol() || var.isGlobal())
+ return false;
+ const Fortran::semantics::Symbol &sym = var.getSymbol();
+ if (Fortran::semantics::IsDummy(sym))
+ return false;
+ if (Fortran::semantics::IsSaved(sym))
+ return false;
+ if (Fortran::semantics::IsAllocatableOrPointer(sym))
+ return false;
+ if (Fortran::lower::hasDefaultInitialization(sym))
+ return false;
+ if (const auto *obj =
+ sym.detailsIf<Fortran::semantics::ObjectEntityDetails>())
+ if (obj->init())
+ return false;
+ if (Fortran::semantics::FindEquivalenceSet(sym))
+ return false;
+ // Cray pointees own no storage of their own; their FIR base is a
+ // pointer-box descriptor. Initializing it would overwrite the
+ // descriptor, not the pointee storage.
+ if (sym.test(Fortran::semantics::Symbol::Flag::CrayPointee))
+ return false;
+ // CUDA storage accessibility:
+ // constant / shared / usedevice: always unreachable by a plain fir.store
+ // from the host -- skip.
+ // device in a HOST context: lives in global device memory; a host
+ // fir.store cannot reach it -- skip.
+ // device in a DEVICE subprogram: implicitly set by SetImplicitCUDADevice
+ // for every local in a device kernel; these are per-thread stack
+ // allocations reachable by a device fir.store -- initialize.
+ // managed / unified / pinned: host-accessible unified memory --
initialize.
+ if (auto cudaAttr = Fortran::semantics::GetCUDADataAttr(&sym)) {
+ switch (*cudaAttr) {
+ case Fortran::common::CUDADataAttr::Constant:
+ case Fortran::common::CUDADataAttr::Shared:
+ case Fortran::common::CUDADataAttr::UseDevice:
+ return false;
+ case Fortran::common::CUDADataAttr::Device:
+ // In a device subprogram the attribute is implicit
+ // (SetImplicitCUDADevice) and the variable is a thread-local stack
+ // allocation -- initialize it.
+ if (!Fortran::semantics::IsCUDADeviceContext(&sym.owner()))
+ return false;
+ break;
+ default:
+ break;
+ }
+ }
+ return true;
+}
+
+/// Build a constant whose every byte equals \p bytePat.
+/// FP types: bitcast from an integer splat. Complex: apply to both parts.
+/// LOGICAL(k): returns a raw iN integer (caller stores via bitcasted address).
+/// Character: falls back to fir.zero_bits (see TODO). Derived types are
+/// handled by the caller before this function is reached.
+static mlir::Value genByteSplatInit(fir::FirOpBuilder &builder,
+ mlir::Location loc, mlir::Type ty,
+ uint8_t bytePat) {
+ mlir::Type eleTy = fir::unwrapSequenceType(ty);
+
+ // Build an integer constant of the given bit width from a byte splat.
+ auto makeIntCst = [&](unsigned bits) -> mlir::Value {
+ llvm::APInt byteVal(8, bytePat);
+ llvm::APInt splat = llvm::APInt::getSplat(bits, byteVal);
+ mlir::Type intTy = builder.getIntegerType(bits);
+ return mlir::arith::ConstantOp::create(
+ builder, loc, intTy, builder.getIntegerAttr(intTy, splat));
+ };
+
+ if (auto fpTy = mlir::dyn_cast<mlir::FloatType>(eleTy)) {
+ unsigned bits = fpTy.getWidth();
+ mlir::Value intCst = makeIntCst(bits);
+ return mlir::arith::BitcastOp::create(builder, loc, fpTy, intCst);
+ }
+ if (auto intTy = mlir::dyn_cast<mlir::IntegerType>(eleTy)) {
+ return makeIntCst(intTy.getWidth());
+ }
+ // Complex: apply the byte pattern to each (real, imag) part.
+ if (auto cplxTy = mlir::dyn_cast<mlir::ComplexType>(eleTy)) {
+ mlir::Type partTy = cplxTy.getElementType();
+ mlir::Value partVal = genByteSplatInit(builder, loc, partTy, bytePat);
+ return mlir::complex::CreateOp::create(builder, loc, cplxTy, partVal,
+ partVal);
+ }
+ // LOGICAL(k) has a fixed size of k bytes. Return the raw integer splat;
+ // the caller stores it via a bitcasted address to preserve the bit pattern
+ // (fir.convert from integer to !fir.logical normalizes nonzero -> true).
+ if (auto logTy = mlir::dyn_cast<fir::LogicalType>(eleTy)) {
+ return makeIntCst(logTy.getFKind() * 8);
+ }
+ // Fallback (e.g. fir.char reached via an unexpected path): zero-initialise.
+ return fir::ZeroOp::create(builder, loc, eleTy);
----------------
DanielCChen wrote:
> Could hex mode avoid silently initializing unhandled types to zero? A PowerPC
> `vector(real(4))` local passes the eligibility check and lowers to
> `fir::VectorType`. The lowered local reaches this fallback. The fallback
> initializes the vector to zero instead of the requested byte pattern.
>
> The available build lacks a PowerPC target, so I could not run a PowerPC
> reproducer. Could `fir::VectorType` use a byte fill, or could the fallback
> emit a diagnostic for an unhandled type? Could you add a PowerPC LLVM
> regression test for the chosen behavior?
For the scope of this PR, I would prefer to stay with the standard types only.
We can certainly extend the coverage of the extension types like PPC vector
types in a future PR if needed.
https://github.com/llvm/llvm-project/pull/216164
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
