[Flang][MLIR] Add basic initial support for alloca and program address space handling in FIR->LLVMIR codegen (#77518)

This is a slightly more slimmed down and up-to-date version of the older
PR from here: https://reviews.llvm.org/D144203, written by @jsjodin,
which has already under gone some review.

This PR places allocas in the alloca address space specified by the
provided data layout (default is 0 for all address spaces, unless
explicitly specified by the layout), and then will cast these alloca's
to the program address space if this address space is different from the
allocation address space. For most architectures data layouts, this will
be a no-op, as they have a flat address space. But in the case of AMDGPU
it will result in allocas being placed in the correct address space (5,
private), and then casted into the correct program address space (0,
generic). This results in correct (partially, a follow up PR will be
forthcoming soon) generation of allocations inside of device code.

This PR is in addition to the work by @skatrak in this PR:
#69599 and adds seperate and
neccesary functionality of casting alloca's from their address space to
the program address space, both are independent PRs, although there is
some minor overlap e.g. this PR incorporates some of the useful helper
functions from 69599, so whichever lands first will need a minor rebase.

Co-author: jsjodin

Author: agozillon

flang/include/flang/Optimizer/CodeGen/CGPasses.td

@@ -27,6 +27,8 @@ def FIRToLLVMLowering : Pass<"fir-to-llvm-ir", "mlir::ModuleOp"> {
   let options = [
     Option<"forcedTargetTriple", "target", "std::string", /*default=*/"",
            "Override module's target triple.">,
+    Option<"forcedDataLayout", "datalayout", "std::string", /*default=*/"",
+           "Override module's data layout.">,
     Option<"applyTBAA", "apply-tbaa", "bool", /*default=*/"false",
            "Attach TBAA tags to memory accessing operations.">
   ];

flang/lib/Optimizer/CodeGen/CodeGen.cpp

@@ -44,6 +44,7 @@
 #include "mlir/IR/Matchers.h"
 #include "mlir/Pass/Pass.h"
 #include "mlir/Pass/PassManager.h"
+#include "mlir/Target/LLVMIR/Import.h"
 #include "mlir/Target/LLVMIR/ModuleTranslation.h"
 #include "llvm/ADT/ArrayRef.h"
 #include "llvm/ADT/TypeSwitch.h"
@@ -61,14 +62,40 @@ namespace fir {
 
 // TODO: This should really be recovered from the specified target.
 static constexpr unsigned defaultAlign = 8;
+static constexpr unsigned defaultAddressSpace = 0u;
 
 /// `fir.box` attribute values as defined for CFI_attribute_t in
 /// flang/ISO_Fortran_binding.h.
 static constexpr unsigned kAttrPointer = CFI_attribute_pointer;
 static constexpr unsigned kAttrAllocatable = CFI_attribute_allocatable;
 
-static inline mlir::Type getLlvmPtrType(mlir::MLIRContext *context) {
-  return mlir::LLVM::LLVMPointerType::get(context);
+static inline unsigned
+getAllocaAddressSpace(mlir::ConversionPatternRewriter &rewriter) {
+  mlir::Operation *parentOp = rewriter.getInsertionBlock()->getParentOp();
+  assert(parentOp != nullptr &&
+         "expected insertion block to have parent operation");
+  if (auto module = parentOp->getParentOfType<mlir::ModuleOp>())
+    if (mlir::Attribute addrSpace =
+            mlir::DataLayout(module).getAllocaMemorySpace())
+      return llvm::cast<mlir::IntegerAttr>(addrSpace).getUInt();
+  return defaultAddressSpace;
+}
+
+static inline unsigned
+getProgramAddressSpace(mlir::ConversionPatternRewriter &rewriter) {
+  mlir::Operation *parentOp = rewriter.getInsertionBlock()->getParentOp();
+  assert(parentOp != nullptr &&
+         "expected insertion block to have parent operation");
+  if (auto module = parentOp->getParentOfType<mlir::ModuleOp>())
+    if (mlir::Attribute addrSpace =
+            mlir::DataLayout(module).getProgramMemorySpace())
+      return llvm::cast<mlir::IntegerAttr>(addrSpace).getUInt();
+  return defaultAddressSpace;
+}
+
+static inline mlir::Type getLlvmPtrType(mlir::MLIRContext *context,
+                                        unsigned addressSpace = 0) {
+  return mlir::LLVM::LLVMPointerType::get(context, addressSpace);
 }
 
 static inline mlir::Type getI8Type(mlir::MLIRContext *context) {
@@ -368,19 +395,37 @@ class FIROpConversion : public mlir::ConvertOpToLLVMPattern<FromOp> {
     return getBlockForAllocaInsert(op->getParentOp());
   }
 
-  // Generate an alloca of size 1 for an object of type \p llvmObjectTy.
-  mlir::LLVM::AllocaOp
-  genAllocaWithType(mlir::Location loc, mlir::Type llvmObjectTy,
-                    unsigned alignment,
-                    mlir::ConversionPatternRewriter &rewriter) const {
+  // Generate an alloca of size 1 for an object of type \p llvmObjectTy in the
+  // allocation address space provided for the architecture in the DataLayout
+  // specification. If the address space is different from the devices
+  // program address space we perform a cast. In the case of most architectures
+  // the program and allocation address space will be the default of 0 and no
+  // cast will be emitted.
+  mlir::Value genAllocaAndAddrCastWithType(
+      mlir::Location loc, mlir::Type llvmObjectTy, unsigned alignment,
+      mlir::ConversionPatternRewriter &rewriter) const {
     auto thisPt = rewriter.saveInsertionPoint();
     mlir::Operation *parentOp = rewriter.getInsertionBlock()->getParentOp();
     mlir::Block *insertBlock = getBlockForAllocaInsert(parentOp);
     rewriter.setInsertionPointToStart(insertBlock);
     auto size = genI32Constant(loc, rewriter, 1);
-    mlir::Type llvmPtrTy = ::getLlvmPtrType(llvmObjectTy.getContext());
-    auto al = rewriter.create<mlir::LLVM::AllocaOp>(
-        loc, llvmPtrTy, llvmObjectTy, size, alignment);
+    unsigned allocaAs = getAllocaAddressSpace(rewriter);
+    unsigned programAs = getProgramAddressSpace(rewriter);
+
+    mlir::Value al = rewriter.create<mlir::LLVM::AllocaOp>(
+        loc, ::getLlvmPtrType(llvmObjectTy.getContext(), allocaAs),
+        llvmObjectTy, size, alignment);
+
+    // if our allocation address space, is not the same as the program address
+    // space, then we must emit a cast to the program address space before use.
+    // An example case would be on AMDGPU, where the allocation address space is
+    // the numeric value 5 (private), and the program address space is 0
+    // (generic).
+    if (allocaAs != programAs) {
+      al = rewriter.create<mlir::LLVM::AddrSpaceCastOp>(
+          loc, ::getLlvmPtrType(llvmObjectTy.getContext(), programAs), al);
+    }
+
     rewriter.restoreInsertionPoint(thisPt);
     return al;
   }
@@ -532,20 +577,34 @@ struct AllocaOpConversion : public FIROpConversion<fir::AllocaOp> {
         size = rewriter.create<mlir::LLVM::MulOp>(
             loc, ity, size, integerCast(loc, rewriter, ity, operands[i]));
     }
-    mlir::Type llvmPtrTy = ::getLlvmPtrType(alloc.getContext());
+
+    unsigned allocaAs = getAllocaAddressSpace(rewriter);
+    unsigned programAs = getProgramAddressSpace(rewriter);
+
     // NOTE: we used to pass alloc->getAttrs() in the builder for non opaque
     // pointers! Only propagate pinned and bindc_name to help debugging, but
     // this should have no functional purpose (and passing the operand segment
     // attribute like before is certainly bad).
     auto llvmAlloc = rewriter.create<mlir::LLVM::AllocaOp>(
-        loc, llvmPtrTy, llvmObjectType, size);
+        loc, ::getLlvmPtrType(alloc.getContext(), allocaAs), llvmObjectType,
+        size);
     if (alloc.getPinned())
       llvmAlloc->setDiscardableAttr(alloc.getPinnedAttrName(),
                                     alloc.getPinnedAttr());
     if (alloc.getBindcName())
       llvmAlloc->setDiscardableAttr(alloc.getBindcNameAttrName(),
                                     alloc.getBindcNameAttr());
-    rewriter.replaceOp(alloc, llvmAlloc);
+    if (allocaAs == programAs) {
+      rewriter.replaceOp(alloc, llvmAlloc);
+    } else {
+      // if our allocation address space, is not the same as the program address
+      // space, then we must emit a cast to the program address space before
+      // use. An example case would be on AMDGPU, where the allocation address
+      // space is the numeric value 5 (private), and the program address space
+      // is 0 (generic).
+      rewriter.replaceOpWithNewOp<mlir::LLVM::AddrSpaceCastOp>(
+          alloc, ::getLlvmPtrType(alloc.getContext(), programAs), llvmAlloc);
+    }
     return mlir::success();
   }
 };
@@ -1691,8 +1750,8 @@ struct EmboxCommonConversion : public FIROpConversion<OP> {
     if (isInGlobalOp(rewriter))
       return boxValue;
     mlir::Type llvmBoxTy = boxValue.getType();
-    auto alloca =
-        this->genAllocaWithType(loc, llvmBoxTy, defaultAlign, rewriter);
+    auto alloca = this->genAllocaAndAddrCastWithType(loc, llvmBoxTy,
+                                                     defaultAlign, rewriter);
     auto storeOp = rewriter.create<mlir::LLVM::StoreOp>(loc, boxValue, alloca);
     this->attachTBAATag(storeOp, boxTy, boxTy, nullptr);
     return alloca;
@@ -3110,11 +3169,11 @@ struct LoadOpConversion : public FIROpConversion<fir::LoadOp> {
       else
         attachTBAATag(boxValue, boxTy, boxTy, nullptr);
       auto newBoxStorage =
-          genAllocaWithType(loc, llvmLoadTy, defaultAlign, rewriter);
+          genAllocaAndAddrCastWithType(loc, llvmLoadTy, defaultAlign, rewriter);
       auto storeOp =
           rewriter.create<mlir::LLVM::StoreOp>(loc, boxValue, newBoxStorage);
       attachTBAATag(storeOp, boxTy, boxTy, nullptr);
-      rewriter.replaceOp(load, newBoxStorage.getResult());
+      rewriter.replaceOp(load, newBoxStorage);
     } else {
       auto loadOp = rewriter.create<mlir::LLVM::LoadOp>(
           load.getLoc(), llvmLoadTy, adaptor.getOperands(), load->getAttrs());
@@ -3808,6 +3867,11 @@ class FIRToLLVMLowering
     if (!forcedTargetTriple.empty())
       fir::setTargetTriple(mod, forcedTargetTriple);
 
+    if (!forcedDataLayout.empty()) {
+      llvm::DataLayout dl(forcedDataLayout);
+      fir::support::setMLIRDataLayout(mod, dl);
+    }
+
     // Run dynamic pass pipeline for converting Math dialect
     // operations into other dialects (llvm, func, etc.).
     // Some conversions of Math operations cannot be done