[CIR] Make use of !invariant.group metadata for const allocas

This patch updates the LLVM lowering part of load and stores to const allocas
and makes use of the !invariant.group metadata in the result LLVM IR.

The HoistAlloca pass is also updated. The const flag on a hoisted alloca is
removed for now since their uses are not always invariants. Will update in later
patches to teach their invariants.

clang/lib/CIR/Dialect/Transforms/HoistAllocas.cpp

@@ -48,8 +48,14 @@ static void process(cir::FuncOp func) {
 
   mlir::Operation *insertPoint = &*entryBlock.begin();
 
-  for (auto alloca : allocas)
+  for (auto alloca : allocas) {
     alloca->moveBefore(insertPoint);
+    if (alloca.getConstant()) {
+      // For now, we remove the const flag on nested allocas for constant
+      // variables when hoisting them.
+      alloca.setConstant(false);
+    }
+  }
 }
 
 void HoistAllocasPass::runOnOperation() {

clang/lib/CIR/Lowering/DirectToLLVM/LowerToLLVM.cpp

@@ -949,8 +949,7 @@ mlir::LogicalResult CIRToLLVMVTTAddrPointOpLowering::matchAndRewrite(
 
   if (op.getSymAddr()) {
     if (op.getOffset() == 0) {
-      rewriter.replaceAllUsesWith(op, llvmAddr);
-      rewriter.eraseOp(op);
+      rewriter.replaceOp(op, {llvmAddr});
       return mlir::success();
     }
 
@@ -1480,11 +1479,15 @@ mlir::LogicalResult CIRToLLVMLoadOpLowering::matchAndRewrite(
     alignment = *alignOpt;
   }
 
-  // TODO: nontemporal, invariant, syncscope.
+  auto addrAllocaOp =
+      mlir::dyn_cast_if_present<cir::AllocaOp>(op.getAddr().getDefiningOp());
+  auto invariant = addrAllocaOp && addrAllocaOp.getConstant();
+
+  // TODO: nontemporal, syncscope.
   rewriter.replaceOpWithNewOp<mlir::LLVM::LoadOp>(
       op, llvmTy, adaptor.getAddr(), /* alignment */ alignment,
       op.getIsVolatile(), /* nontemporal */ false,
-      /* invariant */ false, /* invariantGroup */ false, ordering);
+      /* invariant */ false, /* invariantGroup */ invariant, ordering);
   return mlir::LogicalResult::success();
 }
 
@@ -1505,10 +1508,14 @@ mlir::LogicalResult CIRToLLVMStoreOpLowering::matchAndRewrite(
     alignment = *alignOpt;
   }
 
+  auto addrAllocaOp =
+      mlir::dyn_cast_if_present<cir::AllocaOp>(op.getAddr().getDefiningOp());
+  auto invariant = addrAllocaOp && addrAllocaOp.getConstant();
+
   // TODO: nontemporal, syncscope.
   rewriter.replaceOpWithNewOp<mlir::LLVM::StoreOp>(
       op, adaptor.getValue(), adaptor.getAddr(), alignment, op.getIsVolatile(),
-      /* nontemporal */ false, /* invariantGroup */ false, ordering);
+      /* nontemporal */ false, /* invariantGroup */ invariant, ordering);
   return mlir::LogicalResult::success();
 }
 

clang/test/CIR/CodeGen/const-alloca.cpp

@@ -1,5 +1,7 @@
 // RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -fclangir -emit-cir %s -o %t.cir
-// RUN: FileCheck --input-file=%t.cir %s
+// RUN: FileCheck --input-file=%t.cir --check-prefix=CIR %s
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -fclangir -emit-llvm %s -o %t.ll
+// RUN: FileCheck --input-file=%t.ll --check-prefix=LLVM %s
 
 int produce_int();
 void blackbox(const int &);
@@ -8,33 +10,33 @@ void local_const_int() {
   const int x = produce_int();
 }
 
-// CHECK-LABEL: @_Z15local_const_intv
-// CHECK:   %{{.+}} = cir.alloca !s32i, !cir.ptr<!s32i>, ["x", init, const]
-// CHECK: }
+// CIR-LABEL: @_Z15local_const_intv
+// CIR:   %{{.+}} = cir.alloca !s32i, !cir.ptr<!s32i>, ["x", init, const]
+// CIR: }
 
 void param_const_int(const int x) {}
 
-// CHECK-LABEL: @_Z15param_const_inti
-// CHECK:  %{{.+}} = cir.alloca !s32i, !cir.ptr<!s32i>, ["x", init, const]
-// CHECK: }
+// CIR-LABEL: @_Z15param_const_inti
+// CIR:  %{{.+}} = cir.alloca !s32i, !cir.ptr<!s32i>, ["x", init, const]
+// CIR: }
 
 void local_constexpr_int() {
   constexpr int x = 42;
   blackbox(x);
 }
 
-// CHECK-LABEL: @_Z19local_constexpr_intv
-// CHECK:   %{{.+}} = cir.alloca !s32i, !cir.ptr<!s32i>, ["x", init, const]
-// CHECK: }
+// CIR-LABEL: @_Z19local_constexpr_intv
+// CIR:   %{{.+}} = cir.alloca !s32i, !cir.ptr<!s32i>, ["x", init, const]
+// CIR: }
 
 void local_reference() {
   int x = 0;
   int &r = x;
 }
 
-// CHECK-LABEL: @_Z15local_referencev
-// CHECK:   %{{.+}} = cir.alloca !cir.ptr<!s32i>, !cir.ptr<!cir.ptr<!s32i>>, ["r", init, const]
-// CHECK: }
+// CIR-LABEL: @_Z15local_referencev
+// CIR:   %{{.+}} = cir.alloca !cir.ptr<!s32i>, !cir.ptr<!cir.ptr<!s32i>>, ["r", init, const]
+// CIR: }
 
 struct Foo {
   int a;
@@ -47,6 +49,22 @@ void local_const_struct() {
   const Foo x = produce_foo();
 }
 
-// CHECK-LABEL: @_Z18local_const_structv
-// CHECK:   %{{.+}} = cir.alloca !ty_Foo, !cir.ptr<!ty_Foo>, ["x", init, const]
-// CHECK: }
+// CIR-LABEL: @_Z18local_const_structv
+// CIR:   %{{.+}} = cir.alloca !ty_Foo, !cir.ptr<!ty_Foo>, ["x", init, const]
+// CIR: }
+
+int local_const_load_store() {
+  const int x = produce_int();
+  int y = x;
+  return y;
+}
+
+// CIR-LABEL: @_Z22local_const_load_storev
+// CIR: %{{.+}} = cir.alloca !s32i, !cir.ptr<!s32i>, ["x", init, const] {alignment = 4 : i64}
+// CIR: }
+
+// LLVM-LABEL: @_Z22local_const_load_storev
+//      LLVM: %[[#INIT:]] = call i32 @_Z11produce_intv()
+// LLVM-NEXT: store i32 %[[#INIT]], ptr %[[#SLOT:]], align 4, !dbg !{{.+}}, !invariant.group !{{.+}}
+// LLVM-NEXT: %{{.+}} = load i32, ptr %[[#SLOT]], align 4, !dbg !{{.+}}, !invariant.group !{{.+}}
+// LLVM: }