[flang-commits] [flang] [flang][cuda] Add address cast for src and dst in TMA operations (PR #170375)

[via flang-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

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

Author: Valentin Clement (バレンタイン クレメン) (clementval)

<details>
<summary>Changes</summary>

src and dst pointer needs to have an address cast

---
Full diff: https://github.com/llvm/llvm-project/pull/170375.diff


2 Files Affected:

- (modified) flang/lib/Optimizer/Builder/CUDAIntrinsicCall.cpp (+4-2) 
- (modified) flang/test/Lower/CUDA/cuda-device-proc.cuf (+7-7) 


``````````diff
diff --git a/flang/lib/Optimizer/Builder/CUDAIntrinsicCall.cpp b/flang/lib/Optimizer/Builder/CUDAIntrinsicCall.cpp
index 270037f5fcb00..d8c22d16388e9 100644
--- a/flang/lib/Optimizer/Builder/CUDAIntrinsicCall.cpp
+++ b/flang/lib/Optimizer/Builder/CUDAIntrinsicCall.cpp
@@ -1496,8 +1496,10 @@ static void genTMABulkLoad(fir::FirOpBuilder &builder, mlir::Location loc,
   mlir::Value size = mlir::arith::MulIOp::create(builder, loc, nelem, eleSize);
   auto llvmPtrTy = mlir::LLVM::LLVMPointerType::get(builder.getContext());
   barrier = builder.createConvert(loc, llvmPtrTy, barrier);
-  dst = builder.createConvert(loc, llvmPtrTy, dst);
-  src = builder.createConvert(loc, llvmPtrTy, src);
+  dst = convertPtrToNVVMSpace(builder, loc, dst,
+                              mlir::NVVM::NVVMMemorySpace::Shared);
+  src = convertPtrToNVVMSpace(builder, loc, src,
+                              mlir::NVVM::NVVMMemorySpace::Shared);
   mlir::NVVM::InlinePtxOp::create(
       builder, loc, mlir::TypeRange{}, {dst, src, size, barrier}, {},
       "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], "
diff --git a/flang/test/Lower/CUDA/cuda-device-proc.cuf b/flang/test/Lower/CUDA/cuda-device-proc.cuf
index 434322ea22265..f8202d882d031 100644
--- a/flang/test/Lower/CUDA/cuda-device-proc.cuf
+++ b/flang/test/Lower/CUDA/cuda-device-proc.cuf
@@ -542,7 +542,7 @@ end subroutine
 ! CHECK: %[[ELEM_SIZE:.*]] = arith.constant 8 : i32
 ! CHECK: %[[SIZE:.*]] = arith.muli %[[COUNT]], %[[ELEM_SIZE]] : i32
 ! CHECK: %[[BARRIER_PTR:.*]] = fir.convert %[[BARRIER]]#0 : (!fir.ref<i64>) -> !llvm.ptr
-! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr, !llvm.ptr, i32, !llvm.ptr)
+! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr<3>, !llvm.ptr<3>, i32, !llvm.ptr)
 ! CHECK: nvvm.inline_ptx "mbarrier.expect_tx.relaxed.cta.shared::cta.b64 [%0], %1;" ro(%[[BARRIER_PTR]], %[[SIZE]] : !llvm.ptr, i32)
 
 attributes(global) subroutine test_tma_bulk_load_c8(a, n)
@@ -561,7 +561,7 @@ end subroutine
 ! CHECK: %[[ELEM_SIZE:.*]] = arith.constant 16 : i32
 ! CHECK: %[[SIZE:.*]] = arith.muli %[[COUNT]], %[[ELEM_SIZE]] : i32
 ! CHECK: %[[BARRIER_PTR:.*]] = fir.convert %[[BARRIER]]#0 : (!fir.ref<i64>) -> !llvm.ptr
-! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr, !llvm.ptr, i32, !llvm.ptr)
+! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr<3>, !llvm.ptr<3>, i32, !llvm.ptr)
 ! CHECK: nvvm.inline_ptx "mbarrier.expect_tx.relaxed.cta.shared::cta.b64 [%0], %1;" ro(%[[BARRIER_PTR]], %[[SIZE]] : !llvm.ptr, i32)
 
 attributes(global) subroutine test_tma_bulk_load_i4(a, n)
@@ -580,7 +580,7 @@ end subroutine
 ! CHECK: %[[ELEM_SIZE:.*]] = arith.constant 4 : i32
 ! CHECK: %[[SIZE:.*]] = arith.muli %[[COUNT]], %[[ELEM_SIZE]] : i32
 ! CHECK: %[[BARRIER_PTR:.*]] = fir.convert %[[BARRIER]]#0 : (!fir.ref<i64>) -> !llvm.ptr
-! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr, !llvm.ptr, i32, !llvm.ptr)
+! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr<3>, !llvm.ptr<3>, i32, !llvm.ptr)
 ! CHECK: nvvm.inline_ptx "mbarrier.expect_tx.relaxed.cta.shared::cta.b64 [%0], %1;" ro(%[[BARRIER_PTR]], %[[SIZE]] : !llvm.ptr, i32)
 
 attributes(global) subroutine test_tma_bulk_load_i8(a, n)
@@ -599,7 +599,7 @@ end subroutine
 ! CHECK: %[[ELEM_SIZE:.*]] = arith.constant 8 : i32
 ! CHECK: %[[SIZE:.*]] = arith.muli %[[COUNT]], %[[ELEM_SIZE]] : i32
 ! CHECK: %[[BARRIER_PTR:.*]] = fir.convert %[[BARRIER]]#0 : (!fir.ref<i64>) -> !llvm.ptr
-! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr, !llvm.ptr, i32, !llvm.ptr)
+! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr<3>, !llvm.ptr<3>, i32, !llvm.ptr)
 ! CHECK: nvvm.inline_ptx "mbarrier.expect_tx.relaxed.cta.shared::cta.b64 [%0], %1;" ro(%[[BARRIER_PTR]], %[[SIZE]] : !llvm.ptr, i32)
 
 attributes(global) subroutine test_tma_bulk_load_r2(a, n)
@@ -618,7 +618,7 @@ end subroutine
 ! CHECK: %[[ELEM_SIZE:.*]] = arith.constant 2 : i32
 ! CHECK: %[[SIZE:.*]] = arith.muli %[[COUNT]], %[[ELEM_SIZE]] : i32
 ! CHECK: %[[BARRIER_PTR:.*]] = fir.convert %[[BARRIER]]#0 : (!fir.ref<i64>) -> !llvm.ptr
-! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr, !llvm.ptr, i32, !llvm.ptr)
+! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr<3>, !llvm.ptr<3>, i32, !llvm.ptr)
 ! CHECK: nvvm.inline_ptx "mbarrier.expect_tx.relaxed.cta.shared::cta.b64 [%0], %1;" ro(%[[BARRIER_PTR]], %[[SIZE]] : !llvm.ptr, i32)
 
 attributes(global) subroutine test_tma_bulk_load_r4(a, n)
@@ -637,7 +637,7 @@ end subroutine
 ! CHECK: %[[ELEM_SIZE:.*]] = arith.constant 4 : i32
 ! CHECK: %[[SIZE:.*]] = arith.muli %[[COUNT]], %[[ELEM_SIZE]] : i32
 ! CHECK: %[[BARRIER_PTR:.*]] = fir.convert %[[BARRIER]]#0 : (!fir.ref<i64>) -> !llvm.ptr
-! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr, !llvm.ptr, i32, !llvm.ptr)
+! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr<3>, !llvm.ptr<3>, i32, !llvm.ptr)
 ! CHECK: nvvm.inline_ptx "mbarrier.expect_tx.relaxed.cta.shared::cta.b64 [%0], %1;" ro(%[[BARRIER_PTR]], %[[SIZE]] : !llvm.ptr, i32)
 
 attributes(global) subroutine test_tma_bulk_load_r8(a, n)
@@ -656,7 +656,7 @@ end subroutine
 ! CHECK: %[[ELEM_SIZE:.*]] = arith.constant 8 : i32
 ! CHECK: %[[SIZE:.*]] = arith.muli %[[COUNT]], %[[ELEM_SIZE]] : i32
 ! CHECK: %[[BARRIER_PTR:.*]] = fir.convert %[[BARRIER]]#0 : (!fir.ref<i64>) -> !llvm.ptr
-! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr, !llvm.ptr, i32, !llvm.ptr)
+! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr<3>, !llvm.ptr<3>, i32, !llvm.ptr)
 ! CHECK: nvvm.inline_ptx "mbarrier.expect_tx.relaxed.cta.shared::cta.b64 [%0], %1;" ro(%[[BARRIER_PTR]], %[[SIZE]] : !llvm.ptr, i32)
 
 attributes(global) subroutine test_tma_bulk_store_c4(c, n)

``````````

</details>


https://github.com/llvm/llvm-project/pull/170375