[Mlir-commits] [mlir] [MLIR] Add optimization attrs for gpu-to-llvmspv function declarations (PR #99301)

llvmlistbot at llvm.org llvmlistbot at llvm.org
Wed Jul 17 03:31:05 PDT 2024


https://github.com/FMarno updated https://github.com/llvm/llvm-project/pull/99301

>From 4c5887ecaf038d3ff1ff9588725b7b7743b528d3 Mon Sep 17 00:00:00 2001
From: Finlay Marno <finlay.marno at codeplay.com>
Date: Tue, 16 Jul 2024 11:15:04 +0100
Subject: [PATCH] set optimization attrs for gpu-to-llvmspv ops

Adds the attributes include nounwind, willreturn and memory(none).
---
 .../Conversion/GPUToLLVMSPV/GPUToLLVMSPV.cpp  |  35 ++++--
 .../GPUToLLVMSPV/gpu-to-llvm-spv.mlir         | 104 ++++++++++++++----
 2 files changed, 108 insertions(+), 31 deletions(-)

diff --git a/mlir/lib/Conversion/GPUToLLVMSPV/GPUToLLVMSPV.cpp b/mlir/lib/Conversion/GPUToLLVMSPV/GPUToLLVMSPV.cpp
index ebeb8f803d71d..81682a52c8c4b 100644
--- a/mlir/lib/Conversion/GPUToLLVMSPV/GPUToLLVMSPV.cpp
+++ b/mlir/lib/Conversion/GPUToLLVMSPV/GPUToLLVMSPV.cpp
@@ -40,11 +40,10 @@ namespace mlir {
 // Helper Functions
 //===----------------------------------------------------------------------===//
 
-static LLVM::LLVMFuncOp lookupOrCreateSPIRVFn(Operation *symbolTable,
-                                              StringRef name,
-                                              ArrayRef<Type> paramTypes,
-                                              Type resultType,
-                                              bool isConvergent = false) {
+static LLVM::LLVMFuncOp
+lookupOrCreateSPIRVFn(Operation *symbolTable, StringRef name,
+                      ArrayRef<Type> paramTypes, Type resultType,
+                      bool hasMemoryEffects = true, bool isConvergent = false) {
   auto func = dyn_cast_or_null<LLVM::LLVMFuncOp>(
       SymbolTable::lookupSymbolIn(symbolTable, name));
   if (!func) {
@@ -53,6 +52,17 @@ static LLVM::LLVMFuncOp lookupOrCreateSPIRVFn(Operation *symbolTable,
         symbolTable->getLoc(), name,
         LLVM::LLVMFunctionType::get(resultType, paramTypes));
     func.setCConv(LLVM::cconv::CConv::SPIR_FUNC);
+    func.setNoUnwind(true);
+    func.setWillReturn(true);
+    if (!hasMemoryEffects) {
+      // no externally observable effects
+      constexpr auto noModRef = mlir::LLVM::ModRefInfo::NoModRef;
+      auto memAttr = b.getAttr<LLVM::MemoryEffectsAttr>(
+          /*other*/ noModRef,
+          /*argMem*/ noModRef, /*inaccessibleMem*/ noModRef);
+      func.setMemoryAttr(memAttr);
+    }
+
     func.setConvergent(isConvergent);
   }
   return func;
@@ -91,8 +101,9 @@ struct GPUBarrierConversion final : ConvertOpToLLVMPattern<gpu::BarrierOp> {
     assert(moduleOp && "Expecting module");
     Type flagTy = rewriter.getI32Type();
     Type voidTy = rewriter.getType<LLVM::LLVMVoidType>();
-    LLVM::LLVMFuncOp func = lookupOrCreateSPIRVFn(
-        moduleOp, funcName, flagTy, voidTy, /*isConvergent=*/true);
+    LLVM::LLVMFuncOp func =
+        lookupOrCreateSPIRVFn(moduleOp, funcName, flagTy, voidTy,
+                              /*hasMemoryEffects*/ true, /*isConvergent=*/true);
 
     // Value used by SPIR-V backend to represent `CLK_LOCAL_MEM_FENCE`.
     // See `llvm/lib/Target/SPIRV/SPIRVBuiltins.td`.
@@ -134,8 +145,8 @@ struct LaunchConfigConversion : ConvertToLLVMPattern {
     assert(moduleOp && "Expecting module");
     Type dimTy = rewriter.getI32Type();
     Type indexTy = getTypeConverter()->getIndexType();
-    LLVM::LLVMFuncOp func =
-        lookupOrCreateSPIRVFn(moduleOp, funcName, dimTy, indexTy);
+    LLVM::LLVMFuncOp func = lookupOrCreateSPIRVFn(
+        moduleOp, funcName, dimTy, indexTy, /*hasMemoryEffects*/ false);
 
     Location loc = op->getLoc();
     gpu::Dimension dim = getDimension(op);
@@ -268,9 +279,9 @@ struct GPUShuffleConversion final : ConvertOpToLLVMPattern<gpu::ShuffleOp> {
     Type valueType = adaptor.getValue().getType();
     Type offsetType = adaptor.getOffset().getType();
     Type resultType = valueType;
-    LLVM::LLVMFuncOp func =
-        lookupOrCreateSPIRVFn(moduleOp, funcName, {valueType, offsetType},
-                              resultType, /*isConvergent=*/true);
+    LLVM::LLVMFuncOp func = lookupOrCreateSPIRVFn(
+        moduleOp, funcName, {valueType, offsetType}, resultType,
+        /*hasMemoryEffects*/ true, /*isConvergent=*/true);
 
     Location loc = op->getLoc();
     std::array<Value, 2> args{adaptor.getValue(), adaptor.getOffset()};
diff --git a/mlir/test/Conversion/GPUToLLVMSPV/gpu-to-llvm-spv.mlir b/mlir/test/Conversion/GPUToLLVMSPV/gpu-to-llvm-spv.mlir
index 1b0f89a9a573e..d094c786414fc 100644
--- a/mlir/test/Conversion/GPUToLLVMSPV/gpu-to-llvm-spv.mlir
+++ b/mlir/test/Conversion/GPUToLLVMSPV/gpu-to-llvm-spv.mlir
@@ -4,16 +4,46 @@
 // RUN: | FileCheck --check-prefixes=CHECK-32,CHECK %s
 
 gpu.module @builtins {
-  // CHECK-64:    llvm.func spir_funccc @_Z14get_num_groupsj(i32) -> i64
-  // CHECK-64:    llvm.func spir_funccc @_Z12get_local_idj(i32) -> i64
-  // CHECK-64:    llvm.func spir_funccc @_Z14get_local_sizej(i32) -> i64
-  // CHECK-64:    llvm.func spir_funccc @_Z13get_global_idj(i32) -> i64
-  // CHECK-64:    llvm.func spir_funccc @_Z12get_group_idj(i32) -> i64
-  // CHECK-32:    llvm.func spir_funccc @_Z14get_num_groupsj(i32) -> i32
-  // CHECK-32:    llvm.func spir_funccc @_Z12get_local_idj(i32) -> i32
-  // CHECK-32:    llvm.func spir_funccc @_Z14get_local_sizej(i32) -> i32
-  // CHECK-32:    llvm.func spir_funccc @_Z13get_global_idj(i32) -> i32
-  // CHECK-32:    llvm.func spir_funccc @_Z12get_group_idj(i32) -> i32
+  // CHECK-64:        llvm.func spir_funccc @_Z14get_num_groupsj(i32) -> i64 attributes {
+  // CHECK-SAME-DAG:  memory = #llvm.memory_effects<other = none, argMem = none, inaccessibleMem = none>
+  // CHECK-SAME-DAG:  no_unwind
+  // CHECK-SAME-DAG:  will_return
+  // CHECK-64:        llvm.func spir_funccc @_Z12get_local_idj(i32) -> i64 attributes {
+  // CHECK-SAME-DAG:  memory = #llvm.memory_effects<other = none, argMem = none, inaccessibleMem = none>
+  // CHECK-SAME-DAG:  no_unwind
+  // CHECK-SAME-DAG:  will_return
+  // CHECK-64:        llvm.func spir_funccc @_Z14get_local_sizej(i32) -> i64 attributes {
+  // CHECK-SAME-DAG:  memory = #llvm.memory_effects<other = none, argMem = none, inaccessibleMem = none>
+  // CHECK-SAME-DAG:  no_unwind
+  // CHECK-SAME-DAG:  will_return
+  // CHECK-64:        llvm.func spir_funccc @_Z13get_global_idj(i32) -> i64 attributes {
+  // CHECK-SAME-DAG:  memory = #llvm.memory_effects<other = none, argMem = none, inaccessibleMem = none>
+  // CHECK-SAME-DAG:  no_unwind
+  // CHECK-SAME-DAG:  will_return
+  // CHECK-64:        llvm.func spir_funccc @_Z12get_group_idj(i32) -> i64 attributes {
+  // CHECK-SAME-DAG:  memory = #llvm.memory_effects<other = none, argMem = none, inaccessibleMem = none>
+  // CHECK-SAME-DAG:  no_unwind
+  // CHECK-SAME-DAG:  will_return
+  // CHECK-32:        llvm.func spir_funccc @_Z14get_num_groupsj(i32) -> i32 attributes {
+  // CHECK-SAME-DAG:  memory = #llvm.memory_effects<other = none, argMem = none, inaccessibleMem = none>
+  // CHECK-SAME-DAG:  no_unwind
+  // CHECK-SAME-DAG:  will_return
+  // CHECK-32:        llvm.func spir_funccc @_Z12get_local_idj(i32) -> i32 attributes {
+  // CHECK-SAME-DAG:  memory = #llvm.memory_effects<other = none, argMem = none, inaccessibleMem = none>
+  // CHECK-SAME-DAG:  no_unwind
+  // CHECK-SAME-DAG:  will_return
+  // CHECK-32:        llvm.func spir_funccc @_Z14get_local_sizej(i32) -> i32 attributes {
+  // CHECK-SAME-DAG:  memory = #llvm.memory_effects<other = none, argMem = none, inaccessibleMem = none>
+  // CHECK-SAME-DAG:  no_unwind
+  // CHECK-SAME-DAG:  will_return
+  // CHECK-32:        llvm.func spir_funccc @_Z13get_global_idj(i32) -> i32 attributes {
+  // CHECK-SAME-DAG:  memory = #llvm.memory_effects<other = none, argMem = none, inaccessibleMem = none>
+  // CHECK-SAME-DAG:  no_unwind
+  // CHECK-SAME-DAG:  will_return
+  // CHECK-32:        llvm.func spir_funccc @_Z12get_group_idj(i32) -> i32 attributes {
+  // CHECK-SAME-DAG:  memory = #llvm.memory_effects<other = none, argMem = none, inaccessibleMem = none>
+  // CHECK-SAME-DAG:  no_unwind
+  // CHECK-SAME-DAG:  will_return
 
   // CHECK-LABEL: gpu_block_id
   func.func @gpu_block_id() -> (index, index, index) {
@@ -104,7 +134,11 @@ gpu.module @builtins {
 // -----
 
 gpu.module @barriers {
-  // CHECK:       llvm.func spir_funccc @_Z7barrierj(i32) attributes {convergent}
+  // CHECK:           llvm.func spir_funccc @_Z7barrierj(i32) attributes {
+  // CHECK-SAME-DAG:  no_unwind
+  // CHECK-SAME-DAG:  convergent
+  // CHECK-NOT:       memory = #llvm.memory_effects
+  // CHECK-SAME:      }
 
   // CHECK-LABEL: gpu_barrier
   func.func @gpu_barrier() {
@@ -120,10 +154,26 @@ gpu.module @barriers {
 // Check `gpu.shuffle` conversion with default subgroup size.
 
 gpu.module @shuffles {
-  // CHECK:       llvm.func spir_funccc @_Z22sub_group_shuffle_downdj(f64, i32) -> f64 attributes {convergent}
-  // CHECK:       llvm.func spir_funccc @_Z20sub_group_shuffle_upfj(f32, i32) -> f32 attributes {convergent}
-  // CHECK:       llvm.func spir_funccc @_Z21sub_group_shuffle_xorlj(i64, i32) -> i64 attributes {convergent}
-  // CHECK:       llvm.func spir_funccc @_Z17sub_group_shuffleij(i32, i32) -> i32 attributes {convergent}
+  // CHECK:           llvm.func spir_funccc @_Z22sub_group_shuffle_downdj(f64, i32) -> f64 attributes {
+  // CHECK-SAME-DAG:  no_unwind
+  // CHECK-SAME-DAG:  convergent
+  // CHECK-NOT:       memory = #llvm.memory_effects
+  // CHECK-SAME:      }
+  // CHECK:           llvm.func spir_funccc @_Z20sub_group_shuffle_upfj(f32, i32) -> f32 attributes {
+  // CHECK-SAME-DAG:  no_unwind
+  // CHECK-SAME-DAG:  convergent
+  // CHECK-NOT:       memory = #llvm.memory_effects
+  // CHECK-SAME:      }
+  // CHECK:           llvm.func spir_funccc @_Z21sub_group_shuffle_xorlj(i64, i32) -> i64 attributes {
+  // CHECK-SAME-DAG:  no_unwind
+  // CHECK-SAME-DAG:  convergent
+  // CHECK-NOT:       memory = #llvm.memory_effects
+  // CHECK-SAME:      }
+  // CHECK:           llvm.func spir_funccc @_Z17sub_group_shuffleij(i32, i32) -> i32 attributes {
+  // CHECK-SAME-DAG:  no_unwind
+  // CHECK-SAME-DAG:  convergent
+  // CHECK-NOT:       memory = #llvm.memory_effects
+  // CHECK-SAME:      }
 
   // CHECK-LABEL: gpu_shuffles
   // CHECK-SAME:              (%[[VAL_0:.*]]: i32, %[[VAL_1:.*]]: i32, %[[VAL_2:.*]]: i64, %[[VAL_3:.*]]: i32, %[[VAL_4:.*]]: f32, %[[VAL_5:.*]]: i32, %[[VAL_6:.*]]: f64, %[[VAL_7:.*]]: i32)
@@ -155,10 +205,26 @@ gpu.module @shuffles {
 gpu.module @shuffles attributes {
   spirv.target_env = #spirv.target_env<#spirv.vce<v1.4, [Kernel, Addresses, GroupNonUniformShuffle, Int64], []>, #spirv.resource_limits<subgroup_size = 16>>
 } {
-  // CHECK:       llvm.func spir_funccc @_Z22sub_group_shuffle_downdj(f64, i32) -> f64 attributes {convergent}
-  // CHECK:       llvm.func spir_funccc @_Z20sub_group_shuffle_upfj(f32, i32) -> f32 attributes {convergent}
-  // CHECK:       llvm.func spir_funccc @_Z21sub_group_shuffle_xorlj(i64, i32) -> i64 attributes {convergent}
-  // CHECK:       llvm.func spir_funccc @_Z17sub_group_shuffleij(i32, i32) -> i32 attributes {convergent}
+  // CHECK:           llvm.func spir_funccc @_Z22sub_group_shuffle_downdj(f64, i32) -> f64 attributes {
+  // CHECK-SAME-DAG:  no_unwind
+  // CHECK-SAME-DAG:  convergent
+  // CHECK-NOT:       memory = #llvm.memory_effects
+  // CHECK-SAME:      }
+  // CHECK:           llvm.func spir_funccc @_Z20sub_group_shuffle_upfj(f32, i32) -> f32 attributes {
+  // CHECK-SAME-DAG:  no_unwind
+  // CHECK-SAME-DAG:  convergent
+  // CHECK-NOT:       memory = #llvm.memory_effects
+  // CHECK-SAME:      }
+  // CHECK:           llvm.func spir_funccc @_Z21sub_group_shuffle_xorlj(i64, i32) -> i64 attributes {
+  // CHECK-SAME-DAG:  no_unwind
+  // CHECK-SAME-DAG:  convergent
+  // CHECK-NOT:       memory = #llvm.memory_effects
+  // CHECK-SAME:      }
+  // CHECK:           llvm.func spir_funccc @_Z17sub_group_shuffleij(i32, i32) -> i32 attributes {
+  // CHECK-SAME-DAG:  no_unwind
+  // CHECK-SAME-DAG:  convergent
+  // CHECK-NOT:       memory = #llvm.memory_effects
+  // CHECK-SAME:      }
 
   // CHECK-LABEL: gpu_shuffles
   // CHECK-SAME:              (%[[VAL_0:.*]]: i32, %[[VAL_1:.*]]: i32, %[[VAL_2:.*]]: i64, %[[VAL_3:.*]]: i32, %[[VAL_4:.*]]: f32, %[[VAL_5:.*]]: i32, %[[VAL_6:.*]]: f64, %[[VAL_7:.*]]: i32)



More information about the Mlir-commits mailing list