[llvm-branch-commits] [clang] clang/AMDGPU: Emit atomicrmw for __builtin_amdgcn_global_atomic_fadd_{f32|f64} (PR #96872)

via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Thu Jun 27 02:12:54 PDT 2024


llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-llvm-transforms

Author: Matt Arsenault (arsenm)

<details>
<summary>Changes</summary>

Need to emit syncscope and new metadata to get the native instruction,
most of the time.

---
Full diff: https://github.com/llvm/llvm-project/pull/96872.diff


5 Files Affected:

- (modified) clang/lib/CodeGen/CGBuiltin.cpp (+27-12) 
- (modified) clang/test/CodeGenOpenCL/builtins-amdgcn-gfx11.cl (+1-1) 
- (modified) clang/test/CodeGenOpenCL/builtins-fp-atomics-gfx12.cl (+2-2) 
- (modified) clang/test/CodeGenOpenCL/builtins-fp-atomics-gfx90a.cl (+2-2) 
- (modified) clang/test/CodeGenOpenCL/builtins-fp-atomics-gfx940.cl (+2-2) 


``````````diff
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index 96dcf6283f9f8..d90762748d925 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -58,6 +58,7 @@
 #include "llvm/IR/MDBuilder.h"
 #include "llvm/IR/MatrixBuilder.h"
 #include "llvm/IR/MemoryModelRelaxationAnnotations.h"
+#include "llvm/Support/AMDGPUAddrSpace.h"
 #include "llvm/Support/ConvertUTF.h"
 #include "llvm/Support/MathExtras.h"
 #include "llvm/Support/ScopedPrinter.h"
@@ -18654,8 +18655,6 @@ Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID,
     Src0 = Builder.CreatePointerBitCastOrAddrSpaceCast(Src0, PTy);
     return Builder.CreateCall(F, { Src0, Src1, Src2, Src3, Src4 });
   }
-  case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_f64:
-  case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_f32:
   case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_v2f16:
   case AMDGPU::BI__builtin_amdgcn_global_atomic_fmin_f64:
   case AMDGPU::BI__builtin_amdgcn_global_atomic_fmax_f64:
@@ -18667,18 +18666,11 @@ Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID,
     Intrinsic::ID IID;
     llvm::Type *ArgTy = llvm::Type::getDoubleTy(getLLVMContext());
     switch (BuiltinID) {
-    case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_f32:
-      ArgTy = llvm::Type::getFloatTy(getLLVMContext());
-      IID = Intrinsic::amdgcn_global_atomic_fadd;
-      break;
     case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_v2f16:
       ArgTy = llvm::FixedVectorType::get(
           llvm::Type::getHalfTy(getLLVMContext()), 2);
       IID = Intrinsic::amdgcn_global_atomic_fadd;
       break;
-    case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_f64:
-      IID = Intrinsic::amdgcn_global_atomic_fadd;
-      break;
     case AMDGPU::BI__builtin_amdgcn_global_atomic_fmin_f64:
       IID = Intrinsic::amdgcn_global_atomic_fmin;
       break;
@@ -19091,7 +19083,9 @@ Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID,
   case AMDGPU::BI__builtin_amdgcn_ds_atomic_fadd_f64:
   case AMDGPU::BI__builtin_amdgcn_ds_atomic_fadd_f32:
   case AMDGPU::BI__builtin_amdgcn_ds_atomic_fadd_v2f16:
-  case AMDGPU::BI__builtin_amdgcn_ds_atomic_fadd_v2bf16: {
+  case AMDGPU::BI__builtin_amdgcn_ds_atomic_fadd_v2bf16:
+  case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_f32:
+  case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_f64: {
     llvm::AtomicRMWInst::BinOp BinOp;
     switch (BuiltinID) {
     case AMDGPU::BI__builtin_amdgcn_atomic_inc32:
@@ -19107,6 +19101,8 @@ Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID,
     case AMDGPU::BI__builtin_amdgcn_ds_atomic_fadd_f32:
     case AMDGPU::BI__builtin_amdgcn_ds_atomic_fadd_v2f16:
     case AMDGPU::BI__builtin_amdgcn_ds_atomic_fadd_v2bf16:
+    case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_f32:
+    case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_f64:
       BinOp = llvm::AtomicRMWInst::FAdd;
       break;
     }
@@ -19133,8 +19129,13 @@ Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID,
       ProcessOrderScopeAMDGCN(EmitScalarExpr(E->getArg(2)),
                               EmitScalarExpr(E->getArg(3)), AO, SSID);
     } else {
-      // The ds_atomic_fadd_* builtins do not have syncscope/order arguments.
-      SSID = llvm::SyncScope::System;
+      // Most of the builtins do not have syncscope/order arguments. For DS
+      // atomics the scope doesn't really matter, as they implicitly operate at
+      // workgroup scope.
+      //
+      // The global/flat cases need to use agent scope to consistently produce
+      // the native instruction instead of a cmpxchg expansion.
+      SSID = getLLVMContext().getOrInsertSyncScopeID("agent");
       AO = AtomicOrdering::SequentiallyConsistent;
 
       // The v2bf16 builtin uses i16 instead of a natural bfloat type.
@@ -19149,6 +19150,20 @@ Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID,
         Builder.CreateAtomicRMW(BinOp, Ptr, Val, AO, SSID);
     if (Volatile)
       RMW->setVolatile(true);
+
+    unsigned AddrSpace = Ptr.getType()->getAddressSpace();
+    if (AddrSpace != llvm::AMDGPUAS::LOCAL_ADDRESS) {
+      // Most targets require "amdgpu.no.fine.grained.memory" to emit the native
+      // instruction for flat and global operations.
+      llvm::MDTuple *EmptyMD = MDNode::get(getLLVMContext(), {});
+      RMW->setMetadata("amdgpu.no.fine.grained.memory", EmptyMD);
+
+      // Most targets require "amdgpu.ignore.denormal.mode" to emit the native
+      // instruction, but this only matters for float fadd.
+      if (BinOp == llvm::AtomicRMWInst::FAdd && Val->getType()->isFloatTy())
+        RMW->setMetadata("amdgpu.ignore.denormal.mode", EmptyMD);
+    }
+
     return Builder.CreateBitCast(RMW, OrigTy);
   }
   case AMDGPU::BI__builtin_amdgcn_s_sendmsg_rtn:
diff --git a/clang/test/CodeGenOpenCL/builtins-amdgcn-gfx11.cl b/clang/test/CodeGenOpenCL/builtins-amdgcn-gfx11.cl
index f444657463568..e8889f57432f5 100644
--- a/clang/test/CodeGenOpenCL/builtins-amdgcn-gfx11.cl
+++ b/clang/test/CodeGenOpenCL/builtins-amdgcn-gfx11.cl
@@ -49,7 +49,7 @@ void test_s_wait_event_export_ready() {
 }
 
 // CHECK-LABEL: @test_global_add_f32
-// CHECK: {{.*}}call{{.*}} float @llvm.amdgcn.global.atomic.fadd.f32.p1.f32(ptr addrspace(1) %{{.*}}, float %{{.*}})
+// CHECK: = atomicrmw fadd ptr addrspace(1) %addr, float %x syncscope("agent") seq_cst, align 4, !amdgpu.no.fine.grained.memory !{{[0-9]+}}, !amdgpu.ignore.denormal.mode !{{[0-9]+$}}
 void test_global_add_f32(float *rtn, global float *addr, float x) {
   *rtn = __builtin_amdgcn_global_atomic_fadd_f32(addr, x);
 }
diff --git a/clang/test/CodeGenOpenCL/builtins-fp-atomics-gfx12.cl b/clang/test/CodeGenOpenCL/builtins-fp-atomics-gfx12.cl
index 63381942eaba5..21c1c38bc78dc 100644
--- a/clang/test/CodeGenOpenCL/builtins-fp-atomics-gfx12.cl
+++ b/clang/test/CodeGenOpenCL/builtins-fp-atomics-gfx12.cl
@@ -11,7 +11,7 @@ typedef short __attribute__((ext_vector_type(2))) short2;
 
 // CHECK-LABEL: test_local_add_2bf16
 // CHECK: [[BC0:%.+]] = bitcast <2 x i16> {{.+}} to <2 x bfloat>
-// CHECK: [[RMW:%.+]] = atomicrmw fadd ptr addrspace(3) %{{.+}}, <2 x bfloat> [[BC0]] seq_cst, align 4
+// CHECK: [[RMW:%.+]] = atomicrmw fadd ptr addrspace(3) %{{.+}}, <2 x bfloat> [[BC0]] syncscope("agent") seq_cst, align 4
 // CHECK-NEXT: bitcast <2 x bfloat> [[RMW]] to <2 x i16>
 
 // GFX12-LABEL:  test_local_add_2bf16
@@ -22,7 +22,7 @@ short2 test_local_add_2bf16(__local short2 *addr, short2 x) {
 
 // CHECK-LABEL: test_local_add_2bf16_noret
 // CHECK: [[BC0:%.+]] = bitcast <2 x i16> {{.+}} to <2 x bfloat>
-// CHECK: [[RMW:%.+]] = atomicrmw fadd ptr addrspace(3) %{{.+}}, <2 x bfloat> [[BC0]] seq_cst, align 4
+// CHECK: [[RMW:%.+]] = atomicrmw fadd ptr addrspace(3) %{{.+}}, <2 x bfloat> [[BC0]] syncscope("agent") seq_cst, align 4
 // CHECK-NEXT: bitcast <2 x bfloat> [[RMW]] to <2 x i16>
 
 // GFX12-LABEL:  test_local_add_2bf16_noret
diff --git a/clang/test/CodeGenOpenCL/builtins-fp-atomics-gfx90a.cl b/clang/test/CodeGenOpenCL/builtins-fp-atomics-gfx90a.cl
index e2117f11858f7..3778f65feaad4 100644
--- a/clang/test/CodeGenOpenCL/builtins-fp-atomics-gfx90a.cl
+++ b/clang/test/CodeGenOpenCL/builtins-fp-atomics-gfx90a.cl
@@ -9,7 +9,7 @@
 typedef half __attribute__((ext_vector_type(2))) half2;
 
 // CHECK-LABEL: test_global_add_f64
-// CHECK: call double @llvm.amdgcn.global.atomic.fadd.f64.p1.f64(ptr addrspace(1) %{{.*}}, double %{{.*}})
+// CHECK: = atomicrmw fadd ptr addrspace(1) %{{.+}}, double %{{.+}} syncscope("agent") seq_cst, align 8, !amdgpu.no.fine.grained.memory !{{[0-9]+$}}
 // GFX90A-LABEL:  test_global_add_f64$local:
 // GFX90A:  global_atomic_add_f64
 void test_global_add_f64(__global double *addr, double x) {
@@ -117,7 +117,7 @@ void test_ds_addf_local_f32(__local float *addr, float x){
 }
 
 // CHECK-LABEL: @test_global_add_f32
-// CHECK: call float @llvm.amdgcn.global.atomic.fadd.f32.p1.f32(ptr addrspace(1) %{{.*}}, float %{{.*}})
+// CHECK: = atomicrmw fadd ptr addrspace(1) %{{.+}}, float %{{.+}} syncscope("agent") seq_cst, align 4, !amdgpu.no.fine.grained.memory !{{[0-9]+}}, !amdgpu.ignore.denormal.mode !{{[0-9]+$}}
 void test_global_add_f32(float *rtn, global float *addr, float x) {
   *rtn = __builtin_amdgcn_global_atomic_fadd_f32(addr, x);
 }
diff --git a/clang/test/CodeGenOpenCL/builtins-fp-atomics-gfx940.cl b/clang/test/CodeGenOpenCL/builtins-fp-atomics-gfx940.cl
index 92a33ceac2290..d4ca32b9b3cd6 100644
--- a/clang/test/CodeGenOpenCL/builtins-fp-atomics-gfx940.cl
+++ b/clang/test/CodeGenOpenCL/builtins-fp-atomics-gfx940.cl
@@ -44,7 +44,7 @@ short2 test_global_add_2bf16(__global short2 *addr, short2 x) {
 // CHECK-LABEL: test_local_add_2bf16
 
 // CHECK: [[BC0:%.+]] = bitcast <2 x i16> {{.+}} to <2 x bfloat>
-// CHECK: [[RMW:%.+]] = atomicrmw fadd ptr addrspace(3) %{{.+}}, <2 x bfloat> [[BC0]] seq_cst, align 4
+// CHECK: [[RMW:%.+]] = atomicrmw fadd ptr addrspace(3) %{{.+}}, <2 x bfloat> [[BC0]] syncscope("agent") seq_cst, align 4{{$}}
 // CHECK-NEXT: bitcast <2 x bfloat> [[RMW]] to <2 x i16>
 
 // GFX940-LABEL:  test_local_add_2bf16
@@ -70,7 +70,7 @@ void test_local_add_2f16_noret(__local half2 *addr, half2 x) {
 }
 
 // CHECK-LABEL: @test_global_add_f32
-// CHECK: call float @llvm.amdgcn.global.atomic.fadd.f32.p1.f32(ptr addrspace(1) %{{.*}}, float %{{.*}})
+// CHECK: = atomicrmw fadd ptr addrspace(1) %{{.+}}, float %{{.+}} syncscope("agent") seq_cst, align 4, !amdgpu.no.fine.grained.memory !{{[0-9]+}}, !amdgpu.ignore.denormal.mode !{{[0-9]+$}}
 void test_global_add_f32(float *rtn, global float *addr, float x) {
   *rtn = __builtin_amdgcn_global_atomic_fadd_f32(addr, x);
 }

``````````

</details>


https://github.com/llvm/llvm-project/pull/96872


More information about the llvm-branch-commits mailing list