[clang] [clang][SPIR-V] Treat AMDGCN builtin pointers in AS 0 as generic (PR #234504)
Arseniy Obolenskiy via cfe-commits
cfe-commits at lists.llvm.org
Thu Oct 22 02:27:55 PDT 2026
https://github.com/aobolensk updated https://github.com/llvm/llvm-project/pull/234504
>From cff662c8fa9bffb0d0ac559be8c8d9b595d23ee0 Mon Sep 17 00:00:00 2001
From: Arseniy Obolenskiy <arseniy.obolenskiy at amd.com>
Date: Thu, 22 Oct 2026 08:11:26 +0200
Subject: [PATCH] [clang][SPIR-V] Treat AMDGCN builtin pointers in AS 0 as
generic
In AMDGCN builtins AS 0 means flat, but SPIR-V reads it as Function
This broke flat atomics such as __builtin_amdgcn_flat_atomic_fadd_v2f16
---
clang/lib/Basic/Targets/SPIR.h | 9 ++
.../CodeGenHIP/amdgpu-flat-atomic-fadd.hip | 132 +++++++++++++++---
2 files changed, 121 insertions(+), 20 deletions(-)
diff --git a/clang/lib/Basic/Targets/SPIR.h b/clang/lib/Basic/Targets/SPIR.h
index c24a55ecfc9704..43a94cb9d31c89 100644
--- a/clang/lib/Basic/Targets/SPIR.h
+++ b/clang/lib/Basic/Targets/SPIR.h
@@ -507,6 +507,15 @@ class LLVM_LIBRARY_VISIBILITY SPIRV64AMDGCNTargetInfo final
bool hasInt128Type() const override { return TargetInfo::hasInt128Type(); }
+ // Builtin address spaces use AMDGPU numbering, where 0 is flat.
+ LangAS getOpenCLBuiltinAddressSpace(unsigned AS) const override {
+ return AS == 0 ? LangAS::opencl_generic : getLangASFromTargetAS(AS);
+ }
+
+ LangAS getCUDABuiltinAddressSpace(unsigned AS) const override {
+ return AS == 0 ? LangAS::Default : getLangASFromTargetAS(AS);
+ }
+
// This is only needed for validating arguments passed to
// __builtin_amdgcn_processor_is
bool isValidCPUName(StringRef Name) const override;
diff --git a/clang/test/CodeGenHIP/amdgpu-flat-atomic-fadd.hip b/clang/test/CodeGenHIP/amdgpu-flat-atomic-fadd.hip
index 41da90d9da4b56..c7f831d2ec98ed 100644
--- a/clang/test/CodeGenHIP/amdgpu-flat-atomic-fadd.hip
+++ b/clang/test/CodeGenHIP/amdgpu-flat-atomic-fadd.hip
@@ -44,10 +44,9 @@ __device__ double global_double;
// SPIRV-NEXT: br i1 [[TMP0]], label %[[IF_THEN:.*]], label %[[IF_END:.*]]
// SPIRV: [[IF_THEN]]:
// SPIRV-NEXT: [[TMP1:%.*]] = load ptr addrspace(4), ptr addrspace(4) [[PTR_ADDR_ASCAST]], align 8
-// SPIRV-NEXT: [[TMP2:%.*]] = addrspacecast ptr addrspace(4) [[TMP1]] to ptr
-// SPIRV-NEXT: [[TMP3:%.*]] = load float, ptr addrspace(4) [[VAL_ADDR_ASCAST]], align 4
-// SPIRV-NEXT: [[TMP4:%.*]] = atomicrmw fadd ptr [[TMP2]], float [[TMP3]] syncscope("device") monotonic, align 4, !atomic.ignore.denormal.mode [[META5:![0-9]+]], !amdgpu.no.fine.grained.memory [[META5]]
-// SPIRV-NEXT: store float [[TMP4]], ptr addrspace(4) [[RESULT_ASCAST]], align 4
+// SPIRV-NEXT: [[TMP2:%.*]] = load float, ptr addrspace(4) [[VAL_ADDR_ASCAST]], align 4
+// SPIRV-NEXT: [[TMP3:%.*]] = atomicrmw fadd ptr addrspace(4) [[TMP1]], float [[TMP2]] syncscope("device") monotonic, align 4, !atomic.ignore.denormal.mode [[META5:![0-9]+]], !amdgpu.no.fine.grained.memory [[META5]]
+// SPIRV-NEXT: store float [[TMP3]], ptr addrspace(4) [[RESULT_ASCAST]], align 4
// SPIRV-NEXT: br label %[[IF_END]]
// SPIRV: [[IF_END]]:
// SPIRV-NEXT: ret void
@@ -89,10 +88,9 @@ __device__ void test_flat_atomic_fadd_f32_flat(float *ptr, float val) {
// SPIRV-NEXT: br i1 [[TMP0]], label %[[IF_THEN:.*]], label %[[IF_END:.*]]
// SPIRV: [[IF_THEN]]:
// SPIRV-NEXT: [[TMP1:%.*]] = load ptr addrspace(4), ptr addrspace(4) [[PTR_ADDR_ASCAST]], align 8
-// SPIRV-NEXT: [[TMP2:%.*]] = addrspacecast ptr addrspace(4) [[TMP1]] to ptr
-// SPIRV-NEXT: [[TMP3:%.*]] = load double, ptr addrspace(4) [[VAL_ADDR_ASCAST]], align 8
-// SPIRV-NEXT: [[TMP4:%.*]] = atomicrmw fadd ptr [[TMP2]], double [[TMP3]] syncscope("device") monotonic, align 8, !amdgpu.no.fine.grained.memory [[META5]]
-// SPIRV-NEXT: store double [[TMP4]], ptr addrspace(4) [[RESULT_ASCAST]], align 8
+// SPIRV-NEXT: [[TMP2:%.*]] = load double, ptr addrspace(4) [[VAL_ADDR_ASCAST]], align 8
+// SPIRV-NEXT: [[TMP3:%.*]] = atomicrmw fadd ptr addrspace(4) [[TMP1]], double [[TMP2]] syncscope("device") monotonic, align 8, !amdgpu.no.fine.grained.memory [[META5]]
+// SPIRV-NEXT: store double [[TMP3]], ptr addrspace(4) [[RESULT_ASCAST]], align 8
// SPIRV-NEXT: br label %[[IF_END]]
// SPIRV: [[IF_END]]:
// SPIRV-NEXT: ret void
@@ -134,10 +132,9 @@ __device__ void test_flat_atomic_fadd_f64_flat(double *ptr, double val) {
// SPIRV-NEXT: br i1 [[TMP0]], label %[[IF_THEN:.*]], label %[[IF_END:.*]]
// SPIRV: [[IF_THEN]]:
// SPIRV-NEXT: [[TMP1:%.*]] = load ptr addrspace(4), ptr addrspace(4) [[PTR_ADDR_ASCAST]], align 8
-// SPIRV-NEXT: [[TMP2:%.*]] = addrspacecast ptr addrspace(4) [[TMP1]] to ptr
-// SPIRV-NEXT: [[TMP3:%.*]] = load float, ptr addrspace(4) [[VAL_ADDR_ASCAST]], align 4
-// SPIRV-NEXT: [[TMP4:%.*]] = atomicrmw fadd ptr [[TMP2]], float [[TMP3]] syncscope("device") monotonic, align 4, !atomic.ignore.denormal.mode [[META5]], !amdgpu.no.fine.grained.memory [[META5]]
-// SPIRV-NEXT: store float [[TMP4]], ptr addrspace(4) [[RESULT_ASCAST]], align 4
+// SPIRV-NEXT: [[TMP2:%.*]] = load float, ptr addrspace(4) [[VAL_ADDR_ASCAST]], align 4
+// SPIRV-NEXT: [[TMP3:%.*]] = atomicrmw fadd ptr addrspace(4) [[TMP1]], float [[TMP2]] syncscope("device") monotonic, align 4, !atomic.ignore.denormal.mode [[META5]], !amdgpu.no.fine.grained.memory [[META5]]
+// SPIRV-NEXT: store float [[TMP3]], ptr addrspace(4) [[RESULT_ASCAST]], align 4
// SPIRV-NEXT: br label %[[IF_END]]
// SPIRV: [[IF_END]]:
// SPIRV-NEXT: ret void
@@ -179,10 +176,9 @@ __device__ void test_flat_atomic_fadd_f32_shared(__shared__ float *ptr, float va
// SPIRV-NEXT: br i1 [[TMP0]], label %[[IF_THEN:.*]], label %[[IF_END:.*]]
// SPIRV: [[IF_THEN]]:
// SPIRV-NEXT: [[TMP1:%.*]] = load ptr addrspace(4), ptr addrspace(4) [[PTR_ADDR_ASCAST]], align 8
-// SPIRV-NEXT: [[TMP2:%.*]] = addrspacecast ptr addrspace(4) [[TMP1]] to ptr
-// SPIRV-NEXT: [[TMP3:%.*]] = load double, ptr addrspace(4) [[VAL_ADDR_ASCAST]], align 8
-// SPIRV-NEXT: [[TMP4:%.*]] = atomicrmw fadd ptr [[TMP2]], double [[TMP3]] syncscope("device") monotonic, align 8, !amdgpu.no.fine.grained.memory [[META5]]
-// SPIRV-NEXT: store double [[TMP4]], ptr addrspace(4) [[RESULT_ASCAST]], align 8
+// SPIRV-NEXT: [[TMP2:%.*]] = load double, ptr addrspace(4) [[VAL_ADDR_ASCAST]], align 8
+// SPIRV-NEXT: [[TMP3:%.*]] = atomicrmw fadd ptr addrspace(4) [[TMP1]], double [[TMP2]] syncscope("device") monotonic, align 8, !amdgpu.no.fine.grained.memory [[META5]]
+// SPIRV-NEXT: store double [[TMP3]], ptr addrspace(4) [[RESULT_ASCAST]], align 8
// SPIRV-NEXT: br label %[[IF_END]]
// SPIRV: [[IF_END]]:
// SPIRV-NEXT: ret void
@@ -217,7 +213,7 @@ __device__ void test_flat_atomic_fadd_f64_shared(__shared__ double *ptr, double
// SPIRV-NEXT: br i1 [[TMP0]], label %[[IF_THEN:.*]], label %[[IF_END:.*]]
// SPIRV: [[IF_THEN]]:
// SPIRV-NEXT: [[TMP1:%.*]] = load float, ptr addrspace(4) [[VAL_ADDR_ASCAST]], align 4
-// SPIRV-NEXT: [[TMP2:%.*]] = atomicrmw fadd ptr addrspacecast (ptr addrspace(1) @const_float to ptr), float [[TMP1]] syncscope("device") monotonic, align 4, !atomic.ignore.denormal.mode [[META5]], !amdgpu.no.fine.grained.memory [[META5]]
+// SPIRV-NEXT: [[TMP2:%.*]] = atomicrmw fadd ptr addrspace(4) addrspacecast (ptr addrspace(1) @const_float to ptr addrspace(4)), float [[TMP1]] syncscope("device") monotonic, align 4, !atomic.ignore.denormal.mode [[META5]], !amdgpu.no.fine.grained.memory [[META5]]
// SPIRV-NEXT: store float [[TMP2]], ptr addrspace(4) [[RESULT_ASCAST]], align 4
// SPIRV-NEXT: br label %[[IF_END]]
// SPIRV: [[IF_END]]:
@@ -253,7 +249,7 @@ __device__ void test_flat_atomic_fadd_f32_constant(float val) {
// SPIRV-NEXT: br i1 [[TMP0]], label %[[IF_THEN:.*]], label %[[IF_END:.*]]
// SPIRV: [[IF_THEN]]:
// SPIRV-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(4) [[VAL_ADDR_ASCAST]], align 8
-// SPIRV-NEXT: [[TMP2:%.*]] = atomicrmw fadd ptr addrspacecast (ptr addrspace(1) @const_double to ptr), double [[TMP1]] syncscope("device") monotonic, align 8, !amdgpu.no.fine.grained.memory [[META5]]
+// SPIRV-NEXT: [[TMP2:%.*]] = atomicrmw fadd ptr addrspace(4) addrspacecast (ptr addrspace(1) @const_double to ptr addrspace(4)), double [[TMP1]] syncscope("device") monotonic, align 8, !amdgpu.no.fine.grained.memory [[META5]]
// SPIRV-NEXT: store double [[TMP2]], ptr addrspace(4) [[RESULT_ASCAST]], align 8
// SPIRV-NEXT: br label %[[IF_END]]
// SPIRV: [[IF_END]]:
@@ -289,7 +285,7 @@ __device__ void test_flat_atomic_fadd_f64_constant(double val) {
// SPIRV-NEXT: br i1 [[TMP0]], label %[[IF_THEN:.*]], label %[[IF_END:.*]]
// SPIRV: [[IF_THEN]]:
// SPIRV-NEXT: [[TMP1:%.*]] = load float, ptr addrspace(4) [[VAL_ADDR_ASCAST]], align 4
-// SPIRV-NEXT: [[TMP2:%.*]] = atomicrmw fadd ptr addrspacecast (ptr addrspace(1) @global_float to ptr), float [[TMP1]] syncscope("device") monotonic, align 4, !atomic.ignore.denormal.mode [[META5]], !amdgpu.no.fine.grained.memory [[META5]]
+// SPIRV-NEXT: [[TMP2:%.*]] = atomicrmw fadd ptr addrspace(4) addrspacecast (ptr addrspace(1) @global_float to ptr addrspace(4)), float [[TMP1]] syncscope("device") monotonic, align 4, !atomic.ignore.denormal.mode [[META5]], !amdgpu.no.fine.grained.memory [[META5]]
// SPIRV-NEXT: store float [[TMP2]], ptr addrspace(4) [[RESULT_ASCAST]], align 4
// SPIRV-NEXT: br label %[[IF_END]]
// SPIRV: [[IF_END]]:
@@ -325,7 +321,7 @@ __device__ void test_flat_atomic_fadd_f32_global(float val) {
// SPIRV-NEXT: br i1 [[TMP0]], label %[[IF_THEN:.*]], label %[[IF_END:.*]]
// SPIRV: [[IF_THEN]]:
// SPIRV-NEXT: [[TMP1:%.*]] = load double, ptr addrspace(4) [[VAL_ADDR_ASCAST]], align 8
-// SPIRV-NEXT: [[TMP2:%.*]] = atomicrmw fadd ptr addrspacecast (ptr addrspace(1) @global_double to ptr), double [[TMP1]] syncscope("device") monotonic, align 8, !amdgpu.no.fine.grained.memory [[META5]]
+// SPIRV-NEXT: [[TMP2:%.*]] = atomicrmw fadd ptr addrspace(4) addrspacecast (ptr addrspace(1) @global_double to ptr addrspace(4)), double [[TMP1]] syncscope("device") monotonic, align 8, !amdgpu.no.fine.grained.memory [[META5]]
// SPIRV-NEXT: store double [[TMP2]], ptr addrspace(4) [[RESULT_ASCAST]], align 8
// SPIRV-NEXT: br label %[[IF_END]]
// SPIRV: [[IF_END]]:
@@ -335,10 +331,106 @@ __device__ void test_flat_atomic_fadd_f64_global(double val) {
if (__builtin_amdgcn_is_invocable(__builtin_amdgcn_flat_atomic_fadd_f64))
double result = __builtin_amdgcn_flat_atomic_fadd_f64(&global_double, val);
}
+
+typedef _Float16 __attribute__((ext_vector_type(2))) half2;
+typedef short __attribute__((ext_vector_type(2))) short2;
+
+// CHECK-LABEL: define dso_local void @_Z32test_flat_atomic_fadd_v2f16_flatPDv2_DF16_S_(
+// CHECK-SAME: ptr noundef [[PTR:%.*]], <2 x half> noundef [[VAL:%.*]]) #[[ATTR0]] {
+// CHECK-NEXT: [[ENTRY:.*:]]
+// CHECK-NEXT: [[PTR_ADDR:%.*]] = alloca ptr, align 8, addrspace(5)
+// CHECK-NEXT: [[VAL_ADDR:%.*]] = alloca <2 x half>, align 4, addrspace(5)
+// CHECK-NEXT: [[RESULT:%.*]] = alloca <2 x half>, align 4, addrspace(5)
+// CHECK-NEXT: [[PTR_ADDR_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[PTR_ADDR]] to ptr
+// CHECK-NEXT: [[VAL_ADDR_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[VAL_ADDR]] to ptr
+// CHECK-NEXT: [[RESULT_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[RESULT]] to ptr
+// CHECK-NEXT: store ptr [[PTR]], ptr [[PTR_ADDR_ASCAST]], align 8
+// CHECK-NEXT: store <2 x half> [[VAL]], ptr [[VAL_ADDR_ASCAST]], align 4
+// CHECK-NEXT: [[TMP0:%.*]] = load ptr, ptr [[PTR_ADDR_ASCAST]], align 8
+// CHECK-NEXT: [[TMP1:%.*]] = load <2 x half>, ptr [[VAL_ADDR_ASCAST]], align 4
+// CHECK-NEXT: [[TMP2:%.*]] = atomicrmw fadd ptr [[TMP0]], <2 x half> [[TMP1]] syncscope("agent") monotonic, align 4, !amdgpu.no.fine.grained.memory [[META3]]
+// CHECK-NEXT: store <2 x half> [[TMP2]], ptr [[RESULT_ASCAST]], align 4
+// CHECK-NEXT: ret void
+//
+// SPIRV-LABEL: define spir_func void @_Z32test_flat_atomic_fadd_v2f16_flatPDv2_DF16_S_(
+// SPIRV-SAME: ptr addrspace(4) noundef [[PTR:%.*]], <2 x half> noundef [[VAL:%.*]]) addrspace(4) #[[ATTR0]] {
+// SPIRV-NEXT: [[ENTRY:.*:]]
+// SPIRV-NEXT: [[PTR_ADDR:%.*]] = alloca ptr addrspace(4), align 8
+// SPIRV-NEXT: [[VAL_ADDR:%.*]] = alloca <2 x half>, align 4
+// SPIRV-NEXT: [[RESULT:%.*]] = alloca <2 x half>, align 4
+// SPIRV-NEXT: [[PTR_ADDR_ASCAST:%.*]] = addrspacecast ptr [[PTR_ADDR]] to ptr addrspace(4)
+// SPIRV-NEXT: [[VAL_ADDR_ASCAST:%.*]] = addrspacecast ptr [[VAL_ADDR]] to ptr addrspace(4)
+// SPIRV-NEXT: [[RESULT_ASCAST:%.*]] = addrspacecast ptr [[RESULT]] to ptr addrspace(4)
+// SPIRV-NEXT: store ptr addrspace(4) [[PTR]], ptr addrspace(4) [[PTR_ADDR_ASCAST]], align 8
+// SPIRV-NEXT: store <2 x half> [[VAL]], ptr addrspace(4) [[VAL_ADDR_ASCAST]], align 4
+// SPIRV-NEXT: [[TMP0:%.*]] = call addrspace(4) i1 @llvm.spv.named.boolean.spec.constant(i32 -1, i1 false, metadata [[META7:![0-9]+]])
+// SPIRV-NEXT: br i1 [[TMP0]], label %[[IF_THEN:.*]], label %[[IF_END:.*]]
+// SPIRV: [[IF_THEN]]:
+// SPIRV-NEXT: [[TMP1:%.*]] = load ptr addrspace(4), ptr addrspace(4) [[PTR_ADDR_ASCAST]], align 8
+// SPIRV-NEXT: [[TMP2:%.*]] = load <2 x half>, ptr addrspace(4) [[VAL_ADDR_ASCAST]], align 4
+// SPIRV-NEXT: [[TMP3:%.*]] = atomicrmw fadd ptr addrspace(4) [[TMP1]], <2 x half> [[TMP2]] syncscope("device") monotonic, align 4, !amdgpu.no.fine.grained.memory [[META5]]
+// SPIRV-NEXT: store <2 x half> [[TMP3]], ptr addrspace(4) [[RESULT_ASCAST]], align 4
+// SPIRV-NEXT: br label %[[IF_END]]
+// SPIRV: [[IF_END]]:
+// SPIRV-NEXT: ret void
+//
+__device__ void test_flat_atomic_fadd_v2f16_flat(half2 *ptr, half2 val) {
+ if (__builtin_amdgcn_is_invocable(__builtin_amdgcn_flat_atomic_fadd_v2f16))
+ half2 result = __builtin_amdgcn_flat_atomic_fadd_v2f16(ptr, val);
+}
+
+// CHECK-LABEL: define dso_local void @_Z33test_flat_atomic_fadd_v2bf16_flatPDv2_sS_(
+// CHECK-SAME: ptr noundef [[PTR:%.*]], <2 x i16> noundef [[VAL:%.*]]) #[[ATTR0]] {
+// CHECK-NEXT: [[ENTRY:.*:]]
+// CHECK-NEXT: [[PTR_ADDR:%.*]] = alloca ptr, align 8, addrspace(5)
+// CHECK-NEXT: [[VAL_ADDR:%.*]] = alloca <2 x i16>, align 4, addrspace(5)
+// CHECK-NEXT: [[RESULT:%.*]] = alloca <2 x i16>, align 4, addrspace(5)
+// CHECK-NEXT: [[PTR_ADDR_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[PTR_ADDR]] to ptr
+// CHECK-NEXT: [[VAL_ADDR_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[VAL_ADDR]] to ptr
+// CHECK-NEXT: [[RESULT_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[RESULT]] to ptr
+// CHECK-NEXT: store ptr [[PTR]], ptr [[PTR_ADDR_ASCAST]], align 8
+// CHECK-NEXT: store <2 x i16> [[VAL]], ptr [[VAL_ADDR_ASCAST]], align 4
+// CHECK-NEXT: [[TMP0:%.*]] = load ptr, ptr [[PTR_ADDR_ASCAST]], align 8
+// CHECK-NEXT: [[TMP1:%.*]] = load <2 x i16>, ptr [[VAL_ADDR_ASCAST]], align 4
+// CHECK-NEXT: [[TMP2:%.*]] = bitcast <2 x i16> [[TMP1]] to <2 x bfloat>
+// CHECK-NEXT: [[TMP3:%.*]] = atomicrmw fadd ptr [[TMP0]], <2 x bfloat> [[TMP2]] syncscope("agent") monotonic, align 4, !amdgpu.no.fine.grained.memory [[META3]]
+// CHECK-NEXT: [[TMP4:%.*]] = bitcast <2 x bfloat> [[TMP3]] to <2 x i16>
+// CHECK-NEXT: store <2 x i16> [[TMP4]], ptr [[RESULT_ASCAST]], align 4
+// CHECK-NEXT: ret void
+//
+// SPIRV-LABEL: define spir_func void @_Z33test_flat_atomic_fadd_v2bf16_flatPDv2_sS_(
+// SPIRV-SAME: ptr addrspace(4) noundef [[PTR:%.*]], <2 x i16> noundef [[VAL:%.*]]) addrspace(4) #[[ATTR0]] {
+// SPIRV-NEXT: [[ENTRY:.*:]]
+// SPIRV-NEXT: [[PTR_ADDR:%.*]] = alloca ptr addrspace(4), align 8
+// SPIRV-NEXT: [[VAL_ADDR:%.*]] = alloca <2 x i16>, align 4
+// SPIRV-NEXT: [[RESULT:%.*]] = alloca <2 x i16>, align 4
+// SPIRV-NEXT: [[PTR_ADDR_ASCAST:%.*]] = addrspacecast ptr [[PTR_ADDR]] to ptr addrspace(4)
+// SPIRV-NEXT: [[VAL_ADDR_ASCAST:%.*]] = addrspacecast ptr [[VAL_ADDR]] to ptr addrspace(4)
+// SPIRV-NEXT: [[RESULT_ASCAST:%.*]] = addrspacecast ptr [[RESULT]] to ptr addrspace(4)
+// SPIRV-NEXT: store ptr addrspace(4) [[PTR]], ptr addrspace(4) [[PTR_ADDR_ASCAST]], align 8
+// SPIRV-NEXT: store <2 x i16> [[VAL]], ptr addrspace(4) [[VAL_ADDR_ASCAST]], align 4
+// SPIRV-NEXT: [[TMP0:%.*]] = call addrspace(4) i1 @llvm.spv.named.boolean.spec.constant(i32 -1, i1 false, metadata [[META7]])
+// SPIRV-NEXT: br i1 [[TMP0]], label %[[IF_THEN:.*]], label %[[IF_END:.*]]
+// SPIRV: [[IF_THEN]]:
+// SPIRV-NEXT: [[TMP1:%.*]] = load ptr addrspace(4), ptr addrspace(4) [[PTR_ADDR_ASCAST]], align 8
+// SPIRV-NEXT: [[TMP2:%.*]] = load <2 x i16>, ptr addrspace(4) [[VAL_ADDR_ASCAST]], align 4
+// SPIRV-NEXT: [[TMP3:%.*]] = bitcast <2 x i16> [[TMP2]] to <2 x bfloat>
+// SPIRV-NEXT: [[TMP4:%.*]] = atomicrmw fadd ptr addrspace(4) [[TMP1]], <2 x bfloat> [[TMP3]] syncscope("device") monotonic, align 4, !amdgpu.no.fine.grained.memory [[META5]]
+// SPIRV-NEXT: [[TMP5:%.*]] = bitcast <2 x bfloat> [[TMP4]] to <2 x i16>
+// SPIRV-NEXT: store <2 x i16> [[TMP5]], ptr addrspace(4) [[RESULT_ASCAST]], align 4
+// SPIRV-NEXT: br label %[[IF_END]]
+// SPIRV: [[IF_END]]:
+// SPIRV-NEXT: ret void
+//
+__device__ void test_flat_atomic_fadd_v2bf16_flat(short2 *ptr, short2 val) {
+ if (__builtin_amdgcn_is_invocable(__builtin_amdgcn_flat_atomic_fadd_v2bf16))
+ short2 result = __builtin_amdgcn_flat_atomic_fadd_v2bf16(ptr, val);
+}
//.
// CHECK: [[META3]] = !{}
//.
// SPIRV: [[META4]] = !{!"has.gfx940-insts"}
// SPIRV: [[META5]] = !{}
// SPIRV: [[META6]] = !{!"has.gfx90a-insts"}
+// SPIRV: [[META7]] = !{!"has.atomic-flat-pk-add-16-insts"}
//.
More information about the cfe-commits
mailing list