================
@@ -0,0 +1,324 @@
+// RUN: %clang_cc1 -triple amdgpu-amd-amdhsa -fsycl-is-device
-disable-llvm-passes -Wno-deprecated-attributes -fclangir -emit-cir %s -o %t.cir
+// RUN: FileCheck --input-file=%t.cir %s -check-prefix=CIR
+// RUN: %clang_cc1 -triple amdgpu-amd-amdhsa -fsycl-is-device
-disable-llvm-passes -Wno-deprecated-attributes -fclangir -emit-llvm %s -o
%t-cir.ll
+// RUN: FileCheck --input-file=%t-cir.ll %s --check-prefixes=LLVM,LLVM-CIR
+// RUN: %clang_cc1 -triple amdgpu-amd-amdhsa -fsycl-is-device
-disable-llvm-passes -Wno-deprecated-attributes -emit-llvm %s -o %t.ll
+// RUN: FileCheck --input-file=%t.ll %s --check-prefixes=LLVM,OGCG
+
+// Port of clang/test/CodeGenSYCL/amd-address-space-conversions.cpp.
+
+void bar(int &Data) {}
+void bar2(int &Data) {}
+void bar(int [[clang::sycl_local]] &Data) {}
+void foo(int *Data) {}
+void foo2(int *Data) {}
+void foo(int [[clang::sycl_local]] *Data) {}
+
+template <typename T>
+void tmpl(T t) {}
+
+[[clang::sycl_external]] void usages() {
+ int *NoAS;
+ int [[clang::sycl_global]] *GLOB;
+ int [[clang::sycl_local]] *LOC;
+ int [[clang::sycl_private]] *PRIV;
+ int __attribute__((opencl_global_device)) *GLOBDEVICE;
+ int __attribute__((opencl_global_host)) *GLOBHOST;
+ LOC = nullptr;
+ GLOB = nullptr;
+ NoAS = (int *)GLOB;
+ NoAS = (int *)LOC;
+ NoAS = (int *)PRIV;
+ GLOB = (int [[clang::sycl_global]] *)NoAS;
+ LOC = (int [[clang::sycl_local]] *)NoAS;
+ PRIV = (int [[clang::sycl_private]] *)NoAS;
+ GLOB = (int [[clang::sycl_global]] *)GLOBDEVICE;
+ GLOB = (int [[clang::sycl_global]] *)GLOBHOST;
+ bar(*GLOB);
+ bar2(*GLOB);
+ bar(*LOC);
+ bar2(*LOC);
+ bar(*NoAS);
+ bar2(*NoAS);
+ foo(GLOB);
+ foo2(GLOB);
+ foo(LOC);
+ foo2(LOC);
+ foo(NoAS);
+ foo2(NoAS);
+
+ // Ensure that we still get 3 different template instantiations.
+ tmpl(GLOB);
+ tmpl(LOC);
+ tmpl(PRIV);
+ tmpl(NoAS);
+}
+
+// CIR-LABEL: cir.func {{.*}}@_Z6usagesv(
+// CIR-NEXT: %[[NOAS:.*]] = cir.alloca "NoAS" align(8) :
!cir.ptr<!cir.ptr<!s32i>, target_address_space(5)>
+// CIR-NEXT: %[[GLOB:.*]] = cir.alloca "GLOB" align(8) :
!cir.ptr<!cir.ptr<!s32i, target_address_space(1)>, target_address_space(5)>
+// CIR-NEXT: %[[LOC:.*]] = cir.alloca "LOC" align(4) :
!cir.ptr<!cir.ptr<!s32i, target_address_space(3)>, target_address_space(5)>
+// CIR-NEXT: %[[PRIV:.*]] = cir.alloca "PRIV" align(4) :
!cir.ptr<!cir.ptr<!s32i, target_address_space(5)>, target_address_space(5)>
+// CIR-NEXT: %[[GLOBDEVICE:.*]] = cir.alloca "GLOBDEVICE" align(8) :
!cir.ptr<!cir.ptr<!s32i, target_address_space(1)>, target_address_space(5)>
+// CIR-NEXT: %[[GLOBHOST:.*]] = cir.alloca "GLOBHOST" align(8) :
!cir.ptr<!cir.ptr<!s32i, target_address_space(1)>, target_address_space(5)>
+// CIR-NEXT: %[[GLOBHOST_ASCAST:.*]] = cir.cast address_space %[[GLOBHOST]] :
!cir.ptr<!cir.ptr<!s32i, target_address_space(1)>, target_address_space(5)> ->
!cir.ptr<!cir.ptr<!s32i, target_address_space(1)>>
+// CIR-NEXT: %[[GLOBDEVICE_ASCAST:.*]] = cir.cast address_space
%[[GLOBDEVICE]] : !cir.ptr<!cir.ptr<!s32i, target_address_space(1)>,
target_address_space(5)> -> !cir.ptr<!cir.ptr<!s32i, target_address_space(1)>>
+// CIR-NEXT: %[[PRIV_ASCAST:.*]] = cir.cast address_space %[[PRIV]] :
!cir.ptr<!cir.ptr<!s32i, target_address_space(5)>, target_address_space(5)> ->
!cir.ptr<!cir.ptr<!s32i, target_address_space(5)>>
+// CIR-NEXT: %[[LOC_ASCAST:.*]] = cir.cast address_space %[[LOC]] :
!cir.ptr<!cir.ptr<!s32i, target_address_space(3)>, target_address_space(5)> ->
!cir.ptr<!cir.ptr<!s32i, target_address_space(3)>>
+// CIR-NEXT: %[[GLOB_ASCAST:.*]] = cir.cast address_space %[[GLOB]] :
!cir.ptr<!cir.ptr<!s32i, target_address_space(1)>, target_address_space(5)> ->
!cir.ptr<!cir.ptr<!s32i, target_address_space(1)>>
+// CIR-NEXT: %[[NOAS_ASCAST:.*]] = cir.cast address_space %[[NOAS]] :
!cir.ptr<!cir.ptr<!s32i>, target_address_space(5)> -> !cir.ptr<!cir.ptr<!s32i>>
+// The null value of the local address space is ~0 on AMDGPU, so a null
+// local pointer is an address space cast of the generic null pointer.
+// CIR-NEXT: %[[NULL:.*]] = cir.const #cir.ptr<null> : !cir.ptr<!s32i>
+// CIR-NEXT: %[[LOC_NULL:.*]] = cir.cast address_space %[[NULL]] :
!cir.ptr<!s32i> -> !cir.ptr<!s32i, target_address_space(3)>
+// CIR-NEXT: cir.store align(4) %[[LOC_NULL]], %[[LOC_ASCAST]] :
!cir.ptr<!s32i, target_address_space(3)>, !cir.ptr<!cir.ptr<!s32i,
target_address_space(3)>>
+// CIR-NEXT: %[[GLOB_NULL:.*]] = cir.const #cir.ptr<null> : !cir.ptr<!s32i,
target_address_space(1)>
+// CIR-NEXT: cir.store align(8) %[[GLOB_NULL]], %[[GLOB_ASCAST]] :
!cir.ptr<!s32i, target_address_space(1)>, !cir.ptr<!cir.ptr<!s32i,
target_address_space(1)>>
+
+// CIR-LABEL: cir.func {{.*}}@_Z3barRi(
+// CIR-NEXT: %[[DATA:.*]] = cir.alloca "Data" align(8) init const :
!cir.ptr<!cir.ptr<!s32i>, target_address_space(5)>
+// CIR-NEXT: %[[DATA_ASCAST:.*]] = cir.cast address_space %[[DATA]] :
!cir.ptr<!cir.ptr<!s32i>, target_address_space(5)> -> !cir.ptr<!cir.ptr<!s32i>>
+// CIR-NEXT: cir.store %arg0, %[[DATA_ASCAST]] : !cir.ptr<!s32i>,
!cir.ptr<!cir.ptr<!s32i>>
+// CIR-NEXT: cir.return
+
+// CIR-LABEL: cir.func {{.*}}@_Z3barRU3AS3i(
+// CIR-NEXT: %[[DATA:.*]] = cir.alloca "Data" align(4) init const :
!cir.ptr<!cir.ptr<!s32i, target_address_space(3)>, target_address_space(5)>
+// CIR-NEXT: %[[DATA_ASCAST:.*]] = cir.cast address_space %[[DATA]] :
!cir.ptr<!cir.ptr<!s32i, target_address_space(3)>, target_address_space(5)> ->
!cir.ptr<!cir.ptr<!s32i, target_address_space(3)>>
+// CIR-NEXT: cir.store %arg0, %[[DATA_ASCAST]] : !cir.ptr<!s32i,
target_address_space(3)>, !cir.ptr<!cir.ptr<!s32i, target_address_space(3)>>
+// CIR-NEXT: cir.return
+
+// CIR-LABEL: cir.func {{.*}}@_Z3fooPi(
+// CIR-NEXT: %[[DATA:.*]] = cir.alloca "Data" align(8) init :
!cir.ptr<!cir.ptr<!s32i>, target_address_space(5)>
+// CIR-NEXT: %[[DATA_ASCAST:.*]] = cir.cast address_space %[[DATA]] :
!cir.ptr<!cir.ptr<!s32i>, target_address_space(5)> -> !cir.ptr<!cir.ptr<!s32i>>
+// CIR-NEXT: cir.store %arg0, %[[DATA_ASCAST]] : !cir.ptr<!s32i>,
!cir.ptr<!cir.ptr<!s32i>>
+// CIR-NEXT: cir.return
+
+// CIR-LABEL: cir.func {{.*}}@_Z3fooPU3AS3i(
+// CIR-NEXT: %[[DATA:.*]] = cir.alloca "Data" align(4) init :
!cir.ptr<!cir.ptr<!s32i, target_address_space(3)>, target_address_space(5)>
+// CIR-NEXT: %[[DATA_ASCAST:.*]] = cir.cast address_space %[[DATA]] :
!cir.ptr<!cir.ptr<!s32i, target_address_space(3)>, target_address_space(5)> ->
!cir.ptr<!cir.ptr<!s32i, target_address_space(3)>>
+// CIR-NEXT: cir.store %arg0, %[[DATA_ASCAST]] : !cir.ptr<!s32i,
target_address_space(3)>, !cir.ptr<!cir.ptr<!s32i, target_address_space(3)>>
+// CIR-NEXT: cir.return
+
+// LLVM-CIR-LABEL: define {{.*}}void @_Z6usagesv(
+// LLVM-CIR: %[[NOAS:.*]] = alloca ptr, align 8, addrspace(5)
+// LLVM-CIR-NEXT: %[[GLOB:.*]] = alloca ptr addrspace(1), align 8, addrspace(5)
+// LLVM-CIR-NEXT: %[[LOC:.*]] = alloca ptr addrspace(3), align 4, addrspace(5)
+// LLVM-CIR-NEXT: %[[PRIV:.*]] = alloca ptr addrspace(5), align 4, addrspace(5)
+// LLVM-CIR-NEXT: %[[GLOBDEVICE:.*]] = alloca ptr addrspace(1), align 8,
addrspace(5)
+// LLVM-CIR-NEXT: %[[GLOBHOST:.*]] = alloca ptr addrspace(1), align 8,
addrspace(5)
+// LLVM-CIR-NEXT: %[[GLOBHOST_ASCAST:.*]] = addrspacecast ptr addrspace(5)
%[[GLOBHOST]] to ptr
+// LLVM-CIR-NEXT: %[[GLOBDEVICE_ASCAST:.*]] = addrspacecast ptr addrspace(5)
%[[GLOBDEVICE]] to ptr
+// LLVM-CIR-NEXT: %[[PRIV_ASCAST:.*]] = addrspacecast ptr addrspace(5)
%[[PRIV]] to ptr
+// LLVM-CIR-NEXT: %[[LOC_ASCAST:.*]] = addrspacecast ptr addrspace(5) %[[LOC]]
to ptr
+// LLVM-CIR-NEXT: %[[GLOB_ASCAST:.*]] = addrspacecast ptr addrspace(5)
%[[GLOB]] to ptr
+// LLVM-CIR-NEXT: %[[NOAS_ASCAST:.*]] = addrspacecast ptr addrspace(5)
%[[NOAS]] to ptr
+// LLVM-CIR-NEXT: store ptr addrspace(3) addrspacecast (ptr null to ptr
addrspace(3)), ptr %[[LOC_ASCAST]], align 4
+// LLVM-CIR-NEXT: store ptr addrspace(1) null, ptr %[[GLOB_ASCAST]], align 8
+// LLVM-CIR-NEXT: %[[TMP0:.*]] = load ptr addrspace(1), ptr %[[GLOB_ASCAST]],
align 8
+// LLVM-CIR-NEXT: %[[TMP1:.*]] = addrspacecast ptr addrspace(1) %[[TMP0]] to
ptr
+// LLVM-CIR-NEXT: store ptr %[[TMP1]], ptr %[[NOAS_ASCAST]], align 8
+// LLVM-CIR-NEXT: %[[TMP2:.*]] = load ptr addrspace(3), ptr %[[LOC_ASCAST]],
align 4
+// LLVM-CIR-NEXT: %[[TMP3:.*]] = addrspacecast ptr addrspace(3) %[[TMP2]] to
ptr
+// LLVM-CIR-NEXT: store ptr %[[TMP3]], ptr %[[NOAS_ASCAST]], align 8
+// LLVM-CIR-NEXT: %[[TMP4:.*]] = load ptr addrspace(5), ptr %[[PRIV_ASCAST]],
align 4
+// LLVM-CIR-NEXT: %[[TMP5:.*]] = addrspacecast ptr addrspace(5) %[[TMP4]] to
ptr
+// LLVM-CIR-NEXT: store ptr %[[TMP5]], ptr %[[NOAS_ASCAST]], align 8
+// LLVM-CIR-NEXT: %[[TMP6:.*]] = load ptr, ptr %[[NOAS_ASCAST]], align 8
+// LLVM-CIR-NEXT: %[[TMP7:.*]] = addrspacecast ptr %[[TMP6]] to ptr
addrspace(1)
+// LLVM-CIR-NEXT: store ptr addrspace(1) %[[TMP7]], ptr %[[GLOB_ASCAST]],
align 8
+// LLVM-CIR-NEXT: %[[TMP8:.*]] = load ptr, ptr %[[NOAS_ASCAST]], align 8
+// LLVM-CIR-NEXT: %[[TMP9:.*]] = addrspacecast ptr %[[TMP8]] to ptr
addrspace(3)
+// LLVM-CIR-NEXT: store ptr addrspace(3) %[[TMP9]], ptr %[[LOC_ASCAST]], align
4
+// LLVM-CIR-NEXT: %[[TMP10:.*]] = load ptr, ptr %[[NOAS_ASCAST]], align 8
+// LLVM-CIR-NEXT: %[[TMP11:.*]] = addrspacecast ptr %[[TMP10]] to ptr
addrspace(5)
+// LLVM-CIR-NEXT: store ptr addrspace(5) %[[TMP11]], ptr %[[PRIV_ASCAST]],
align 4
+// LLVM-CIR-NEXT: %[[TMP12:.*]] = load ptr addrspace(1), ptr
%[[GLOBDEVICE_ASCAST]], align 8
+// LLVM-CIR-NEXT: store ptr addrspace(1) %[[TMP12]], ptr %[[GLOB_ASCAST]],
align 8
+// LLVM-CIR-NEXT: %[[TMP13:.*]] = load ptr addrspace(1), ptr
%[[GLOBHOST_ASCAST]], align 8
+// LLVM-CIR-NEXT: store ptr addrspace(1) %[[TMP13]], ptr %[[GLOB_ASCAST]],
align 8
+// LLVM-CIR-NEXT: %[[TMP14:.*]] = load ptr addrspace(1), ptr %[[GLOB_ASCAST]],
align 8
+// LLVM-CIR-NEXT: %[[TMP15:.*]] = addrspacecast ptr addrspace(1) %[[TMP14]] to
ptr
+// LLVM-CIR-NEXT: call void @_Z3barRi(ptr noundef nonnull align 4
dereferenceable(4) %[[TMP15]])
+// LLVM-CIR-NEXT: %[[TMP16:.*]] = load ptr addrspace(1), ptr %[[GLOB_ASCAST]],
align 8
+// LLVM-CIR-NEXT: %[[TMP17:.*]] = addrspacecast ptr addrspace(1) %[[TMP16]] to
ptr
+// LLVM-CIR-NEXT: call void @_Z4bar2Ri(ptr noundef nonnull align 4
dereferenceable(4) %[[TMP17]])
+// LLVM-CIR-NEXT: %[[TMP18:.*]] = load ptr addrspace(3), ptr %[[LOC_ASCAST]],
align 4
+// LLVM-CIR-NEXT: call void @_Z3barRU3AS3i(ptr addrspace(3) noundef align 4
dereferenceable(4) %[[TMP18]])
+// LLVM-CIR-NEXT: %[[TMP19:.*]] = load ptr addrspace(3), ptr %[[LOC_ASCAST]],
align 4
+// LLVM-CIR-NEXT: %[[TMP20:.*]] = addrspacecast ptr addrspace(3) %[[TMP19]] to
ptr
+// LLVM-CIR-NEXT: call void @_Z4bar2Ri(ptr noundef nonnull align 4
dereferenceable(4) %[[TMP20]])
+// LLVM-CIR-NEXT: %[[TMP21:.*]] = load ptr, ptr %[[NOAS_ASCAST]], align 8
+// LLVM-CIR-NEXT: call void @_Z3barRi(ptr noundef nonnull align 4
dereferenceable(4) %[[TMP21]])
+// LLVM-CIR-NEXT: %[[TMP22:.*]] = load ptr, ptr %[[NOAS_ASCAST]], align 8
+// LLVM-CIR-NEXT: call void @_Z4bar2Ri(ptr noundef nonnull align 4
dereferenceable(4) %[[TMP22]])
+// LLVM-CIR-NEXT: %[[TMP23:.*]] = load ptr addrspace(1), ptr %[[GLOB_ASCAST]],
align 8
+// LLVM-CIR-NEXT: %[[TMP24:.*]] = addrspacecast ptr addrspace(1) %[[TMP23]] to
ptr
+// LLVM-CIR-NEXT: call void @_Z3fooPi(ptr noundef %[[TMP24]])
+// LLVM-CIR-NEXT: %[[TMP25:.*]] = load ptr addrspace(1), ptr %[[GLOB_ASCAST]],
align 8
+// LLVM-CIR-NEXT: %[[TMP26:.*]] = addrspacecast ptr addrspace(1) %[[TMP25]] to
ptr
+// LLVM-CIR-NEXT: call void @_Z4foo2Pi(ptr noundef %[[TMP26]])
+// LLVM-CIR-NEXT: %[[TMP27:.*]] = load ptr addrspace(3), ptr %[[LOC_ASCAST]],
align 4
+// LLVM-CIR-NEXT: call void @_Z3fooPU3AS3i(ptr addrspace(3) noundef %[[TMP27]])
+// LLVM-CIR-NEXT: %[[TMP28:.*]] = load ptr addrspace(3), ptr %[[LOC_ASCAST]],
align 4
+// LLVM-CIR-NEXT: %[[TMP29:.*]] = addrspacecast ptr addrspace(3) %[[TMP28]] to
ptr
+// LLVM-CIR-NEXT: call void @_Z4foo2Pi(ptr noundef %[[TMP29]])
+// LLVM-CIR-NEXT: %[[TMP30:.*]] = load ptr, ptr %[[NOAS_ASCAST]], align 8
+// LLVM-CIR-NEXT: call void @_Z3fooPi(ptr noundef %[[TMP30]])
+// LLVM-CIR-NEXT: %[[TMP31:.*]] = load ptr, ptr %[[NOAS_ASCAST]], align 8
+// LLVM-CIR-NEXT: call void @_Z4foo2Pi(ptr noundef %[[TMP31]])
+// LLVM-CIR-NEXT: %[[TMP32:.*]] = load ptr addrspace(1), ptr %[[GLOB_ASCAST]],
align 8
+// LLVM-CIR-NEXT: call void @_Z4tmplIPU3AS1iEvT_(ptr addrspace(1) noundef
%[[TMP32]])
+// LLVM-CIR-NEXT: %[[TMP33:.*]] = load ptr addrspace(3), ptr %[[LOC_ASCAST]],
align 4
+// LLVM-CIR-NEXT: call void @_Z4tmplIPU3AS3iEvT_(ptr addrspace(3) noundef
%[[TMP33]])
+// LLVM-CIR-NEXT: %[[TMP34:.*]] = load ptr addrspace(5), ptr %[[PRIV_ASCAST]],
align 4
+// LLVM-CIR-NEXT: call void @_Z4tmplIPU3AS5iEvT_(ptr addrspace(5) noundef
%[[TMP34]])
+// LLVM-CIR-NEXT: %[[TMP35:.*]] = load ptr, ptr %[[NOAS_ASCAST]], align 8
+// LLVM-CIR-NEXT: call void @_Z4tmplIPiEvT_(ptr noundef %[[TMP35]])
+// LLVM-CIR-NEXT: ret void
+
+// OGCG-LABEL: define {{.*}}void @_Z6usagesv(
+// OGCG: %[[NOAS:.*]] = alloca ptr, align 8, addrspace(5)
+// OGCG-NEXT: %[[GLOB:.*]] = alloca ptr addrspace(1), align 8, addrspace(5)
+// OGCG-NEXT: %[[LOC:.*]] = alloca ptr addrspace(3), align 4, addrspace(5)
+// OGCG-NEXT: %[[PRIV:.*]] = alloca ptr addrspace(5), align 4, addrspace(5)
+// OGCG-NEXT: %[[GLOBDEVICE:.*]] = alloca ptr addrspace(1), align 8,
addrspace(5)
+// OGCG-NEXT: %[[GLOBHOST:.*]] = alloca ptr addrspace(1), align 8, addrspace(5)
+// OGCG-NEXT: %[[NOAS_ASCAST:.*]] = addrspacecast ptr addrspace(5) %[[NOAS]]
to ptr
+// OGCG-NEXT: %[[GLOB_ASCAST:.*]] = addrspacecast ptr addrspace(5) %[[GLOB]]
to ptr
+// OGCG-NEXT: %[[LOC_ASCAST:.*]] = addrspacecast ptr addrspace(5) %[[LOC]] to
ptr
+// OGCG-NEXT: %[[PRIV_ASCAST:.*]] = addrspacecast ptr addrspace(5) %[[PRIV]]
to ptr
+// OGCG-NEXT: %[[GLOBDEVICE_ASCAST:.*]] = addrspacecast ptr addrspace(5)
%[[GLOBDEVICE]] to ptr
+// OGCG-NEXT: %[[GLOBHOST_ASCAST:.*]] = addrspacecast ptr addrspace(5)
%[[GLOBHOST]] to ptr
----------------
steffenlarsen wrote:
Is this the only difference between `OGCG` and `CIR-LLVM`? Could `LLVM-DAG` be
used to ignore the ordering?
https://github.com/llvm/llvm-project/pull/237387
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
