saiislam created this revision.
saiislam added reviewers: JonChesterfield, sameerds, yaxunl, gregrodgers,
b-sumner.
Herald added subscribers: cfe-commits, jfb.
Herald added a project: clang.
"__buitlin_memory_fence(<ordering>, <target-scope>)" in clang gets
mapped to "fence [syncscope("<target-scope>")] <ordering>; yields void"
in IR.
Repository:
rC Clang
https://reviews.llvm.org/D75917
Files:
clang/include/clang/Basic/Builtins.def
clang/include/clang/Basic/DiagnosticSemaKinds.td
clang/lib/CodeGen/CGBuiltin.cpp
clang/lib/Sema/SemaChecking.cpp
clang/test/CodeGenOpenCL/atomic-ops.cl
clang/test/SemaOpenCL/atomic-ops.cl
Index: clang/test/SemaOpenCL/atomic-ops.cl
===================================================================
--- clang/test/SemaOpenCL/atomic-ops.cl
+++ clang/test/SemaOpenCL/atomic-ops.cl
@@ -194,3 +194,21 @@
// The 'expected' pointer shouldn't be NULL.
(void)__opencl_atomic_compare_exchange_strong(Ap, (void *)0, val, memory_order_relaxed, memory_order_relaxed, memory_scope_work_group); // expected-warning {{null passed to a callee that requires a non-null argument}}
}
+
+void memory_fence_errors() {
+ __builtin_memory_fence(memory_order_seq_cst + 1, memory_scope_work_group); // expected-error {{memory order argument to fence operation is invalid}}
+
+ __builtin_memory_fence(memory_order_seq_cst, memory_scope_sub_group + 1); // expected-error {{synchronization scope argument to fence operation is invalid}}
+
+ __builtin_memory_fence(memory_scope_work_group); // expected-error {{too few arguments to function call, expected 2}}
+
+ __builtin_memory_fence(2, 2, 2); // expected-error {{too many arguments to function call, expected 2}}
+
+ __builtin_memory_fence("string", memory_scope_work_group); // expected-error {{used type '__constant char [7]' where unsigned is required}}
+
+ __builtin_memory_fence(3.14, memory_scope_work_group); // expected-error {{used type 'double' where unsigned is required}}
+
+ __builtin_memory_fence(memory_order_seq_cst, "string"); // expected-error {{used type '__constant char [7]' where unsigned is required}}
+
+ __builtin_memory_fence(memory_order_seq_cst, 3.14); // expected-error {{used type 'double' where unsigned is required}}
+}
Index: clang/test/CodeGenOpenCL/atomic-ops.cl
===================================================================
--- clang/test/CodeGenOpenCL/atomic-ops.cl
+++ clang/test/CodeGenOpenCL/atomic-ops.cl
@@ -288,4 +288,44 @@
return __opencl_atomic_load(i, memory_order_seq_cst, memory_scope_work_group);
}
+void test_memory_fence() {
+ // CHECK-LABEL: @test_memory_fence
+
+ // CHECK: fence syncscope("workgroup-one-as") acquire
+ __builtin_memory_fence(memory_order_acquire, memory_scope_work_group);
+
+ // CHECK: fence syncscope("agent-one-as") acquire
+ __builtin_memory_fence(memory_order_acquire, memory_scope_device);
+
+ // CHECK: fence syncscope("one-as") acquire
+ __builtin_memory_fence(memory_order_acquire, memory_scope_all_svm_devices);
+
+ // CHECK: fence syncscope("wavefront-one-as") acquire
+ __builtin_memory_fence(memory_order_acquire, memory_scope_sub_group);
+
+ // CHECK: fence syncscope("workgroup-one-as") release
+ __builtin_memory_fence(memory_order_release, memory_scope_work_group);
+
+ // CHECK: fence syncscope("agent-one-as") release
+ __builtin_memory_fence(memory_order_release, memory_scope_device);
+
+ // CHECK: fence syncscope("one-as") release
+ __builtin_memory_fence(memory_order_release, memory_scope_all_svm_devices);
+
+ // CHECK: fence syncscope("wavefront-one-as") release
+ __builtin_memory_fence(memory_order_release, memory_scope_sub_group);
+
+ // CHECK: fence syncscope("workgroup") seq_cst
+ __builtin_memory_fence(memory_order_seq_cst, memory_scope_work_group);
+
+ // CHECK: fence syncscope("agent") seq_cst
+ __builtin_memory_fence(memory_order_seq_cst, memory_scope_device);
+
+ // CHECK: fence seq_cst
+ __builtin_memory_fence(memory_order_seq_cst, memory_scope_all_svm_devices);
+
+ // CHECK: fence syncscope("wavefront") seq_cst
+ __builtin_memory_fence(memory_order_seq_cst, memory_scope_sub_group);
+}
+
#endif
Index: clang/lib/Sema/SemaChecking.cpp
===================================================================
--- clang/lib/Sema/SemaChecking.cpp
+++ clang/lib/Sema/SemaChecking.cpp
@@ -1865,6 +1865,60 @@
: "__builtin_frame_address")
<< TheCall->getSourceRange();
break;
+
+ // Clang builtins to expose llvm fence instruction
+ case Builtin::BI__builtin_memory_fence: {
+ if (checkArgCount(*this, TheCall, 2))
+ return true;
+
+ // Order should be the first argument
+ ExprResult Arg = TheCall->getArg(0);
+ auto ArgExpr = Arg.get();
+ auto Ty = ArgExpr->getType();
+
+ // Check if Order is an unsigned
+ if (!Ty->isIntegerType()) {
+ Diag(ArgExpr->getExprLoc(), diag::err_typecheck_expect_uint) << Ty;
+ return ExprError();
+ }
+
+ Expr::EvalResult ArgResult;
+ ArgExpr->EvaluateAsInt(ArgResult, Context);
+ int ord = ArgResult.Val.getInt().getZExtValue();
+
+ // Check if Order is one of the valid types
+ if (!llvm::isValidAtomicOrderingCABI(ord)) {
+ Diag(ArgExpr->getBeginLoc(),
+ diag::err_memory_fence_has_invalid_memory_order)
+ << ArgExpr->getSourceRange();
+ return ExprError();
+ }
+
+ // Sync scope should be the second argument
+ Arg = TheCall->getArg(1);
+ ArgExpr = Arg.get();
+ Ty = ArgExpr->getType();
+
+ // Check if SyncScope is an unsgined
+ if (!Ty->isIntegerType()) {
+ Diag(ArgExpr->getExprLoc(), diag::err_typecheck_expect_uint) << Ty;
+ return ExprError();
+ }
+
+ ArgExpr->EvaluateAsInt(ArgResult, Context);
+ int scp = ArgResult.Val.getInt().getZExtValue();
+ auto SM = AtomicScopeModel::create(AtomicScopeModelKind::OpenCL);
+ AtomicScopeModel *ScopeModel = SM.get();
+
+ // Check if Scope is one of the valid types
+ if (!ScopeModel->isValid(scp)) {
+ Diag(ArgExpr->getBeginLoc(),
+ diag::err_memory_fence_has_invalid_synch_scope)
+ << ArgExpr->getSourceRange();
+ return ExprError();
+ }
+ break;
+ }
}
// Since the target specific builtins for each arch overlap, only check those
Index: clang/lib/CodeGen/CGBuiltin.cpp
===================================================================
--- clang/lib/CodeGen/CGBuiltin.cpp
+++ clang/lib/CodeGen/CGBuiltin.cpp
@@ -3698,6 +3698,42 @@
case Builtin::BI__builtin_coro_param:
return EmitCoroutineIntrinsic(E, Intrinsic::coro_param);
+ // Clang builtins to expose llvm fence instruction
+ case Builtin::BI__builtin_memory_fence: {
+ llvm::AtomicOrdering AO;
+ llvm::SyncScope::ID SSID;
+ Value *Order = EmitScalarExpr(E->getArg(0));
+ Value *Scope = EmitScalarExpr(E->getArg(1));
+ auto ScopeModel = AtomicScopeModel::create(AtomicScopeModelKind::OpenCL);
+
+ if ( isa<llvm::ConstantInt>(Order) && isa<llvm::ConstantInt>(Scope) ) {
+ int scp = cast<llvm::ConstantInt>(Scope)->getZExtValue();
+ int ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
+ switch (ord) {
+ case 0: // memory_order_relaxed
+ default: // invalid order
+ break;
+ case 1: // memory_order_consume
+ case 2: // memory_order_acquire
+ AO = llvm::AtomicOrdering::Acquire;
+ break;
+ case 3: // memory_order_release
+ AO = llvm::AtomicOrdering::Release;
+ break;
+ case 4: // memory_order_acq_rel
+ AO = llvm::AtomicOrdering::AcquireRelease;
+ break;
+ case 5: // memory_order_seq_cst
+ AO = llvm::AtomicOrdering::SequentiallyConsistent;
+ break;
+ }
+ SSID = getTargetHooks().getLLVMSyncScopeID(getLangOpts(), ScopeModel->map(scp), AO, getLLVMContext());
+ Builder.CreateFence(AO, SSID);
+ return RValue::get(nullptr);
+ }
+ LLVM_FALLTHROUGH;
+ }
+
// OpenCL v2.0 s6.13.16.2, Built-in pipe read and write functions
case Builtin::BIread_pipe:
case Builtin::BIwrite_pipe: {
Index: clang/include/clang/Basic/DiagnosticSemaKinds.td
===================================================================
--- clang/include/clang/Basic/DiagnosticSemaKinds.td
+++ clang/include/clang/Basic/DiagnosticSemaKinds.td
@@ -6225,6 +6225,8 @@
"cannot %select{decrement|increment}1 value of type %0">;
def err_typecheck_expect_int : Error<
"used type %0 where integer is required">;
+def err_typecheck_expect_uint : Error<
+ "used type %0 where unsigned is required">;
def err_typecheck_arithmetic_incomplete_type : Error<
"arithmetic on a pointer to an incomplete type %0">;
def err_typecheck_pointer_arith_function_type : Error<
@@ -7855,6 +7857,10 @@
InGroup<DiagGroup<"atomic-memory-ordering">>;
def err_atomic_op_has_invalid_synch_scope : Error<
"synchronization scope argument to atomic operation is invalid">;
+def err_memory_fence_has_invalid_memory_order : Error<
+ "memory order argument to fence operation is invalid">;
+def err_memory_fence_has_invalid_synch_scope : Error<
+ "synchronization scope argument to fence operation is invalid">;
def warn_atomic_implicit_seq_cst : Warning<
"implicit use of sequentially-consistent atomic may incur stronger memory barriers than necessary">,
InGroup<DiagGroup<"atomic-implicit-seq-cst">>, DefaultIgnore;
Index: clang/include/clang/Basic/Builtins.def
===================================================================
--- clang/include/clang/Basic/Builtins.def
+++ clang/include/clang/Basic/Builtins.def
@@ -1513,6 +1513,9 @@
BUILTIN(__builtin_coro_suspend, "cIb", "n")
BUILTIN(__builtin_coro_param, "bv*v*", "n")
+// Builtin to expose llvm fence instruction
+BUILTIN(__builtin_memory_fence, "vUiUi", "t")
+
// OpenCL v2.0 s6.13.16, s9.17.3.5 - Pipe functions.
// We need the generic prototype, since the packet type could be anything.
LANGBUILTIN(read_pipe, "i.", "tn", OCLC20_LANG)
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
