[AMDGPU][SPIRV] Use SPIR-V syncscopes for some AMDGCN BIs (#154867)

AMDGCN flavoured SPIR-V allows AMDGCN specific builtins, including those
for scoped fences and some specific RMWs. However, at present we don't
map syncscopes to their SPIR-V equivalents, but rather use the AMDGCN
ones. This ends up pessimising the resulting code as system scope is
used instead of device (agent) or subgroup (wavefront), so we correct
the behaviour, to ensure that we do the right thing during reverse
translation.

Author: AlexVlx

clang/lib/CodeGen/TargetBuiltins/AMDGPU.cpp

@@ -192,9 +192,17 @@ static Value *emitFPIntBuiltin(CodeGenFunction &CGF,
   return CGF.Builder.CreateCall(F, {Src0, Src1});
 }
 
+static inline StringRef mapScopeToSPIRV(StringRef AMDGCNScope) {
+  if (AMDGCNScope == "agent")
+    return "device";
+  if (AMDGCNScope == "wavefront")
+    return "subgroup";
+  return AMDGCNScope;
+}
+
 // For processing memory ordering and memory scope arguments of various
 // amdgcn builtins.
-// \p Order takes a C++11 comptabile memory-ordering specifier and converts
+// \p Order takes a C++11 compatible memory-ordering specifier and converts
 // it into LLVM's memory ordering specifier using atomic C ABI, and writes
 // to \p AO. \p Scope takes a const char * and converts it into AMDGCN
 // specific SyncScopeID and writes it to \p SSID.
@@ -227,6 +235,8 @@ void CodeGenFunction::ProcessOrderScopeAMDGCN(Value *Order, Value *Scope,
   // Some of the atomic builtins take the scope as a string name.
   StringRef scp;
   if (llvm::getConstantStringInfo(Scope, scp)) {
+    if (getTarget().getTriple().isSPIRV())
+      scp = mapScopeToSPIRV(scp);
     SSID = getLLVMContext().getOrInsertSyncScopeID(scp);
     return;
   }
@@ -238,13 +248,19 @@ void CodeGenFunction::ProcessOrderScopeAMDGCN(Value *Order, Value *Scope,
     SSID = llvm::SyncScope::System;
     break;
   case 1: // __MEMORY_SCOPE_DEVICE
-    SSID = getLLVMContext().getOrInsertSyncScopeID("agent");
+    if (getTarget().getTriple().isSPIRV())
+      SSID = getLLVMContext().getOrInsertSyncScopeID("device");
+    else
+      SSID = getLLVMContext().getOrInsertSyncScopeID("agent");
     break;
   case 2: // __MEMORY_SCOPE_WRKGRP
     SSID = getLLVMContext().getOrInsertSyncScopeID("workgroup");
     break;
   case 3: // __MEMORY_SCOPE_WVFRNT
-    SSID = getLLVMContext().getOrInsertSyncScopeID("wavefront");
+    if (getTarget().getTriple().isSPIRV())
+      SSID = getLLVMContext().getOrInsertSyncScopeID("subgroup");
+    else
+      SSID = getLLVMContext().getOrInsertSyncScopeID("wavefront");
     break;
   case 4: // __MEMORY_SCOPE_SINGLE
     SSID = llvm::SyncScope::SingleThread;
@@ -1510,7 +1526,10 @@ Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID,
       //
       // The global/flat cases need to use agent scope to consistently produce
       // the native instruction instead of a cmpxchg expansion.
-      SSID = getLLVMContext().getOrInsertSyncScopeID("agent");
+      if (getTarget().getTriple().isSPIRV())
+        SSID = getLLVMContext().getOrInsertSyncScopeID("device");
+      else
+        SSID = getLLVMContext().getOrInsertSyncScopeID("agent");
       AO = AtomicOrdering::Monotonic;
 
       // The v2bf16 builtin uses i16 instead of a natural bfloat type.