Introducing DG2 and PVC support for CM

New platforms support in CM compiler

Author: Konstantin Vladimirov

include/clang/Basic/Builtins.def

@@ -872,6 +872,8 @@ BUILTIN(cm_scoreboard_bti, "Ui","ni")
 BUILTIN(cm_scoreboard_deltas, "Q16c", "ni")
 BUILTIN(cm_scoreboard_depcnt, "UiUi", "ni")
 BUILTIN(cm_sbarrier, "vUi", "ni")
+BUILTIN(cm_nbarrier_init, "vUc", "ni")
+BUILTIN(cm_nbarrier_wait, "vUc", "ni")
 BUILTIN(cm_yield, "v", "ni")
 BUILTIN(cm_print_buffer, "LUi", "ni")
 BUILTIN(cm_print_format_index, "icC*", "ni")

include/clang/Basic/TargetOptions.h

@@ -39,6 +39,9 @@ class TargetOptions {
   /// If given, the name of the target CPU to generate code for.
   std::string CPU;
 
+  /// If given, the name of the stepping to generate code for.
+  std::string Stepping;
+
   // If given, revision id to pass to ocloc
   int RevId;
 

include/clang/Driver/CMOptions.td

@@ -238,6 +238,13 @@ def dump_stats_cwd : Flag<["-"], "dump-stats">, Group<cm_Group>,
   HelpText<"Alias for -dump-stats=cwd">, Alias<dump_stats>, AliasArgs<["cwd"]>,
   Flags<[CMOption, CC1Option, CC1AsOption]>;
 
+def Qxcm_doubleGRF : CMIgnoredFlag<"Qxcm_doubleGRF">,
+  HelpText<"increase total GRF number to 256">;
+
+def Qxcm_stepping: CMJoined<"Qxcm_stepping">,
+  HelpText<"Generate code taking platform stepping into account. "
+           "Valid only for some platforms">;
+
 def Qxcm_revid: CMCC1Joined<"Qxcm_revid=">,
   HelpText<"Pass specified revision ID to libocloc "
            "Valid only for some platforms">;

lib/Basic/Targets/GenX.cpp

@@ -61,13 +61,19 @@ bool GenXTargetInfo::handleTargetFeatures(std::vector<std::string> &Features,
                          .Case("RKL", false)
                          .Case("TGLLP", false)
                          .Case("DG1", false)
+                         .Case("ADLS", false)
+                         .Case("ADLP", false)
+                         .Case("DG2", false)
                          .Default(true);
 
   NativeDoubleSupport = llvm::StringSwitch<bool>(CPU)
                             .Case("ICLLP", false)
                             .Case("TGLLP", false)
                             .Case("RKL", false)
                             .Case("DG1", false)
+                            .Case("ADLS", false)
+                            .Case("ADLP", false)
+                            .Case("DG2", false)
                             .Default(true);
 
   // OCL runtime specific headers support
@@ -92,6 +98,11 @@ bool GenXTargetInfo::setCPU(const std::string &Name) {
                       .Case("RKL", true)
                       .Case("DG1", true)
                       .Case("XEHP_SDV", true)
+                      .Case("ADLP", true)
+                      .Case("ADLS", true)
+                      .Case("DG2", true)
+                      .Case("PVC", true)
+                      .Case("PVCXT", true)
                       .Default(false);
 
   if (CPUKnown)

lib/CodeGen/CGBuiltin.cpp

@@ -2295,6 +2295,8 @@ RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl GD, unsigned BuiltinID,
   case Builtin::BIcm_scoreboard_deltas:
   case Builtin::BIcm_scoreboard_depcnt:
   case Builtin::BIcm_sbarrier:
+  case Builtin::BIcm_nbarrier_init:
+  case Builtin::BIcm_nbarrier_wait:
   case Builtin::BIcm_yield :
   case Builtin::BIcm_print_buffer :
   case Builtin::BIcm_print_format_index:

lib/CodeGen/CGCM.cpp

@@ -1423,7 +1423,9 @@ void CGCMRuntime::EmitCMKernelMetadata(const FunctionDecl *FD,
       getMD(llvm::ConstantInt::getNullValue(I32Ty)),
       getMD(llvm::ConstantInt::getNullValue(I32Ty)), // placeholder for arg offsets
       IOKinds,
-      ArgDescs
+      ArgDescs,
+      getMD(llvm::ConstantInt::getNullValue(I32Ty)),
+      getMD(llvm::ConstantInt::getNullValue(I32Ty))
   };
 
   // Add this kernel to the root.

lib/CodeGen/CGCM.h

@@ -634,12 +634,31 @@ class CGCMRuntime {
 
   /// \brief Postprocess cm_bfn implementation builtin
   llvm::Value *HandleBuiltinBFNImpl(CMCallInfo &CallInfo, CMBuiltinKind Kind);
+
+  /// \brief Postprocess cm_srnd implementation builtin
+  llvm::Value *HandleBuiltinSRNDImpl(CMCallInfo &CallInfo, CMBuiltinKind Kind);
   /// \brief Postprocess dpas implementation builtin.
   llvm::Value *HandleBuiltinDPASImpl(CMCallInfo &CallInfo, CMBuiltinKind Kind);
   /// \brief Postprocess dpas2 implementation builtin.
   llvm::Value *HandleBuiltinDPAS2Impl(CMCallInfo &CallInfo, CMBuiltinKind Kind);
   /// \brief Postprocess builtin cm_bf_cvt.
   llvm::Value *HandleBuiltinBFCVTImpl(CMCallInfo &CallInfo, CMBuiltinKind Kind);
+  /// \brief Postprocess builtin cm_qf_cvt.
+  llvm::Value *HandleBuiltinQFCVTImpl(CMCallInfo &CallInfo, CMBuiltinKind Kind);
+  /// \brief Postprocess builtin cm_tf32_cvt.
+  llvm::Value *HandleBuiltinTF32CVTImpl(CMCallInfo &CallInfo, CMBuiltinKind Kind);
+  /// \brief Postprocess block 2d builtins load/store/prefetch.
+  llvm::Value *HandleBuiltinLSC2dImpl(CMCallInfo &CallInfo, CMBuiltinKind Kind);
+
+  /// \brief Postprocess wave 2.5 load/store
+  llvm::Value *HandleBuiltinLSCWaveImpl(CMCallInfo &CallInfo, CMBuiltinKind Kind);
+
+  /// \brief Postprocess BTI-based load/store/prefetch
+  llvm::Value *HandleBuiltinLSCImpl(CMCallInfo &CallInfo, CMBuiltinKind Kind);
+
+  /// \brief Postprocess cm_lsc_fence implementation builtins.
+  llvm::Value *HandleBuiltinLscFenceImpl(CMCallInfo &CallInfo,
+                                         CMBuiltinKind Kind);
 
   /// \brief Emit 1D/2D select expression.
   LValue EmitSelect(CodeGenFunction &CGF, const CMSelectExpr *E, LValue Base);
@@ -712,6 +731,8 @@ class CGCMRuntime {
   /// Emit cm_slm_free builtin call.
   llvm::Value *EmitBuiltinSLMFree(CodeGenFunction &CGF, const CallExpr *E);
 
+  /// Emit cm_nbarrier_init builtin call.
+  void EmitBuiltinNBarrierInit(CodeGenFunction &CGF, const CallExpr *E);
 
   /// \brief Emit one of scatter_scaled, scatter4_scaled.
   llvm::CallInst *EmitScatterScaled(CodeGenFunction &CGF, unsigned IntrinsicID,