IFC @ r222830

7 files changed, 36 insertions, 9 deletions

diff --git a/contrib/llvm/lib/Target/PTX/PTX.td b/contrib/llvm/lib/Target/PTX/PTX.td
index ae8326e..231866a 100644
--- a/contrib/llvm/lib/Target/PTX/PTX.td
+++ b/contrib/llvm/lib/Target/PTX/PTX.td
@@ -24,6 +24,9 @@ include "llvm/Target/Target.td"
 def FeatureDouble : SubtargetFeature<"double", "SupportsDouble", "true",
                                      "Do not demote .f64 to .f32">;
 
+def FeatureNoFMA  : SubtargetFeature<"no-fma","SupportsFMA", "false",
+                                     "Disable Fused-Multiply Add">;
+
 //===- PTX Version --------------------------------------------------------===//
 
 def FeaturePTX20 : SubtargetFeature<"ptx20", "PTXVersion", "PTX_VERSION_2_0",
@@ -38,6 +41,10 @@ def FeaturePTX22 : SubtargetFeature<"ptx22", "PTXVersion", "PTX_VERSION_2_2",
                                     "Use PTX Language Version 2.2",
                                     [FeaturePTX21]>;
 
+def FeaturePTX23 : SubtargetFeature<"ptx23", "PTXVersion", "PTX_VERSION_2_3",
+                                    "Use PTX Language Version 2.3",
+                                    [FeaturePTX22]>;
+
 //===- PTX Shader Model ---------------------------------------------------===//
 
 def FeatureSM10 : SubtargetFeature<"sm10", "PTXShaderModel", "PTX_SM_1_0",
diff --git a/contrib/llvm/lib/Target/PTX/PTXISelLowering.cpp b/contrib/llvm/lib/Target/PTX/PTXISelLowering.cpp
index 23b93da..e9b1d8c 100644
--- a/contrib/llvm/lib/Target/PTX/PTXISelLowering.cpp
+++ b/contrib/llvm/lib/Target/PTX/PTXISelLowering.cpp
@@ -65,7 +65,9 @@ PTXTargetLowering::PTXTargetLowering(TargetMachine &TM)
   
   // need to lower SETCC of Preds into bitwise logic
   setOperationAction(ISD::SETCC, MVT::i1, Custom);
-  
+
+  setMinFunctionAlignment(2);
+
   // Compute derived properties from the register classes
   computeRegisterProperties();
 }
diff --git a/contrib/llvm/lib/Target/PTX/PTXISelLowering.h b/contrib/llvm/lib/Target/PTX/PTXISelLowering.h
index 6a7e3e6..225c000 100644
--- a/contrib/llvm/lib/Target/PTX/PTXISelLowering.h
+++ b/contrib/llvm/lib/Target/PTX/PTXISelLowering.h
@@ -37,9 +37,6 @@ class PTXTargetLowering : public TargetLowering {
 
     virtual const char *getTargetNodeName(unsigned Opcode) const;
 
-    virtual unsigned getFunctionAlignment(const Function *F) const {
-      return 2; }
-
     virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
 
     virtual SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
diff --git a/contrib/llvm/lib/Target/PTX/PTXInstrInfo.td b/contrib/llvm/lib/Target/PTX/PTXInstrInfo.td
index 1ac9d3f..d5d08be 100644
--- a/contrib/llvm/lib/Target/PTX/PTXInstrInfo.td
+++ b/contrib/llvm/lib/Target/PTX/PTXInstrInfo.td
@@ -36,6 +36,12 @@ def SupportsPTX21       : Predicate<"getSubtarget().supportsPTX21()">;
 def DoesNotSupportPTX21 : Predicate<"!getSubtarget().supportsPTX21()">;
 def SupportsPTX22       : Predicate<"getSubtarget().supportsPTX22()">;
 def DoesNotSupportPTX22 : Predicate<"!getSubtarget().supportsPTX22()">;
+def SupportsPTX23       : Predicate<"getSubtarget().supportsPTX23()">;
+def DoesNotSupportPTX23 : Predicate<"!getSubtarget().supportsPTX23()">;
+
+// Fused-Multiply Add
+def SupportsFMA         : Predicate<"getSubtarget().supportsFMA()">;
+def DoesNotSupportFMA   : Predicate<"!getSubtarget().supportsFMA()">;
 
 //===----------------------------------------------------------------------===//
 // Instruction Pattern Stuff
@@ -627,8 +633,8 @@ def FDIVri64SM10 : InstPTX<(outs RRegf64:$d),
 // In the short term, mad is supported on all PTX versions and we use a
 // default rounding mode no matter what shader model or PTX version.
 // TODO: Allow the rounding mode to be selectable through llc.
-defm FMADSM13 : PTX_FLOAT_4OP<"mad.rn", fmul, fadd>, Requires<[SupportsSM13]>;
-defm FMAD : PTX_FLOAT_4OP<"mad", fmul, fadd>, Requires<[DoesNotSupportSM13]>;
+defm FMADSM13 : PTX_FLOAT_4OP<"mad.rn", fmul, fadd>, Requires<[SupportsSM13, SupportsFMA]>;
+defm FMAD : PTX_FLOAT_4OP<"mad", fmul, fadd>, Requires<[DoesNotSupportSM13, SupportsFMA]>;
 
 ///===- Floating-Point Intrinsic Instructions -----------------------------===//
 
@@ -665,6 +671,8 @@ def FCOS64 : InstPTX<(outs RRegf64:$d),
 
 ///===- Comparison and Selection Instructions -----------------------------===//
 
+// .setp
+
 // Compare u16
 
 defm SETPEQu16 : PTX_SETP_I<RRegu16, "u16", i16imm, SETEQ,  "eq">;
diff --git a/contrib/llvm/lib/Target/PTX/PTXRegisterInfo.h b/contrib/llvm/lib/Target/PTX/PTXRegisterInfo.h
index 67e130f..dc56352 100644
--- a/contrib/llvm/lib/Target/PTX/PTXRegisterInfo.h
+++ b/contrib/llvm/lib/Target/PTX/PTXRegisterInfo.h
@@ -57,6 +57,9 @@ struct PTXRegisterInfo : public PTXGenRegisterInfo {
   virtual int getDwarfRegNum(unsigned RegNum, bool isEH) const {
     return PTXGenRegisterInfo::getDwarfRegNumFull(RegNum, 0);
   }
+  virtual int getLLVMRegNum(unsigned RegNum, bool isEH) const {
+    return PTXGenRegisterInfo::getLLVMRegNumFull(RegNum, 0);
+  }
 }; // struct PTXRegisterInfo
 } // namespace llvm
 
diff --git a/contrib/llvm/lib/Target/PTX/PTXSubtarget.cpp b/contrib/llvm/lib/Target/PTX/PTXSubtarget.cpp
index a224f2b..e8a1dfe 100644
--- a/contrib/llvm/lib/Target/PTX/PTXSubtarget.cpp
+++ b/contrib/llvm/lib/Target/PTX/PTXSubtarget.cpp
@@ -21,7 +21,8 @@ PTXSubtarget::PTXSubtarget(const std::string &TT, const std::string &FS,
   : PTXShaderModel(PTX_SM_1_0),
     PTXVersion(PTX_VERSION_2_0),
     SupportsDouble(false),
-    Is64Bit(is64Bit) {
+    SupportsFMA(true),
+    Is64Bit(is64Bit) {	
   std::string TARGET = "generic";
   ParseSubtargetFeatures(FS, TARGET);
 }
@@ -41,6 +42,7 @@ std::string PTXSubtarget::getPTXVersionString() const {
     case PTX_VERSION_2_0: return "2.0";
     case PTX_VERSION_2_1: return "2.1";
     case PTX_VERSION_2_2: return "2.2";
+    case PTX_VERSION_2_3: return "2.3";
   }
 }
 
diff --git a/contrib/llvm/lib/Target/PTX/PTXSubtarget.h b/contrib/llvm/lib/Target/PTX/PTXSubtarget.h
index 47d9842..59fa696 100644
--- a/contrib/llvm/lib/Target/PTX/PTXSubtarget.h
+++ b/contrib/llvm/lib/Target/PTX/PTXSubtarget.h
@@ -37,7 +37,8 @@ namespace llvm {
       enum PTXVersionEnum {
         PTX_VERSION_2_0,  /*< PTX Version 2.0 */
         PTX_VERSION_2_1,  /*< PTX Version 2.1 */
-        PTX_VERSION_2_2   /*< PTX Version 2.2 */
+        PTX_VERSION_2_2,  /*< PTX Version 2.2 */
+        PTX_VERSION_2_3   /*< PTX Version 2.3 */
       };
 
       /// Shader Model supported on the target GPU.
@@ -48,7 +49,10 @@ namespace llvm {
 
       // The native .f64 type is supported on the hardware.
       bool SupportsDouble;
-
+      
+      // Support the fused-multiply add (FMA) and multiply-add (MAD) instructions
+      bool SupportsFMA;
+      
       // Use .u64 instead of .u32 for addresses.
       bool Is64Bit;
 
@@ -63,6 +67,8 @@ namespace llvm {
 
       bool is64Bit() const { return Is64Bit; }
 
+      bool supportsFMA() const { return SupportsFMA; }
+      
       bool supportsSM13() const { return PTXShaderModel >= PTX_SM_1_3; }
 
       bool supportsSM20() const { return PTXShaderModel >= PTX_SM_2_0; }
@@ -71,6 +77,8 @@ namespace llvm {
 
       bool supportsPTX22() const { return PTXVersion >= PTX_VERSION_2_2; }
 
+      bool supportsPTX23() const { return PTXVersion >= PTX_VERSION_2_3; }
+
       std::string ParseSubtargetFeatures(const std::string &FS,
                                          const std::string &CPU);
   }; // class PTXSubtarget