Remove -post-RA-schedule flag and add a TargetSubtarget method to enable post-register-allocation scheduling. By default it is off. For ARM, enable/disable with -mattr=+/-postrasched. Enable by default for cortex-a8.

llvm-svn: 83122

llvm/include/llvm/Target/TargetSubtarget.h

@@ -39,10 +39,14 @@ public:
   /// should be attempted.
   virtual unsigned getSpecialAddressLatency() const { return 0; }
 
+  // enablePostRAScheduler - Return true to enable
+  // post-register-allocation scheduling.
+  virtual bool enablePostRAScheduler() const { return false; }
+
   // adjustSchedDependency - Perform target specific adjustments to
   // the latency of a schedule dependency.
   virtual void adjustSchedDependency(SUnit *def, SUnit *use, 
-                                     SDep& dep) const { };
+                                     SDep& dep) const { }
 };
 
 } // End llvm namespace

llvm/lib/CodeGen/LLVMTargetMachine.cpp

@@ -45,14 +45,6 @@ static cl::opt<bool> VerifyMachineCode("verify-machineinstrs", cl::Hidden,
     cl::desc("Verify generated machine code"),
     cl::init(getenv("LLVM_VERIFY_MACHINEINSTRS")!=NULL));
 
-// This is not enabled by default due to 1) high compile time cost, 2) it's not
-// beneficial to all targets. The plan is to let targets decide whether this
-// is enabled.
-static cl::opt<bool>
-EnablePostRAScheduler("post-RA-scheduler",
-                       cl::desc("Enable scheduling after register allocation"),
-                       cl::init(false));
-
 // Enable or disable FastISel. Both options are needed, because
 // FastISel is enabled by default with -fast, and we wish to be
 // able to enable or disable fast-isel independently from -O0.
@@ -326,7 +318,7 @@ bool LLVMTargetMachine::addCommonCodeGenPasses(PassManagerBase &PM,
   printAndVerify(PM);
 
   // Second pass scheduler.
-  if (OptLevel != CodeGenOpt::None && EnablePostRAScheduler) {
+  if (OptLevel != CodeGenOpt::None) {
     PM.add(createPostRAScheduler());
     printAndVerify(PM);
   }

llvm/lib/CodeGen/PostRASchedulerList.cpp

@@ -34,6 +34,7 @@
 #include "llvm/Target/TargetMachine.h"
 #include "llvm/Target/TargetInstrInfo.h"
 #include "llvm/Target/TargetRegisterInfo.h"
+#include "llvm/Target/TargetSubtarget.h"
 #include "llvm/Support/Compiler.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Support/ErrorHandling.h"
@@ -209,6 +210,11 @@ static bool isSchedulingBoundary(const MachineInstr *MI,
 }
 
 bool PostRAScheduler::runOnMachineFunction(MachineFunction &Fn) {
+  // Check that post-RA scheduling is enabled for this function
+  const TargetSubtarget &ST = Fn.getTarget().getSubtarget<TargetSubtarget>();
+  if (!ST.enablePostRAScheduler())
+    return true;
+
   DEBUG(errs() << "PostRAScheduler\n");
 
   const MachineLoopInfo &MLI = getAnalysis<MachineLoopInfo>();

llvm/lib/Target/ARM/ARM.td

@@ -43,6 +43,9 @@ def FeatureThumb2 : SubtargetFeature<"thumb2", "ThumbMode", "Thumb2",
 def FeatureNEONFP : SubtargetFeature<"neonfp", "UseNEONForSinglePrecisionFP",
                                      "true",
                                      "Use NEON for single-precision FP">;
+def FeaturePostRASched : SubtargetFeature<"postrasched", "PostRAScheduler",
+                                     "true",
+                                     "Use Post-Register-Allocation Scheduler">;
 
 //===----------------------------------------------------------------------===//
 // ARM Processors supported.
@@ -105,7 +108,8 @@ def : ProcNoItin<"arm1156t2f-s",    [ArchV6T2, FeatureThumb2, FeatureVFP2]>;
 
 // V7 Processors.
 def : Processor<"cortex-a8",        CortexA8Itineraries,
-                [ArchV7A, FeatureThumb2, FeatureNEON, FeatureNEONFP]>;
+                [ArchV7A, FeatureThumb2, FeatureNEON, FeatureNEONFP,
+                 FeaturePostRASched]>;
 def : ProcNoItin<"cortex-a9",       [ArchV7A, FeatureThumb2, FeatureNEON]>;
 
 //===----------------------------------------------------------------------===//

llvm/lib/Target/ARM/ARMSubtarget.cpp

@@ -29,6 +29,7 @@ ARMSubtarget::ARMSubtarget(const std::string &TT, const std::string &FS,
   , UseNEONForSinglePrecisionFP(false)
   , IsThumb(isThumb)
   , ThumbMode(Thumb1)
+  , PostRAScheduler(false)
   , IsR9Reserved(ReserveR9)
   , stackAlignment(4)
   , CPUString("generic")

llvm/lib/Target/ARM/ARMSubtarget.h

@@ -55,6 +55,9 @@ protected:
   /// ThumbMode - Indicates supported Thumb version.
   ThumbTypeEnum ThumbMode;
 
+  /// PostRAScheduler - True if using post-register-allocation scheduler.
+  bool PostRAScheduler;
+
   /// IsR9Reserved - True if R9 is a not available as general purpose register.
   bool IsR9Reserved;
 
@@ -123,6 +126,10 @@ protected:
 
   const std::string & getCPUString() const { return CPUString; }
   
+  /// enablePostRAScheduler - From TargetSubtarget, return true to
+  /// enable post-RA scheduler.
+  bool enablePostRAScheduler() const { return PostRAScheduler; }
+
   /// getInstrItins - Return the instruction itineraies based on subtarget
   /// selection.
   const InstrItineraryData &getInstrItineraryData() const { return InstrItins; }

llvm/test/CodeGen/ARM/2009-08-21-PostRAKill.ll

@@ -1,4 +1,4 @@
-; RUN: llc < %s -march=arm -mattr=+vfp2 -mcpu=cortex-a8 -post-RA-scheduler
+; RUN: llc < %s -march=arm -mattr=+vfp2,+postrasched -mcpu=cortex-a8
 
 ; ModuleID = '<stdin>'
 target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:32-f32:32:32-f64:32:32-v64:64:64-v128:128:128-a0:0:64"

llvm/test/CodeGen/ARM/2009-08-21-PostRAKill2.ll

@@ -1,4 +1,4 @@
-; RUN: llc < %s -asm-verbose=false -O3 -relocation-model=pic -disable-fp-elim -mtriple=thumbv7-apple-darwin -mcpu=cortex-a8 -post-RA-scheduler
+; RUN: llc < %s -asm-verbose=false -O3 -relocation-model=pic -disable-fp-elim -mtriple=thumbv7-apple-darwin -mcpu=cortex-a8 -mattr=+postrasched
 
 ; ModuleID = '<stdin>'
 target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:32-f32:32:32-f64:32:32-v64:64:64-v128:128:128-a0:0:64"

llvm/test/CodeGen/ARM/2009-08-21-PostRAKill3.ll

@@ -1,4 +1,4 @@
-; RUN: llc < %s -asm-verbose=false -O3 -relocation-model=pic -disable-fp-elim -mtriple=thumbv7-apple-darwin -mcpu=cortex-a8 -post-RA-scheduler
+; RUN: llc < %s -asm-verbose=false -O3 -relocation-model=pic -disable-fp-elim -mtriple=thumbv7-apple-darwin -mcpu=cortex-a8 -mattr=+postrasched
 
 ; ModuleID = '<stdin>'
 target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:32-f32:32:32-f64:32:32-v64:64:64-v128:128:128-a0:0:64"

llvm/test/CodeGen/ARM/2009-08-21-PostRAKill4.ll

@@ -1,4 +1,4 @@
-; RUN: llc < %s -asm-verbose=false -O3 -relocation-model=pic -disable-fp-elim -mtriple=thumbv7-apple-darwin -mcpu=cortex-a8 -post-RA-scheduler
+; RUN: llc < %s -asm-verbose=false -O3 -relocation-model=pic -disable-fp-elim -mtriple=thumbv7-apple-darwin -mcpu=cortex-a8 -mattr=+postrasched
 
 ; ModuleID = '<stdin>'
 target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:32-f32:32:32-f64:32:32-v64:64:64-v128:128:128-a0:0:64"

llvm/test/CodeGen/ARM/2009-09-01-PostRAProlog.ll

@@ -1,5 +1,5 @@
 ; XFAIL: *
-; RUN: llvm-as < %s | llc -asm-verbose=false -O3 -relocation-model=pic -disable-fp-elim -mtriple=thumbv7-apple-darwin -mcpu=cortex-a8 -post-RA-scheduler | FileCheck %s
+; RUN: llvm-as < %s | llc -asm-verbose=false -O3 -relocation-model=pic -disable-fp-elim -mtriple=thumbv7-apple-darwin -mcpu=cortex-a8 -mattr=+postrasched | FileCheck %s
 
 
 ; ModuleID = '<stdin>'