[ARM] Do not test for CPUs, use SubtargetFeatures. Also remove 1 flag

[rovka]

Summary

This is a follow-up for r273544.

The end goal is to get rid of the isSwift / isCortexXY / isWhatever methods.

This commit also removes a command line flag that isn't used in any of the tests:
check-vmlx-hazards. It can be replaced easily with the mattr mechanism, since
this is now a subtarget feature.

I have mixed feelings about the FeatureExpandMLx. I'm not sure I understand the
history there. In particular, in the past MLx expansion was enabled for
subtargets with hasVFP2(), until r129775 [1] switched from that to isCortexA9,
without too much justification.

In spite of that, the code performing MLx expansion still contains calls to
isSwift/isLikeA9, although the results of those are pretty clear given that
we're only enabling it for the A9. Should we try to enable it for other targets?
If so, do we want to have a feature for it (as it is in this patch), or should
we control it based on other features? Is FeatureHasVMLxHazards a good candidate
for this? Any thoughts appreciated, thanks.

[1] http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20110418/119725.html

[rengolin]

In particular, in the past MLx expansion was enabled for
subtargets with hasVFP2(), until r129775 [1] switched from that to isCortexA9,
without too much justification.

Now I remember this patch! This is about the VMLA+VMLA having a huge pipeline stall on A9, so VMLA+VMUL+VADD is preferred.

This is *exactly* the semantics of VMLxHazard and can be encoded in the way you do in this patch.

I remember this being a problem in A9, it may very well be present in A8 (though I don't remember exactly). I'm just surprised that A7 has the hazard as well. I'd have thought that, if A15 has fixed it, A7 would have carried the fix with it.

Regardless, this is a conversation for another patch. I think your fix is correct and it retains the current behaviour.

cheers,
--renato

[rovka]

Ping? Can we commit this as-is?
If we want to switch to using HasVMLxHazards instead of ExpandMLx, I'm going to need some help with testing (at least on Swift), and it won't be NFC anymore (at least not on Swift, A7 and A8). I can bring it up on llvm-dev if there's interest.

[MatzeB]

LGTM

llvm/trunk/lib/Target/ARM/ARM.td

 def FeatureSlowLoadDSubreg : SubtargetFeature<"slow-load-D-subreg",
                                               "SlowLoadDSubregister", "true",
                                               "Loading into D subregs is slow">;
 // Some targets (e.g. Cortex-A15) never want VMOVS to be widened to VMOVD.
 def FeatureDontWidenVMOVS : SubtargetFeature<"dont-widen-vmovs",
                                              "DontWidenVMOVS", "true",
                                              "Don't widen VMOVS to VMOVD">;
 
+// Whether or not it is profitable to expand VFP/NEON MLA/MLS instructions.
+def FeatureExpandMLx : SubtargetFeature<"expand-fp-mlx", "ExpandMLx", "true",
+                                        "Expand VFP/NEON MLA/MLS instructions">;
+
+// Some targets have special RAW hazards for VFP/NEON VMLA/VMLS.
+def FeatureHasVMLxHazards : SubtargetFeature<"vmlx-hazards", "HasVMLxHazards",
+                                             "true", "Has VMLx hazards">;
+
 // Some targets (e.g. Cortex-A9) want to convert VMOVRS, VMOVSR and VMOVS from
 // VFP to NEON, as an execution domain optimization.
 def FeatureNEONForFPMovs : SubtargetFeature<"neon-fpmovs", "UseNEONForFPMovs",
                               "true", "Convert VMOVSR, VMOVRS, VMOVS to NEON">;
 
 // Some processors benefit from using NEON instructions for scalar
 // single-precision FP operations. This affects instruction selection and should
 // only be enabled if the handling of denormals is not important.
def : ProcessorModel<"cortex-a5",   CortexA8Model,      [ARMv7a, ProcA5,
                                                          FeatureT2XtPk,
                                                          FeatureMP,
                                                          FeatureVFP4]>;
 
 def : ProcessorModel<"cortex-a7",   CortexA8Model,      [ARMv7a, ProcA7,
                                                          FeatureHasRetAddrStack,
                                                          FeatureTrustZone,
                                                          FeatureSlowFPBrcc,
+                                                         FeatureHasVMLxHazards,
                                                          FeatureHasSlowFPVMLx,
                                                          FeatureVMLxForwarding,
                                                          FeatureT2XtPk,
                                                          FeatureMP,
                                                          FeatureVFP4,
                                                          FeatureHWDiv,
                                                          FeatureHWDivARM,
                                                          FeatureVirtualization]>;
 
 def : ProcessorModel<"cortex-a8",   CortexA8Model,      [ARMv7a, ProcA8,
                                                          FeatureHasRetAddrStack,
                                                          FeatureNonpipelinedVFP,
                                                          FeatureTrustZone,
                                                          FeatureSlowFPBrcc,
+                                                         FeatureHasVMLxHazards,
                                                          FeatureHasSlowFPVMLx,
                                                          FeatureVMLxForwarding,
                                                          FeatureT2XtPk]>;
 
 def : ProcessorModel<"cortex-a9",   CortexA9Model,      [ARMv7a, ProcA9,
                                                          FeatureHasRetAddrStack,
                                                          FeatureTrustZone,
+                                                         FeatureHasVMLxHazards,
                                                          FeatureVMLxForwarding,
                                                          FeatureT2XtPk,
                                                          FeatureFP16,
                                                          FeatureAvoidPartialCPSR,
+                                                         FeatureExpandMLx,
                                                          FeaturePreferVMOVSR,
                                                          FeatureMuxedUnits,
                                                          FeatureNEONForFPMovs,
                                                          FeatureCheckVLDnAlign,
                                                          FeatureMP]>;
 
 // FIXME: A12 has currently the same Schedule model as A9
 def : ProcessorModel<"cortex-a12",  CortexA9Model,      [ARMv7a, ProcA12,
def : ProcessorModel<"swift",       SwiftModel,         [ARMv7a, ProcSwift,
                                                          FeatureT2XtPk,
                                                          FeatureVFP4,
                                                          FeatureMP,
                                                          FeatureHWDiv,
                                                          FeatureHWDivARM,
                                                          FeatureAvoidPartialCPSR,
                                                          FeatureAvoidMOVsShOp,
                                                          FeatureHasSlowFPVMLx,
+                                                         FeatureHasVMLxHazards,
                                                          FeatureProfUnpredicate,
                                                          FeaturePrefISHSTBarrier,
                                                          FeatureSlowOddRegister,
                                                          FeatureSlowLoadDSubreg,
                                                          FeatureSlowVGETLNi32,
                                                          FeatureSlowVDUP32]>;
 
 // FIXME: R4 has currently the same ProcessorModel as A8.

llvm/trunk/lib/Target/ARM/ARMISelDAGToDAG.cpp

 
 #define DEBUG_TYPE "arm-isel"
 
 static cl::opt<bool>
 DisableShifterOp("disable-shifter-op", cl::Hidden,
   cl::desc("Disable isel of shifter-op"),
   cl::init(false));
 
-static cl::opt<bool>
-CheckVMLxHazard("check-vmlx-hazard", cl::Hidden,
-  cl::desc("Check fp vmla / vmls hazard at isel time"),
-  cl::init(true));
-
 //===--------------------------------------------------------------------===//
 /// ARMDAGToDAGISel - ARM specific code to select ARM machine
 /// instructions for SelectionDAG operations.
 ///
 namespace {
 
 enum AddrMode2Type {
   AM2_BASE, // Simple AM2 (+-imm12)

 
 /// hasNoVMLxHazardUse - Return true if it's desirable to select a FP MLA / MLS
 /// node. VFP / NEON fp VMLA / VMLS instructions have special RAW hazards (at
 /// least on current ARM implementations) which should be avoidded.
 bool ARMDAGToDAGISel::hasNoVMLxHazardUse(SDNode *N) const {
   if (OptLevel == CodeGenOpt::None)
     return true;
 
-  if (!CheckVMLxHazard)
-    return true;
-
-  if (!Subtarget->isCortexA7() && !Subtarget->isCortexA8() &&
-      !Subtarget->isCortexA9() && !Subtarget->isSwift())
+  if (!Subtarget->hasVMLxHazards())
     return true;
 
   if (!N->hasOneUse())
     return false;
 
   SDNode *Use = *N->use_begin();
   if (Use->getOpcode() == ISD::CopyToReg)
     return true;

llvm/trunk/lib/Target/ARM/ARMSubtarget.h

protected:
   bool SlowLoadDSubregister = false;
 
   /// If true, the AGU and NEON/FPU units are multiplexed.
   bool HasMuxedUnits = false;
 
   /// If true, VMOVS will never be widened to VMOVD
   bool DontWidenVMOVS = false;
 
+  /// If true, run the MLx expansion pass.
+  bool ExpandMLx = false;
+
+  /// If true, VFP/NEON VMLA/VMLS have special RAW hazards.
+  bool HasVMLxHazards = false;
+
   /// If true, VMOVRS, VMOVSR and VMOVS will be converted from VFP to NEON.
   bool UseNEONForFPMovs = false;
 
   /// If true, VLDn instructions take an extra cycle for unaligned accesses.
   bool CheckVLDnAlign = false;
 
   /// If true, VFP instructions are not pipelined.
   bool NonpipelinedVFP = false;
public:
   bool hasTrustZone() const { return HasTrustZone; }
   bool has8MSecExt() const { return Has8MSecExt; }
   bool hasZeroCycleZeroing() const { return HasZeroCycleZeroing; }
   bool isProfitableToUnpredicate() const { return IsProfitableToUnpredicate; }
   bool hasSlowVGETLNi32() const { return HasSlowVGETLNi32; }
   bool hasSlowVDUP32() const { return HasSlowVDUP32; }
   bool preferVMOVSR() const { return PreferVMOVSR; }
   bool preferISHSTBarriers() const { return PreferISHST; }
+  bool expandMLx() const { return ExpandMLx; }
+  bool hasVMLxHazards() const { return HasVMLxHazards; }
   bool hasSlowOddRegister() const { return SlowOddRegister; }
   bool hasSlowLoadDSubregister() const { return SlowLoadDSubregister; }
   bool hasMuxedUnits() const { return HasMuxedUnits; }
   bool dontWidenVMOVS() const { return DontWidenVMOVS; }
   bool useNEONForFPMovs() const { return UseNEONForFPMovs; }
   bool checkVLDnAccessAlignment() const { return CheckVLDnAlign; }
   bool nonpipelinedVFP() const { return NonpipelinedVFP; }
   bool prefers32BitThumb() const { return Pref32BitThumb; }

llvm/trunk/lib/Target/ARM/MLxExpansionPass.cpp

 bool MLxExpansion::runOnMachineFunction(MachineFunction &Fn) {
   if (skipFunction(*Fn.getFunction()))
     return false;
 
   TII = static_cast<const ARMBaseInstrInfo *>(Fn.getSubtarget().getInstrInfo());
   TRI = Fn.getSubtarget().getRegisterInfo();
   MRI = &Fn.getRegInfo();
   const ARMSubtarget *STI = &Fn.getSubtarget<ARMSubtarget>();
-  // Only run this for CortexA9.
-  if (!STI->isCortexA9())
+  if (!STI->expandMLx())
     return false;
   isLikeA9 = STI->isLikeA9() || STI->isSwift();
   isSwift = STI->isSwift();
 
   bool Modified = false;
   for (MachineBasicBlock &MBB : Fn)
     Modified |= ExpandFPMLxInstructions(MBB);
 
   return Modified;
 }
 
 FunctionPass *llvm::createMLxExpansionPass() {
   return new MLxExpansion();
 }