This patch combines several changes from Evan Cheng for rdar://8659675.

Making use of VFP / NEON floating point multiply-accumulate / subtraction is
difficult on current ARM implementations for a few reasons.
1. Even though a single vmla has latency that is one cycle shorter than a pair
   of vmul + vadd, a RAW hazard during the first (4? on Cortex-a8) can cause
   additional pipeline stall. So it's frequently better to single codegen
   vmul + vadd.
2. A vmla folowed by a vmul, vmadd, or vsub causes the second fp instruction to
   stall for 4 cycles. We need to schedule them apart.
3. A vmla followed vmla is a special case. Obvious issuing back to back RAW
   vmla + vmla is very bad. But this isn't ideal either:
     vmul
     vadd
     vmla
   Instead, we want to expand the second vmla:
     vmla
     vmul
     vadd
   Even with the 4 cycle vmul stall, the second sequence is still 2 cycles
   faster.

Up to now, isel simply avoid codegen'ing fp vmla / vmls. This works well enough
but it isn't the optimial solution. This patch attempts to make it possible to
use vmla / vmls in cases where it is profitable.

A. Add missing isel predicates which cause vmla to be codegen'ed.
B. Make sure the fmul in (fadd (fmul)) has a single use. We don't want to
   compute a fmul and a fmla.
C. Add additional isel checks for vmla, avoid cases where vmla is feeding into
   fp instructions (except for the #3 exceptional case).
D. Add ARM hazard recognizer to model the vmla / vmls hazards.
E. Add a special pre-regalloc case to expand vmla / vmls when it's likely the
   vmla / vmls will trigger one of the special hazards.

Enable these fp vmlx codegen changes for Cortex-A9.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@129775 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/ARM/ARM.td

@@ -118,7 +118,7 @@ def ProcA8      : SubtargetFeature<"a8", "ARMProcFamily", "CortexA8",
                                     FeatureT2XtPk]>;
 def ProcA9      : SubtargetFeature<"a9", "ARMProcFamily", "CortexA9",
                                    "Cortex-A9 ARM processors",
-                                   [FeatureHasSlowFPVMLx, FeatureVMLxForwarding,
+                                   [FeatureVMLxForwarding,
                                     FeatureT2XtPk, FeatureFP16,
                                     FeatureAvoidPartialCPSR]>;
 

lib/Target/ARM/ARMHazardRecognizer.cpp

@@ -49,6 +49,8 @@ ARMHazardRecognizer::getHazardType(SUnit *SU, int Stalls) {
       const TargetInstrDesc &LastTID = LastMI->getDesc();
       // Skip over one non-VFP / NEON instruction.
       if (!LastTID.isBarrier() &&
+          // On A9, AGU and NEON/FPU are muxed.
+          !(STI.isCortexA9() && (LastTID.mayLoad() || LastTID.mayStore())) &&
           (LastTID.TSFlags & ARMII::DomainMask) == ARMII::DomainGeneral) {
         MachineBasicBlock::iterator I = LastMI;
         if (I != LastMI->getParent()->begin()) {

lib/Target/ARM/ARMISelDAGToDAG.cpp

@@ -45,7 +45,7 @@ DisableShifterOp("disable-shifter-op", cl::Hidden,
 static cl::opt<bool>
 CheckVMLxHazard("check-vmlx-hazard", cl::Hidden,
   cl::desc("Check fp vmla / vmls hazard at isel time"),
-  cl::init(false));
+  cl::init(true));
 
 //===--------------------------------------------------------------------===//
 /// ARMDAGToDAGISel - ARM specific code to select ARM machine

lib/Target/ARM/ARMTargetMachine.cpp

@@ -22,8 +22,6 @@
 #include "llvm/Target/TargetRegistry.h"
 using namespace llvm;
 
-static cl::opt<bool>ExpandMLx("expand-fp-mlx", cl::init(false), cl::Hidden);
-
 static MCAsmInfo *createMCAsmInfo(const Target &T, StringRef TT) {
   Triple TheTriple(TT);
   switch (TheTriple.getOS()) {
@@ -148,8 +146,7 @@ bool ARMBaseTargetMachine::addPreRegAlloc(PassManagerBase &PM,
   // FIXME: temporarily disabling load / store optimization pass for Thumb1.
   if (OptLevel != CodeGenOpt::None && !Subtarget.isThumb1Only())
     PM.add(createARMLoadStoreOptimizationPass(true));
-  if (ExpandMLx &&
-      OptLevel != CodeGenOpt::None && Subtarget.hasVFP2())
+  if (OptLevel != CodeGenOpt::None && Subtarget.isCortexA9())
     PM.add(createMLxExpansionPass());
 
   return true;

lib/Target/ARM/MLxExpansionPass.cpp

@@ -15,11 +15,13 @@
 #define DEBUG_TYPE "mlx-expansion"
 #include "ARM.h"
 #include "ARMBaseInstrInfo.h"
+#include "ARMSubtarget.h"
 #include "llvm/CodeGen/MachineInstr.h"
 #include "llvm/CodeGen/MachineInstrBuilder.h"
 #include "llvm/CodeGen/MachineFunctionPass.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
 #include "llvm/Target/TargetRegisterInfo.h"
+#include "llvm/ADT/SmallPtrSet.h"
 #include "llvm/ADT/Statistic.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/Debug.h"
@@ -49,15 +51,17 @@ namespace {
     const TargetRegisterInfo *TRI;
     MachineRegisterInfo *MRI;
 
+    bool isA9;
     unsigned MIIdx;
     MachineInstr* LastMIs[4];
+    SmallPtrSet<MachineInstr*, 4> IgnoreStall;
 
     void clearStack();
     void pushStack(MachineInstr *MI);
     MachineInstr *getAccDefMI(MachineInstr *MI) const;
     unsigned getDefReg(MachineInstr *MI) const;
     bool hasRAWHazard(unsigned Reg, MachineInstr *MI) const;
-    bool FindMLxHazard(MachineInstr *MI) const;
+    bool FindMLxHazard(MachineInstr *MI);
     void ExpandFPMLxInstruction(MachineBasicBlock &MBB, MachineInstr *MI,
                                 unsigned MulOpc, unsigned AddSubOpc,
                                 bool NegAcc, bool HasLane);
@@ -146,7 +150,7 @@ bool MLxExpansion::hasRAWHazard(unsigned Reg, MachineInstr *MI) const {
 }
 
 
-bool MLxExpansion::FindMLxHazard(MachineInstr *MI) const {
+bool MLxExpansion::FindMLxHazard(MachineInstr *MI) {
   if (NumExpand >= ExpandLimit)
     return false;
 
@@ -154,7 +158,7 @@ bool MLxExpansion::FindMLxHazard(MachineInstr *MI) const {
     return true;
 
   MachineInstr *DefMI = getAccDefMI(MI);
-  if (TII->isFpMLxInstruction(DefMI->getOpcode()))
+  if (TII->isFpMLxInstruction(DefMI->getOpcode())) {
     // r0 = vmla
     // r3 = vmla r0, r1, r2
     // takes 16 - 17 cycles
@@ -163,24 +167,33 @@ bool MLxExpansion::FindMLxHazard(MachineInstr *MI) const {
     // r4 = vmul r1, r2
     // r3 = vadd r0, r4
     // takes about 14 - 15 cycles even with vmul stalling for 4 cycles.
+    IgnoreStall.insert(DefMI);
     return true;
+  }
+
+  if (IgnoreStall.count(MI))
+    return false;
 
   // If a VMLA.F is followed by an VADD.F or VMUL.F with no RAW hazard, the
   // VADD.F or VMUL.F will stall 4 cycles before issue. The 4 cycle stall
   // preserves the in-order retirement of the instructions.
   // Look at the next few instructions, if *most* of them can cause hazards,
   // then the scheduler can't *fix* this, we'd better break up the VMLA.
+  unsigned Limit1 = isA9 ? 1 : 4;
+  unsigned Limit2 = isA9 ? 1 : 4;
   for (unsigned i = 1; i <= 4; ++i) {
     int Idx = ((int)MIIdx - i + 4) % 4;
     MachineInstr *NextMI = LastMIs[Idx];
     if (!NextMI)
       continue;
 
-    if (TII->canCauseFpMLxStall(NextMI->getOpcode()))
-      return true;
+    if (TII->canCauseFpMLxStall(NextMI->getOpcode())) {
+      if (i <= Limit1)
+        return true;
+    }
 
     // Look for VMLx RAW hazard.
-    if (hasRAWHazard(getDefReg(MI), NextMI))
+    if (i <= Limit2 && hasRAWHazard(getDefReg(MI), NextMI))
       return true;
   }
 
@@ -248,6 +261,7 @@ bool MLxExpansion::ExpandFPMLxInstructions(MachineBasicBlock &MBB) {
   bool Changed = false;
 
   clearStack();
+  IgnoreStall.clear();
 
   unsigned Skip = 0;
   MachineBasicBlock::reverse_iterator MII = MBB.rbegin(), E = MBB.rend();
@@ -299,6 +313,8 @@ bool MLxExpansion::runOnMachineFunction(MachineFunction &Fn) {
   TII = static_cast<const ARMBaseInstrInfo*>(Fn.getTarget().getInstrInfo());
   TRI = Fn.getTarget().getRegisterInfo();
   MRI = &Fn.getRegInfo();
+  const ARMSubtarget *STI = &Fn.getTarget().getSubtarget<ARMSubtarget>();
+  isA9 = STI->isCortexA9();
 
   bool Modified = false;
   for (MachineFunction::iterator MFI = Fn.begin(), E = Fn.end(); MFI != E;

test/CodeGen/ARM/fmacs.ll

@@ -1,6 +1,8 @@
 ; RUN: llc < %s -march=arm -mattr=+vfp2 | FileCheck %s -check-prefix=VFP2
 ; RUN: llc < %s -march=arm -mattr=+neon | FileCheck %s -check-prefix=NEON
 ; RUN: llc < %s -march=arm -mcpu=cortex-a8 | FileCheck %s -check-prefix=A8
+; RUN: llc < %s -march=arm -mcpu=cortex-a9 | FileCheck %s -check-prefix=A9
+; RUN: llc < %s -mtriple=arm-linux-gnueabi -mcpu=cortex-a9 -float-abi=hard | FileCheck %s -check-prefix=HARD
 
 define float @t1(float %acc, float %a, float %b) {
 entry:
@@ -49,3 +51,54 @@ entry:
         %1 = fadd float %0, %acc
 	ret float %1
 }
+
+; It's possible to make use of fp vmla / vmls on Cortex-A9.
+; rdar://8659675
+define void @t4(float %acc1, float %a, float %b, float %acc2, float %c, float* %P1, float* %P2) {
+entry:
+; A8: t4:
+; A8: vmul.f32
+; A8: vmul.f32
+; A8: vadd.f32
+; A8: vadd.f32
+
+; Two vmla with now RAW hazard
+; A9: t4:
+; A9: vmla.f32
+; A9: vmla.f32
+
+; HARD: t4:
+; HARD: vmla.f32 s0, s1, s2
+; HARD: vmla.f32 s3, s1, s4
+  %0 = fmul float %a, %b
+  %1 = fadd float %acc1, %0
+  %2 = fmul float %a, %c
+  %3 = fadd float %acc2, %2
+  store float %1, float* %P1
+  store float %3, float* %P2
+  ret void
+}
+
+define float @t5(float %a, float %b, float %c, float %d, float %e) {
+entry:
+; A8: t5:
+; A8: vmul.f32
+; A8: vmul.f32
+; A8: vadd.f32
+; A8: vadd.f32
+
+; A9: t5:
+; A9: vmla.f32
+; A9: vmul.f32
+; A9: vadd.f32
+
+; HARD: t5:
+; HARD: vmla.f32 s4, s0, s1
+; HARD: vmul.f32 s0, s2, s3
+; HARD: vadd.f32 s0, s4, s0
+  %0 = fmul float %a, %b
+  %1 = fadd float %e, %0
+  %2 = fmul float %c, %d
+  %3 = fadd float %1, %2
+  ret float %3
+}