Avoid write-after-write issue hazards for Cortex-A9.

Add a avoidWriteAfterWrite() target hook to identify register classes that
suffer from write-after-write hazards. For those register classes, try to avoid
writing the same register in two consecutive instructions.

This is currently disabled by default.  We should not spill to avoid hazards!
The command line flag -avoid-waw-hazard can be used to enable waw avoidance.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@129772 91177308-0d34-0410-b5e6-96231b3b80d8

9 files changed, 65 insertions, 18 deletions

diff --git a/include/llvm/Target/TargetRegisterInfo.h b/include/llvm/Target/TargetRegisterInfo.h
index d06cdc4e8e..0fa3b38561 100644
--- a/include/llvm/Target/TargetRegisterInfo.h
+++ b/include/llvm/Target/TargetRegisterInfo.h
@@ -624,6 +624,14 @@ public:
     return 0;
   }
 
+  /// avoidWriteAfterWrite - Return true if the register allocator should avoid
+  /// writing a register from RC in two consecutive instructions.
+  /// This can avoid pipeline stalls on certain architectures.
+  /// It does cause increased register pressure, though.
+  virtual bool avoidWriteAfterWrite(const TargetRegisterClass *RC) const {
+    return false;
+  }
+
   /// UpdateRegAllocHint - A callback to allow target a chance to update
   /// register allocation hints when a register is "changed" (e.g. coalesced)
   /// to another register. e.g. On ARM, some virtual registers should target
diff --git a/lib/CodeGen/RegAllocLinearScan.cpp b/lib/CodeGen/RegAllocLinearScan.cpp
index b8cb5a7c92..b91f312b72 100644
--- a/lib/CodeGen/RegAllocLinearScan.cpp
+++ b/lib/CodeGen/RegAllocLinearScan.cpp
@@ -67,6 +67,11 @@ TrivCoalesceEnds("trivial-coalesce-ends",
                   cl::desc("Attempt trivial coalescing of interval ends"),
                   cl::init(false), cl::Hidden);
 
+static cl::opt<bool>
+AvoidWAWHazard("avoid-waw-hazard",
+               cl::desc("Avoid write-write hazards for some register classes"),
+               cl::init(false), cl::Hidden);
+
 static RegisterRegAlloc
 linearscanRegAlloc("linearscan", "linear scan register allocator",
                    createLinearScanRegisterAllocator);
@@ -110,6 +115,7 @@ namespace {
       if (NumRecentlyUsedRegs > 0)
         RecentRegs.resize(NumRecentlyUsedRegs, 0);
       RecentNext = RecentRegs.begin();
+      avoidWAW_ = 0;
     }
 
     typedef std::pair<LiveInterval*, LiveInterval::iterator> IntervalPtr;
@@ -180,6 +186,9 @@ namespace {
     SmallVector<unsigned, 4> RecentRegs;
     SmallVector<unsigned, 4>::iterator RecentNext;
 
+    // Last write-after-write register written.
+    unsigned avoidWAW_;
+
     // Record that we just picked this register.
     void recordRecentlyUsed(unsigned reg) {
       assert(reg != 0 && "Recently used register is NOREG!");
@@ -227,8 +236,8 @@ namespace {
 
     // Determine if we skip this register due to its being recently used.
     bool isRecentlyUsed(unsigned reg) const {
-      return std::find(RecentRegs.begin(), RecentRegs.end(), reg) !=
-             RecentRegs.end();
+      return reg == avoidWAW_ ||
+       std::find(RecentRegs.begin(), RecentRegs.end(), reg) != RecentRegs.end();
     }
 
   private:
@@ -1116,6 +1125,12 @@ void RALinScan::assignRegOrStackSlotAtInterval(LiveInterval* cur) {
     active_.push_back(std::make_pair(cur, cur->begin()));
     handled_.push_back(cur);
 
+    // Remember physReg for avoiding a write-after-write hazard in the next
+    // instruction.
+    if (AvoidWAWHazard &&
+        tri_->avoidWriteAfterWrite(mri_->getRegClass(cur->reg)))
+      avoidWAW_ = physReg;
+
     // "Upgrade" the physical register since it has been allocated.
     UpgradeRegister(physReg);
     if (LiveInterval *NextReloadLI = hasNextReloadInterval(cur)) {
@@ -1446,7 +1461,7 @@ unsigned RALinScan::getFreePhysReg(LiveInterval* cur,
     if (reservedRegs_.test(Reg))
       continue;
     // Skip recently allocated registers.
-    if (isRegAvail(Reg) && !isRecentlyUsed(Reg)) {
+    if (isRegAvail(Reg) && (!SkipDGRegs || !isRecentlyUsed(Reg))) {
       FreeReg = Reg;
       if (FreeReg < inactiveCounts.size())
         FreeRegInactiveCount = inactiveCounts[FreeReg];
@@ -1477,7 +1492,8 @@ unsigned RALinScan::getFreePhysReg(LiveInterval* cur,
     if (reservedRegs_.test(Reg))
       continue;
     if (isRegAvail(Reg) && Reg < inactiveCounts.size() &&
-        FreeRegInactiveCount < inactiveCounts[Reg] && !isRecentlyUsed(Reg)) {
+        FreeRegInactiveCount < inactiveCounts[Reg] &&
+        (!SkipDGRegs || !isRecentlyUsed(Reg))) {
       FreeReg = Reg;
       FreeRegInactiveCount = inactiveCounts[Reg];
       if (FreeRegInactiveCount == MaxInactiveCount)
@@ -1528,12 +1544,10 @@ unsigned RALinScan::getFreePhysReg(LiveInterval *cur) {
       return Preference;
   }
 
-  if (!DowngradedRegs.empty()) {
-    unsigned FreeReg = getFreePhysReg(cur, RC, MaxInactiveCount, inactiveCounts,
-                                      true);
-    if (FreeReg)
-      return FreeReg;
-  }
+  unsigned FreeReg = getFreePhysReg(cur, RC, MaxInactiveCount, inactiveCounts,
+                                    true);
+  if (FreeReg)
+    return FreeReg;
   return getFreePhysReg(cur, RC, MaxInactiveCount, inactiveCounts, false);
 }
 
diff --git a/lib/Target/ARM/ARMBaseRegisterInfo.cpp b/lib/Target/ARM/ARMBaseRegisterInfo.cpp
index 3d1eaf0891..6eb9002df8 100644
--- a/lib/Target/ARM/ARMBaseRegisterInfo.cpp
+++ b/lib/Target/ARM/ARMBaseRegisterInfo.cpp
@@ -554,6 +554,29 @@ ARMBaseRegisterInfo::UpdateRegAllocHint(unsigned Reg, unsigned NewReg,
   }
 }
 
+bool
+ARMBaseRegisterInfo::avoidWriteAfterWrite(const TargetRegisterClass *RC) const {
+  // CortexA9 has a Write-after-write hazard for NEON registers.
+  if (!STI.isCortexA9())
+    return false;
+
+  switch (RC->getID()) {
+  case ARM::DPRRegClassID:
+  case ARM::DPR_8RegClassID:
+  case ARM::DPR_VFP2RegClassID:
+  case ARM::QPRRegClassID:
+  case ARM::QPR_8RegClassID:
+  case ARM::QPR_VFP2RegClassID:
+  case ARM::SPRRegClassID:
+  case ARM::SPR_8RegClassID:
+    // Avoid reusing S, D, and Q registers.
+    // Don't increase register pressure for QQ and QQQQ.
+    return true;
+  default:
+    return false;
+  }
+}
+
 bool ARMBaseRegisterInfo::hasBasePointer(const MachineFunction &MF) const {
   const MachineFrameInfo *MFI = MF.getFrameInfo();
   const ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
diff --git a/lib/Target/ARM/ARMBaseRegisterInfo.h b/lib/Target/ARM/ARMBaseRegisterInfo.h
index 0507396f2c..480892ed3e 100644
--- a/lib/Target/ARM/ARMBaseRegisterInfo.h
+++ b/lib/Target/ARM/ARMBaseRegisterInfo.h
@@ -142,6 +142,8 @@ public:
   void UpdateRegAllocHint(unsigned Reg, unsigned NewReg,
                           MachineFunction &MF) const;
 
+  virtual bool avoidWriteAfterWrite(const TargetRegisterClass *RC) const;
+
   bool hasBasePointer(const MachineFunction &MF) const;
 
   bool canRealignStack(const MachineFunction &MF) const;
diff --git a/test/CodeGen/ARM/fabss.ll b/test/CodeGen/ARM/fabss.ll
index f03282bdab..51efe51bf1 100644
--- a/test/CodeGen/ARM/fabss.ll
+++ b/test/CodeGen/ARM/fabss.ll
@@ -24,4 +24,4 @@ declare float @fabsf(float)
 ; CORTEXA8: test:
 ; CORTEXA8: 	vabs.f32	d1, d1
 ; CORTEXA9: test:
-; CORTEXA9: 	vabs.f32	s1, s1
+; CORTEXA9: 	vabs.f32	s{{.}}, s{{.}}
diff --git a/test/CodeGen/ARM/fadds.ll b/test/CodeGen/ARM/fadds.ll
index 749690e98d..e35103c045 100644
--- a/test/CodeGen/ARM/fadds.ll
+++ b/test/CodeGen/ARM/fadds.ll
@@ -20,4 +20,4 @@ entry:
 ; CORTEXA8: test:
 ; CORTEXA8: 	vadd.f32	d0, d1, d0
 ; CORTEXA9: test:
-; CORTEXA9: 	vadd.f32	s0, s1, s0
+; CORTEXA9: 	vadd.f32	s{{.}}, s{{.}}, s{{.}}
diff --git a/test/CodeGen/ARM/fdivs.ll b/test/CodeGen/ARM/fdivs.ll
index 0c31495792..31c1ca9405 100644
--- a/test/CodeGen/ARM/fdivs.ll
+++ b/test/CodeGen/ARM/fdivs.ll
@@ -20,4 +20,4 @@ entry:
 ; CORTEXA8: test:
 ; CORTEXA8: 	vdiv.f32	s0, s1, s0
 ; CORTEXA9: test:
-; CORTEXA9: 	vdiv.f32	s0, s1, s0
+; CORTEXA9: 	vdiv.f32	s{{.}}, s{{.}}, s{{.}}
diff --git a/test/CodeGen/ARM/fmuls.ll b/test/CodeGen/ARM/fmuls.ll
index ef4e3e5281..bc118b8cb2 100644
--- a/test/CodeGen/ARM/fmuls.ll
+++ b/test/CodeGen/ARM/fmuls.ll
@@ -20,4 +20,4 @@ entry:
 ; CORTEXA8: test:
 ; CORTEXA8: 	vmul.f32	d0, d1, d0
 ; CORTEXA9: test:
-; CORTEXA9: 	vmul.f32	s0, s1, s0
+; CORTEXA9: 	vmul.f32	s{{.}}, s{{.}}, s{{.}}
diff --git a/test/CodeGen/ARM/fp_convert.ll b/test/CodeGen/ARM/fp_convert.ll
index 1ef9f7f321..86c06f1ddd 100644
--- a/test/CodeGen/ARM/fp_convert.ll
+++ b/test/CodeGen/ARM/fp_convert.ll
@@ -5,7 +5,7 @@
 
 define i32 @test1(float %a, float %b) {
 ; VFP2: test1:
-; VFP2: vcvt.s32.f32 s0, s0
+; VFP2: vcvt.s32.f32 s{{.}}, s{{.}}
 ; NEON: test1:
 ; NEON: vcvt.s32.f32 d0, d0
 entry:
@@ -16,7 +16,7 @@ entry:
 
 define i32 @test2(float %a, float %b) {
 ; VFP2: test2:
-; VFP2: vcvt.u32.f32 s0, s0
+; VFP2: vcvt.u32.f32 s{{.}}, s{{.}}
 ; NEON: test2:
 ; NEON: vcvt.u32.f32 d0, d0
 entry:
@@ -27,7 +27,7 @@ entry:
 
 define float @test3(i32 %a, i32 %b) {
 ; VFP2: test3:
-; VFP2: vcvt.f32.u32 s0, s0
+; VFP2: vcvt.f32.u32 s{{.}}, s{{.}}
 ; NEON: test3:
 ; NEON: vcvt.f32.u32 d0, d0
 entry:
@@ -38,7 +38,7 @@ entry:
 
 define float @test4(i32 %a, i32 %b) {
 ; VFP2: test4:
-; VFP2: vcvt.f32.s32 s0, s0
+; VFP2: vcvt.f32.s32 s{{.}}, s{{.}}
 ; NEON: test4:
 ; NEON: vcvt.f32.s32 d0, d0
 entry: