Replace vfmaddxx213 instructions with their 231-type equivalents in accumulator

loops. Writing back to the accumulator (231-type) allows the coalescer to
eliminate an extra copy.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@199933 91177308-0d34-0410-b5e6-96231b3b80d8

3 files changed, 107 insertions, 3 deletions

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index ede17c1750..3e641cdba9 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -15963,6 +15963,81 @@ X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
   return MBB;
 }
 
+// Replace 213-type (isel default) FMA3 instructions with 231-type for
+// accumulator loops. Writing back to the accumulator allows the coalescer
+// to remove extra copies in the loop.   
+MachineBasicBlock *
+X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
+                                 MachineBasicBlock *MBB) const {
+  MachineOperand &AddendOp = MI->getOperand(3);
+
+  // Bail out early if the addend isn't a register - we can't switch these.
+  if (!AddendOp.isReg())
+    return MBB;
+
+  MachineFunction &MF = *MBB->getParent();
+  MachineRegisterInfo &MRI = MF.getRegInfo();
+
+  // Check whether the addend is defined by a PHI:
+  assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
+  MachineInstr &AddendDef = *MRI.def_begin(AddendOp.getReg());
+  if (!AddendDef.isPHI())
+    return MBB;
+
+  // Look for the following pattern:
+  // loop:
+  //   %addend = phi [%entry, 0], [%loop, %result]
+  //   ...
+  //   %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
+
+  // Replace with:
+  //   loop:
+  //   %addend = phi [%entry, 0], [%loop, %result]
+  //   ...
+  //   %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
+
+  for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
+    assert(AddendDef.getOperand(i).isReg());
+    MachineOperand PHISrcOp = AddendDef.getOperand(i);
+    MachineInstr &PHISrcInst = *MRI.def_begin(PHISrcOp.getReg());
+    if (&PHISrcInst == MI) {
+      // Found a matching instruction.
+      unsigned NewFMAOpc = 0;
+      switch (MI->getOpcode()) {
+        case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
+        case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
+        case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
+        case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
+        case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
+        case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
+        case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
+        case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
+        case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
+        case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
+        case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
+        case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
+        case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
+        case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
+        case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
+        case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
+        default: llvm_unreachable("Unrecognized FMA variant.");
+      }
+
+      const TargetInstrInfo &TII = *MF.getTarget().getInstrInfo();
+      MachineInstrBuilder MIB =
+        BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
+        .addOperand(MI->getOperand(0))
+        .addOperand(MI->getOperand(3))
+        .addOperand(MI->getOperand(2))
+        .addOperand(MI->getOperand(1));
+      MBB->insert(MachineBasicBlock::iterator(MI), MIB);
+      MI->eraseFromParent();
+    }
+  }
+
+  return MBB;
+}
+
 MachineBasicBlock *
 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
                                                MachineBasicBlock *BB) const {
@@ -16194,6 +16269,32 @@ X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
   case TargetOpcode::STACKMAP:
   case TargetOpcode::PATCHPOINT:
     return emitPatchPoint(MI, BB);
+
+  case X86::VFMADDPDr213r:
+  case X86::VFMADDPSr213r:
+  case X86::VFMADDSDr213r:
+  case X86::VFMADDSSr213r:
+  case X86::VFMSUBPDr213r:
+  case X86::VFMSUBPSr213r:
+  case X86::VFMSUBSDr213r:
+  case X86::VFMSUBSSr213r:
+  case X86::VFNMADDPDr213r:
+  case X86::VFNMADDPSr213r:
+  case X86::VFNMADDSDr213r:
+  case X86::VFNMADDSSr213r:
+  case X86::VFNMSUBPDr213r:
+  case X86::VFNMSUBPSr213r:
+  case X86::VFNMSUBSDr213r:
+  case X86::VFNMSUBSSr213r:
+  case X86::VFMADDPDr213rY:
+  case X86::VFMADDPSr213rY:
+  case X86::VFMSUBPDr213rY:
+  case X86::VFMSUBPSr213rY:
+  case X86::VFNMADDPDr213rY:
+  case X86::VFNMADDPSr213rY:
+  case X86::VFNMSUBPDr213rY:
+  case X86::VFNMSUBPSr213rY:
+    return emitFMA3Instr(MI, BB);
   }
 }
 
diff --git a/lib/Target/X86/X86ISelLowering.h b/lib/Target/X86/X86ISelLowering.h
index 9b32d12101..d985c98875 100644
--- a/lib/Target/X86/X86ISelLowering.h
+++ b/lib/Target/X86/X86ISelLowering.h
@@ -972,6 +972,9 @@ namespace llvm {
     MachineBasicBlock *emitEHSjLjLongJmp(MachineInstr *MI,
                                          MachineBasicBlock *MBB) const;
 
+    MachineBasicBlock *emitFMA3Instr(MachineInstr *MI,
+                                     MachineBasicBlock *MBB) const;
+
     /// Emit nodes that will be selected as "test Op0,Op0", or something
     /// equivalent, for use with the given x86 condition code.
     SDValue EmitTest(SDValue Op0, unsigned X86CC, SelectionDAG &DAG) const;
diff --git a/lib/Target/X86/X86InstrFMA.td b/lib/Target/X86/X86InstrFMA.td
index b2cc8209bf..206f7b600f 100644
--- a/lib/Target/X86/X86InstrFMA.td
+++ b/lib/Target/X86/X86InstrFMA.td
@@ -20,7 +20,7 @@ multiclass fma3p_rm<bits<8> opc, string OpcodeStr,
                     PatFrag MemFrag128, PatFrag MemFrag256,
                     ValueType OpVT128, ValueType OpVT256,
                     SDPatternOperator Op = null_frag> {
-  let isCommutable = 1 in
+  let isCommutable = 1, usesCustomInserter = 1 in
   def r     : FMA3<opc, MRMSrcReg, (outs VR128:$dst),
                    (ins VR128:$src1, VR128:$src2, VR128:$src3),
                    !strconcat(OpcodeStr,
@@ -36,7 +36,7 @@ multiclass fma3p_rm<bits<8> opc, string OpcodeStr,
                    [(set VR128:$dst, (OpVT128 (Op VR128:$src2, VR128:$src1,
                                                (MemFrag128 addr:$src3))))]>;
 
-  let isCommutable = 1 in
+  let isCommutable = 1, usesCustomInserter = 1 in
   def rY    : FMA3<opc, MRMSrcReg, (outs VR256:$dst),
                    (ins VR256:$src1, VR256:$src2, VR256:$src3),
                    !strconcat(OpcodeStr,
@@ -118,7 +118,7 @@ let Constraints = "$src1 = $dst" in {
 multiclass fma3s_rm<bits<8> opc, string OpcodeStr, X86MemOperand x86memop,
                     RegisterClass RC, ValueType OpVT, PatFrag mem_frag,
                     SDPatternOperator OpNode = null_frag> {
-  let isCommutable = 1 in
+  let isCommutable = 1, usesCustomInserter = 1 in
   def r     : FMA3<opc, MRMSrcReg, (outs RC:$dst),
                    (ins RC:$src1, RC:$src2, RC:$src3),
                    !strconcat(OpcodeStr,