[mips] Add option -mno-ldc1-sdc1.

This option is used when the user wants to avoid emitting double precision FP
loads and stores. Double precision FP loads and stores are expanded to single
precision instructions after register allocation.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181718 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/Mips/MipsInstrFPU.td

@@ -154,6 +154,7 @@ class LW_FT<string opstr, RegisterClass RC, InstrItinClass Itin,
   InstSE<(outs RC:$rt), (ins MemOpnd:$addr), !strconcat(opstr, "\t$rt, $addr"),
          [(set RC:$rt, (OpNode addrDefault:$addr))], Itin, FrmFI> {
   let DecoderMethod = "DecodeFMem";
+  let mayLoad = 1;
 }
 
 class SW_FT<string opstr, RegisterClass RC, InstrItinClass Itin,
@@ -161,6 +162,7 @@ class SW_FT<string opstr, RegisterClass RC, InstrItinClass Itin,
   InstSE<(outs), (ins RC:$rt, MemOpnd:$addr), !strconcat(opstr, "\t$rt, $addr"),
          [(OpNode RC:$rt, addrDefault:$addr)], Itin, FrmFI> {
   let DecoderMethod = "DecodeFMem";
+  let mayStore = 1;
 }
 
 class MADDS_FT<string opstr, RegisterClass RC, InstrItinClass Itin,
@@ -314,8 +316,12 @@ let Predicates = [NotN64, HasMips64, HasStdEnc],
 }
 
 let Predicates = [NotN64, NotMips64, HasStdEnc] in {
-  def LDC1 : LW_FT<"ldc1", AFGR64, IILoad, mem, load>, LW_FM<0x35>;
-  def SDC1 : SW_FT<"sdc1", AFGR64, IIStore, mem, store>, LW_FM<0x3d>;
+  let isPseudo = 1, isCodeGenOnly = 1 in {
+    def PseudoLDC1 : LW_FT<"", AFGR64, IILoad, mem, load>;
+    def PseudoSDC1 : SW_FT<"", AFGR64, IIStore, mem, store>;
+  }
+  def LDC1 : LW_FT<"ldc1", AFGR64, IILoad, mem>, LW_FM<0x35>;
+  def SDC1 : SW_FT<"sdc1", AFGR64, IIStore, mem>, LW_FM<0x3d>;
 }
 
 // Indexed loads and stores.
@@ -523,7 +529,7 @@ let AddedComplexity = 40 in {
   }
 
   let Predicates = [NotN64, NotMips64, HasStdEnc] in {
-    def : LoadRegImmPat<LDC1, f64, load>;
-    def : StoreRegImmPat<SDC1, f64>;
+    def : LoadRegImmPat<PseudoLDC1, f64, load>;
+    def : StoreRegImmPat<PseudoSDC1, f64>;
   }
 }

lib/Target/Mips/MipsSEInstrInfo.cpp

@@ -18,11 +18,17 @@
 #include "llvm/ADT/STLExtras.h"
 #include "llvm/CodeGen/MachineInstrBuilder.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/Support/CommandLine.h"
 #include "llvm/Support/ErrorHandling.h"
 #include "llvm/Support/TargetRegistry.h"
 
 using namespace llvm;
 
+static cl::opt<bool> NoDPLoadStore("mno-ldc1-sdc1", cl::init(false),
+                                   cl::desc("Expand double precision loads and "
+                                            "stores to their single precision "
+                                            "counterparts."));
+
 MipsSEInstrInfo::MipsSEInstrInfo(MipsTargetMachine &tm)
   : MipsInstrInfo(tm,
                   tm.getRelocationModel() == Reloc::PIC_ ? Mips::B : Mips::J),
@@ -253,6 +259,12 @@ bool MipsSEInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
   case Mips::ExtractElementF64:
     expandExtractElementF64(MBB, MI);
     break;
+  case Mips::PseudoLDC1:
+    expandDPLoadStore(MBB, MI, Mips::LDC1, Mips::LWC1);
+    break;
+  case Mips::PseudoSDC1:
+    expandDPLoadStore(MBB, MI, Mips::SDC1, Mips::SWC1);
+    break;
   case Mips::MIPSeh_return32:
   case Mips::MIPSeh_return64:
     expandEhReturn(MBB, MI);
@@ -393,6 +405,56 @@ void MipsSEInstrInfo::expandBuildPairF64(MachineBasicBlock &MBB,
     .addReg(HiReg);
 }
 
+/// Add 4 to the displacement of operand MO.
+static void fixDisp(MachineOperand &MO) {
+  switch (MO.getType()) {
+  default:
+    llvm_unreachable("Unhandled operand type.");
+  case MachineOperand::MO_Immediate:
+    MO.setImm(MO.getImm() + 4);
+    break;
+  case MachineOperand::MO_GlobalAddress:
+  case MachineOperand::MO_ConstantPoolIndex:
+  case MachineOperand::MO_BlockAddress:
+  case MachineOperand::MO_TargetIndex:
+  case MachineOperand::MO_ExternalSymbol:
+    MO.setOffset(MO.getOffset() + 4);
+    break;
+  }
+}
+
+void MipsSEInstrInfo::expandDPLoadStore(MachineBasicBlock &MBB,
+                                        MachineBasicBlock::iterator I,
+                                        unsigned OpcD, unsigned OpcS) const {
+  // If NoDPLoadStore is false, just change the opcode.
+  if (!NoDPLoadStore) {
+    genInstrWithNewOpc(OpcD, I);
+    return;
+  }
+
+  // Expand a double precision FP load or store to two single precision
+  // instructions.
+
+  const TargetRegisterInfo &TRI = getRegisterInfo();
+  const MachineOperand &ValReg = I->getOperand(0);
+  unsigned LoReg = TRI.getSubReg(ValReg.getReg(), Mips::sub_fpeven);
+  unsigned HiReg = TRI.getSubReg(ValReg.getReg(), Mips::sub_fpodd);
+
+  if (!TM.getSubtarget<MipsSubtarget>().isLittle())
+    std::swap(LoReg, HiReg);
+
+  // Create an instruction which loads from or stores to the lower memory
+  // address.
+  MachineInstrBuilder MIB = genInstrWithNewOpc(OpcS, I);
+  MIB->getOperand(0).setReg(LoReg);
+
+  // Create an instruction which loads from or stores to the higher memory
+  // address.
+  MIB = genInstrWithNewOpc(OpcS, I);
+  MIB->getOperand(0).setReg(HiReg);
+  fixDisp(MIB->getOperand(2));
+}
+
 void MipsSEInstrInfo::expandEhReturn(MachineBasicBlock &MBB,
                                      MachineBasicBlock::iterator I) const {
   // This pseudo instruction is generated as part of the lowering of

lib/Target/Mips/MipsSEInstrInfo.h

@@ -87,6 +87,9 @@ private:
                                MachineBasicBlock::iterator I) const;
   void expandBuildPairF64(MachineBasicBlock &MBB,
                           MachineBasicBlock::iterator I) const;
+  void expandDPLoadStore(MachineBasicBlock &MBB,
+                         MachineBasicBlock::iterator I, unsigned OpcD,
+                         unsigned OpcS) const;
   void expandEhReturn(MachineBasicBlock &MBB,
                       MachineBasicBlock::iterator I) const;
 };

test/CodeGen/Mips/mno-ldc1-sdc1.ll

@@ -0,0 +1,45 @@
+; RUN: llc -march=mipsel -relocation-model=pic -mno-ldc1-sdc1 < %s | \
+; RUN: FileCheck %s -check-prefix=LE-PIC
+; RUN: llc -march=mipsel -relocation-model=static -mno-ldc1-sdc1 < %s | \
+; RUN: FileCheck %s -check-prefix=LE-STATIC
+; RUN: llc -march=mips -relocation-model=pic -mno-ldc1-sdc1 < %s | \
+; RUN: FileCheck %s -check-prefix=BE-PIC
+; RUN: llc -march=mipsel < %s | FileCheck %s -check-prefix=CHECK-LDC1-SDC1
+
+@g0 = common global double 0.000000e+00, align 8
+
+; LE-PIC: test_ldc1:
+; LE-PIC: lwc1 $f0, 0(${{[0-9]+}})
+; LE-PIC: lwc1 $f1, 4(${{[0-9]+}})
+; LE-STATIC: test_ldc1:
+; LE-STATIC: lwc1 $f0, %lo(g0)(${{[0-9]+}})
+; LE-STATIC: lwc1 $f1, %lo(g0+4)(${{[0-9]+}})
+; BE-PIC: test_ldc1:
+; BE-PIC: lwc1 $f1, 0(${{[0-9]+}})
+; BE-PIC: lwc1 $f0, 4(${{[0-9]+}})
+; CHECK-LDC1-SDC1: test_ldc1:
+; CHECK-LDC1-SDC1: ldc1 $f{{[0-9]+}}
+
+define double @test_ldc1() {
+entry:
+  %0 = load double* @g0, align 8
+  ret double %0
+}
+
+; LE-PIC: test_sdc1:
+; LE-PIC: swc1 $f12, 0(${{[0-9]+}})
+; LE-PIC: swc1 $f13, 4(${{[0-9]+}})
+; LE-STATIC: test_sdc1:
+; LE-STATIC: swc1 $f12, %lo(g0)(${{[0-9]+}})
+; LE-STATIC: swc1 $f13, %lo(g0+4)(${{[0-9]+}})
+; BE-PIC: test_sdc1:
+; BE-PIC: swc1 $f13, 0(${{[0-9]+}})
+; BE-PIC: swc1 $f12, 4(${{[0-9]+}})
+; CHECK-LDC1-SDC1: test_sdc1:
+; CHECK-LDC1-SDC1: sdc1 $f{{[0-9]+}}
+
+define void @test_sdc1(double %a) {
+entry:
+  store double %a, double* @g0, align 8
+  ret void
+}