Cherrypick upstream ARM FastISel ext patches

Specifically:
  r186489 - Fix ARMFastISel::ARMEmitIntExt shift emission
  r183794 - ARM FastISel fix sext/zext fold
  r183601 - Fix unused variable warning from my previous patch
  r183551 - ARM FastISel integer sext/zext improvements

These should fix some failures that I had run into back then, as well as make ARM FastISel faster because it doesn't go to SelectionDAG.

BUG= https://code.google.com/p/nativeclient/issues/detail?id=3501
R=jvoung@chromium.org
TEST= make check-all

Review URL: https://codereview.chromium.org/19992002

1 files changed, 162 insertions, 57 deletions

diff --git a/lib/Target/ARM/ARMFastISel.cpp b/lib/Target/ARM/ARMFastISel.cpp
index 0b29c74ca2..82518d3a32 100644
--- a/lib/Target/ARM/ARMFastISel.cpp
+++ b/lib/Target/ARM/ARMFastISel.cpp
@@ -20,6 +20,7 @@
 #include "ARMSubtarget.h"
 #include "ARMTargetMachine.h"
 #include "MCTargetDesc/ARMAddressingModes.h"
+#include "llvm/ADT/STLExtras.h"
 #include "llvm/CodeGen/Analysis.h"
 #include "llvm/CodeGen/FastISel.h"
 #include "llvm/CodeGen/FunctionLoweringInfo.h"
@@ -2019,7 +2020,7 @@ bool ARMFastISel::ProcessCallArgs(SmallVectorImpl<Value*> &Args,
       case CCValAssign::ZExt: {
         MVT DestVT = VA.getLocVT();
         Arg = ARMEmitIntExt(ArgVT, Arg, DestVT, /*isZExt*/true);
-        assert (Arg != 0 && "Failed to emit a sext");
+        assert (Arg != 0 && "Failed to emit a zext");
         ArgVT = DestVT;
         break;
       }
@@ -2636,47 +2637,135 @@ unsigned ARMFastISel::ARMEmitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT,
                                     bool isZExt) {
   if (DestVT != MVT::i32 && DestVT != MVT::i16 && DestVT != MVT::i8)
     return 0;
+  if (SrcVT != MVT::i16 && SrcVT != MVT::i8 && SrcVT != MVT::i1)
+    return 0;
 
-  unsigned Opc;
-  bool isBoolZext = false;
-  const TargetRegisterClass *RC;
-  switch (SrcVT.SimpleTy) {
-  default: return 0;
-  case MVT::i16:
-    if (!Subtarget->hasV6Ops()) return 0;
-    RC = isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRnopcRegClass;
-    if (isZExt)
-      Opc = isThumb2 ? ARM::t2UXTH : ARM::UXTH;
-    else
-      Opc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
-    break;
-  case MVT::i8:
-    if (!Subtarget->hasV6Ops()) return 0;
-    RC = isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRnopcRegClass;
-    if (isZExt)
-      Opc = isThumb2 ? ARM::t2UXTB : ARM::UXTB;
-    else
-      Opc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
-    break;
-  case MVT::i1:
-    if (isZExt) {
-      RC = isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRRegClass;
-      Opc = isThumb2 ? ARM::t2ANDri : ARM::ANDri;
-      isBoolZext = true;
-      break;
+  // Table of which combinations can be emitted as a single instruction,
+  // and which will require two.
+  static const uint8_t isSingleInstrTbl[3][2][2][2] = {
+    //            ARM                     Thumb
+    //           !hasV6Ops  hasV6Ops     !hasV6Ops  hasV6Ops
+    //    ext:     s  z      s  z          s  z      s  z
+    /*  1 */ { { { 0, 1 }, { 0, 1 } }, { { 0, 0 }, { 0, 1 } } },
+    /*  8 */ { { { 0, 1 }, { 1, 1 } }, { { 0, 0 }, { 1, 1 } } },
+    /* 16 */ { { { 0, 0 }, { 1, 1 } }, { { 0, 0 }, { 1, 1 } } }
+  };
+
+  // Target registers for:
+  //  - For ARM can never be PC.
+  //  - For 16-bit Thumb are restricted to lower 8 registers.
+  //  - For 32-bit Thumb are restricted to non-SP and non-PC.
+  static const TargetRegisterClass *RCTbl[2][2] = {
+    // Instructions: Two                     Single
+    /* ARM      */ { &ARM::GPRnopcRegClass, &ARM::GPRnopcRegClass },
+    /* Thumb    */ { &ARM::tGPRRegClass,    &ARM::rGPRRegClass    }
+  };
+
+  // Table governing the instruction(s) to be emitted.
+  static const struct InstructionTable {
+    uint32_t Opc   : 16;
+    uint32_t hasS  :  1; // Some instructions have an S bit, always set it to 0.
+    uint32_t Shift :  7; // For shift operand addressing mode, used by MOVsi.
+    uint32_t Imm   :  8; // All instructions have either a shift or a mask.
+  } IT[2][2][3][2] = {
+    { // Two instructions (first is left shift, second is in this table).
+      { // ARM                Opc           S  Shift             Imm
+        /*  1 bit sext */ { { ARM::MOVsi  , 1, ARM_AM::asr     ,  31 },
+        /*  1 bit zext */   { ARM::MOVsi  , 1, ARM_AM::lsr     ,  31 } },
+        /*  8 bit sext */ { { ARM::MOVsi  , 1, ARM_AM::asr     ,  24 },
+        /*  8 bit zext */   { ARM::MOVsi  , 1, ARM_AM::lsr     ,  24 } },
+        /* 16 bit sext */ { { ARM::MOVsi  , 1, ARM_AM::asr     ,  16 },
+        /* 16 bit zext */   { ARM::MOVsi  , 1, ARM_AM::lsr     ,  16 } }
+      },
+      { // Thumb              Opc           S  Shift             Imm
+        /*  1 bit sext */ { { ARM::tASRri , 0, ARM_AM::no_shift,  31 },
+        /*  1 bit zext */   { ARM::tLSRri , 0, ARM_AM::no_shift,  31 } },
+        /*  8 bit sext */ { { ARM::tASRri , 0, ARM_AM::no_shift,  24 },
+        /*  8 bit zext */   { ARM::tLSRri , 0, ARM_AM::no_shift,  24 } },
+        /* 16 bit sext */ { { ARM::tASRri , 0, ARM_AM::no_shift,  16 },
+        /* 16 bit zext */   { ARM::tLSRri , 0, ARM_AM::no_shift,  16 } }
+      }
+    },
+    { // Single instruction.
+      { // ARM                Opc           S  Shift             Imm
+        /*  1 bit sext */ { { ARM::KILL   , 0, ARM_AM::no_shift,   0 },
+        /*  1 bit zext */   { ARM::ANDri  , 1, ARM_AM::no_shift,   1 } },
+        /*  8 bit sext */ { { ARM::SXTB   , 0, ARM_AM::no_shift,   0 },
+        /*  8 bit zext */   { ARM::ANDri  , 1, ARM_AM::no_shift, 255 } },
+        /* 16 bit sext */ { { ARM::SXTH   , 0, ARM_AM::no_shift,   0 },
+        /* 16 bit zext */   { ARM::UXTH   , 0, ARM_AM::no_shift,   0 } }
+      },
+      { // Thumb              Opc           S  Shift             Imm
+        /*  1 bit sext */ { { ARM::KILL   , 0, ARM_AM::no_shift,   0 },
+        /*  1 bit zext */   { ARM::t2ANDri, 1, ARM_AM::no_shift,   1 } },
+        /*  8 bit sext */ { { ARM::t2SXTB , 0, ARM_AM::no_shift,   0 },
+        /*  8 bit zext */   { ARM::t2ANDri, 1, ARM_AM::no_shift, 255 } },
+        /* 16 bit sext */ { { ARM::t2SXTH , 0, ARM_AM::no_shift,   0 },
+        /* 16 bit zext */   { ARM::t2UXTH , 0, ARM_AM::no_shift,   0 } }
+      }
     }
-    return 0;
+  };
+
+  unsigned SrcBits = SrcVT.getSizeInBits();
+  unsigned DestBits = DestVT.getSizeInBits();
+  (void) DestBits;
+  assert((SrcBits < DestBits) && "can only extend to larger types");
+  assert((DestBits == 32 || DestBits == 16 || DestBits == 8) &&
+         "other sizes unimplemented");
+  assert((SrcBits == 16 || SrcBits == 8 || SrcBits == 1) &&
+         "other sizes unimplemented");
+
+  bool hasV6Ops = Subtarget->hasV6Ops();
+  unsigned Bitness = SrcBits / 8;  // {1,8,16}=>{0,1,2}
+  assert((Bitness < 3) && "sanity-check table bounds");
+
+  bool isSingleInstr = isSingleInstrTbl[Bitness][isThumb2][hasV6Ops][isZExt];
+  const TargetRegisterClass *RC = RCTbl[isThumb2][isSingleInstr];
+  const InstructionTable *ITP = &IT[isSingleInstr][isThumb2][Bitness][isZExt];
+  unsigned Opc = ITP->Opc;
+  assert(ARM::KILL != Opc && "Invalid table entry");
+  unsigned hasS = ITP->hasS;
+  ARM_AM::ShiftOpc Shift = (ARM_AM::ShiftOpc) ITP->Shift;
+  assert(((Shift == ARM_AM::no_shift) == (Opc != ARM::MOVsi)) &&
+         "only MOVsi has shift operand addressing mode");
+  unsigned Imm = ITP->Imm;
+
+  // 16-bit Thumb instructions always set CPSR (unless they're in an IT block).
+  bool setsCPSR = &ARM::tGPRRegClass == RC;
+  unsigned LSLOpc = isThumb2 ? ARM::tLSLri : ARM::MOVsi;
+  unsigned ResultReg;
+  // MOVsi encodes shift and immediate in shift operand addressing mode.
+  // The following condition has the same value when emitting two
+  // instruction sequences: both are shifts.
+  bool ImmIsSO = (Shift != ARM_AM::no_shift);
+
+  // Either one or two instructions are emitted.
+  // They're always of the form:
+  //   dst = in OP imm
+  // CPSR is set only by 16-bit Thumb instructions.
+  // Predicate, if any, is AL.
+  // S bit, if available, is always 0.
+  // When two are emitted the first's result will feed as the second's input,
+  // that value is then dead.
+  unsigned NumInstrsEmitted = isSingleInstr ? 1 : 2;
+  for (unsigned Instr = 0; Instr != NumInstrsEmitted; ++Instr) {
+    ResultReg = createResultReg(RC);
+    bool isLsl = (0 == Instr) && !isSingleInstr;
+    unsigned Opcode = isLsl ? LSLOpc : Opc;
+    ARM_AM::ShiftOpc ShiftAM = isLsl ? ARM_AM::lsl : Shift;
+    unsigned ImmEnc = ImmIsSO ? ARM_AM::getSORegOpc(ShiftAM, Imm) : Imm;
+    bool isKill = 1 == Instr;
+    MachineInstrBuilder MIB = BuildMI(
+        *FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opcode), ResultReg);
+    if (setsCPSR)
+      MIB.addReg(ARM::CPSR, RegState::Define);
+    AddDefaultPred(MIB.addReg(SrcReg, isKill * RegState::Kill).addImm(ImmEnc));
+    if (hasS)
+      AddDefaultCC(MIB);
+    // Second instruction consumes the first's result.
+    SrcReg = ResultReg;
   }
 
-  unsigned ResultReg = createResultReg(RC);
-  MachineInstrBuilder MIB;
-  MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), ResultReg)
-        .addReg(SrcReg);
-  if (isBoolZext)
-    MIB.addImm(1);
-  else
-    MIB.addImm(0);
-  AddOptionalDefs(MIB);
   return ResultReg;
 }
 
@@ -2831,6 +2920,25 @@ bool ARMFastISel::TargetSelectInstruction(const Instruction *I) {
   return false;
 }
 
+namespace {
+// This table describes sign- and zero-extend instructions which can be
+// folded into a preceding load. All of these extends have an immediate
+// (sometimes a mask and sometimes a shift) that's applied after
+// extension.
+const struct FoldableLoadExtendsStruct {
+  uint16_t Opc[2];  // ARM, Thumb.
+  uint8_t ExpectedImm;
+  uint8_t isZExt     : 1;
+  uint8_t ExpectedVT : 7;
+} FoldableLoadExtends[] = {
+  { { ARM::SXTH,  ARM::t2SXTH  },   0, 0, MVT::i16 },
+  { { ARM::UXTH,  ARM::t2UXTH  },   0, 1, MVT::i16 },
+  { { ARM::ANDri, ARM::t2ANDri }, 255, 1, MVT::i8  },
+  { { ARM::SXTB,  ARM::t2SXTB  },   0, 0, MVT::i8  },
+  { { ARM::UXTB,  ARM::t2UXTB  },   0, 1, MVT::i8  }
+};
+}
+
 /// \brief The specified machine instr operand is a vreg, and that
 /// vreg is being provided by the specified load instruction.  If possible,
 /// try to fold the load as an operand to the instruction, returning true if
@@ -2846,26 +2954,23 @@ bool ARMFastISel::tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo,
   // ldrb r1, [r0]       ldrb r1, [r0]
   // uxtb r2, r1     =>
   // mov  r3, r2         mov  r3, r1
-  bool isZExt = true;
-  switch(MI->getOpcode()) {
-    default: return false;
-    case ARM::SXTH:
-    case ARM::t2SXTH:
-      isZExt = false;
-    case ARM::UXTH:
-    case ARM::t2UXTH:
-      if (VT != MVT::i16)
-        return false;
-    break;
-    case ARM::SXTB:
-    case ARM::t2SXTB:
-      isZExt = false;
-    case ARM::UXTB:
-    case ARM::t2UXTB:
-      if (VT != MVT::i8)
-        return false;
-    break;
+  if (MI->getNumOperands() < 3 || !MI->getOperand(2).isImm())
+    return false;
+  const uint64_t Imm = MI->getOperand(2).getImm();
+
+  bool Found = false;
+  bool isZExt;
+  for (unsigned i = 0, e = array_lengthof(FoldableLoadExtends);
+       i != e; ++i) {
+    if (FoldableLoadExtends[i].Opc[isThumb2] == MI->getOpcode() &&
+        (uint64_t)FoldableLoadExtends[i].ExpectedImm == Imm &&
+        MVT((MVT::SimpleValueType)FoldableLoadExtends[i].ExpectedVT) == VT) {
+      Found = true;
+      isZExt = FoldableLoadExtends[i].isZExt;
+    }
   }
+  if (!Found) return false;
+
   // See if we can handle this address.
   Address Addr;
   if (!ARMComputeAddress(LI->getOperand(0), Addr)) return false;