[Mips] Add support to match more patterns for DEXT and CINS

[violetav]

Summary

This patch adds support for recognizing more patterns to match to DEXT and CINS instructions.
It finds cases where multiple instructions could be replaced with a single DEXT or CINS instruction. For example, for the following:

define i64 @dext_and32(i64 zeroext %a) {
entry:

%and = and i64 %a, 4294967295
ret i64 %and

}

instead of generating:

0000000000000088 <dext_and32>:

88:   64010001        daddiu  at,zero,1
8c:   0001083c        dsll32  at,at,0x0
90:   6421ffff        daddiu  at,at,-1
94:   03e00008        jr      ra
98:   00811024        and     v0,a0,at
9c:   00000000        nop

the following gets generated:

0000000000000068 <dext_and32>:

68:   03e00008        jr      ra
6c:   7c82f803        dext    v0,a0,0x0,0x20

Cases that are covered:
DEXT:

1. and $src, mask where mask > 0xffff
2. zext $src zero extend from i32 to i64

CINS:

1. and (shl $src, pos), mask
2. shl (and $src, mask), pos
3. zext (shl $src, pos) zero extend from i32 to i64

[sdardis]

Can you upload the patch with more context? http://llvm.org/docs/Phabricator.html#requesting-a-review-via-the-web-interface

[violetav]

Uploaded the patch with more context.

[sdardis]

A few issues, mostly formatting and the correct predicates on the instructions. From my reading of the cnMIPS instruction set, the cins and cins32 instruction require that we're compiling for MIPS64 as they require access to 64 bit operations, so these instructions and patterns need to have ASE_MIPS64_CNMIPS. Additionally, they--along with the dext pattern--needs to be marked as NotInMicroMips through the AdditionalPredicates mechanism.

microMIPS64R6 does have the dext family of instructions but lacks the necessary instruction mappings, so if the instruction was selected by ISel because the pattern matched, the resulting code would be only partially translated to microMIPS64R6. This isn't a problem if the result is an assembly which is then feed back into clang or an external assembler, as the correct instruction would be parsed. It is a problem if direct object emission is used as the resulting object would have instructions from two different ISAs without the correct mode switch.

The new test cases can be cut down by getting rid of the metadata and annotations that clang produces when compiling to llvm-ir.

lib/Target/Mips/Mips64InstrInfo.td

334–336

Formatting. Align the (ins) with the out above, when you have to break the line for the (ins), align the parameters of the (ins) under one another. Align the assembly text with (outs) and the rest of the parameters to InstSE with the outs when a line break is needed. The EXT_FM<3> or ISA_MIPS64R2 is aligned with the InstSE as required.

This requires AdditionalPredicates = [NotInMicroMips].

437–438

Align the EXTS_FM with ExtsCins and ASE_CNMIPS should be ASE_MIPS64_CNMIPS.

437–460

All these defs need to be covered with AdditionalPredicates = [NotInMicroMips].

438–460

Align the EXTS_FM with ExtsCins and ASE_CNMIPS should be ASE_MIPS64_CNMIPS.

444

Align the EXTS_FM with ExtsCins. This definition also requires ASE_MIPS64_CNMIPS.

446–448

Align these lines to under the (outs). This definition requires ASE_MIPS64_CNMIPS.

675–677

Use ISA_MIPS64R2 rather than AdditionalPredicates. E.g. this should be:

def : MipsPat<(i64 (zext GPR32:$src)), (DEXT64_32 GPR32:$src, 0, 32)>, ISA_MIPS64R2;

The AdditionalPredicates should be NotInMicroMips.

678–681

This needs ASE_MIPS64_CNMIPS. The AdditionalPredicates should have NotInMicroMips. The two pattern definitions can be covered with the one "let AdditionalPredicates = [NotInMicroMips]" scope.

lib/Target/Mips/MipsISelLowering.cpp

947–949

You can fold Opc = MipsISD::CIns into it's usage here.

test/CodeGen/Mips/cins.ll

2

See my comments about dext.ll.

test/CodeGen/Mips/dext.ll

2

Change this runline to match the ones found under test/CodeGen/Mips/, that is the architecture, cpu and ABI are specified on the run line.

4–7

Drop all this after you've made the above change as it'll be irrelevant.

9

After you've re-written @dext_index1, you can drop this.

11–12

Can you move the CHECKs into the function near the IR that they are testing the codegen of? It makes it clearer to see what output is expected for a given input.

14

These types of lines can be removed as they're unnecessary for the purposes of this test.

15

local_unnamed_addr #0 can be removed as well.

15–30

Rewrite this so that it takes a i32 signext argument and returns a i64 zext result, the the body can just be

%res = zext i32 %n to i64
ret i64 %res

35–43

This test is redundant as the patterns it's likely to produce are covered by the case above.

138–139

Add whatever relevant attributes are required onto the runline instead of relying on function attributes and remove these attribute declarations.

141–147

All of this can be removed as it's irrelevant.

141–147

This is irrelevant and can be removed.

test/CodeGen/Mips/load-store-left-right.ll

12

This is inconsistent with the little endian test above. Also, there are no check lines for MIPS64R2-EB.

186–188

Add the MIPS64R2-EB variant.

487–489

MIPS64R2-EB produces a dext here, so test for that.

test/MC/Mips/sext_64_32.ll

14–15

Add a CHECK-LABEL: foo_2: before this.

[violetav]

Addressed comments. Thank you for the review.

[violetav]

test/CodeGen/Mips/dext.ll

15–30

There is already a test (dext_zext) with the same body that you are proposing, so I removed the dext_index test that produces the same pattern. As for the dext_index1 test, it produces a different pattern (I simplified the body and renamed the test 'dext_add_zext'). The difference is that in the first case, the 'zero_extend' node gets transformed to 'and' node which is then further combined to dext, and in the other case, the zero_extend node doesn't get transformed to 'and' and later DEXT64_32 gets selected.

[sdardis]

Thanks, LGTM with inline nits addressed.

lib/Target/Mips/Mips64InstrInfo.td

438–459

Indent the contents of the "let AdditionalPredicates = [NotInMicroMips} in" by 2 spaces.

lib/Target/Mips/MipsInstrInfo.td

1176

Overly long line, reformat this to match the formatting of the line above.

[violetav]

Addressed comments.

[violetav]

Rebased patch.

[violetav]

Added missing CHECKs in load-store-left-right.ll test.

lib/Target/Mips/Mips64InstrInfo.td

let AdditionalPredicates = [NotInMicroMips] in {
   def DINS : InsBase<"dins", GPR64Opnd, uimm6, uimm5_inssize_plus1, MipsIns>,
              EXT_FM<7>, ISA_MIPS64R2;
   def DINSU : InsBase<"dinsu", GPR64Opnd, uimm5_plus32, uimm5_inssize_plus1>,
               EXT_FM<6>, ISA_MIPS64R2;
   def DINSM : InsBase<"dinsm", GPR64Opnd, uimm5, uimm5_inssize_plus1>,
               EXT_FM<5>, ISA_MIPS64R2;
 }
 
+let isCodeGenOnly = 1, AdditionalPredicates = [NotInMicroMips] in {
+  def DEXT64_32 : InstSE<(outs GPR64Opnd:$rt),
+                         (ins GPR32Opnd:$rs, uimm5_report_uimm6:$pos,
+                              uimm5_plus1:$size),
+                         "dext $rt, $rs, $pos, $size", [], II_EXT, FrmR, "dext">,
+                  EXT_FM<3>, ISA_MIPS64R2;
+}
+
 let isCodeGenOnly = 1, rs = 0, shamt = 0 in {
   def DSLL64_32 : FR<0x00, 0x3c, (outs GPR64:$rd), (ins GPR32:$rt),
                      "dsll\t$rd, $rt, 32", [], II_DSLL>;
   def SLL64_32 : FR<0x0, 0x00, (outs GPR64:$rd), (ins GPR32:$rt),
                     "sll\t$rd, $rt, 0", [], II_SLL>;
   def SLL64_64 : FR<0x0, 0x00, (outs GPR64:$rd), (ins GPR64:$rt),
                     "sll\t$rd, $rt, 0", [], II_SLL>;
 }
let DecoderNamespace = "CnMips",
     EncodingPredicates = []<Predicate> in {
 
 class Count1s<string opstr, RegisterOperand RO>:
   InstSE<(outs RO:$rd), (ins RO:$rs), !strconcat(opstr, "\t$rd, $rs"),
          [(set RO:$rd, (ctpop RO:$rs))], II_POP, FrmR, opstr> {
   let TwoOperandAliasConstraint = "$rd = $rs";
 }
 
-class ExtsCins<string opstr, InstrItinClass itin,
-               SDPatternOperator Op = null_frag>:
-  InstSE<(outs GPR64Opnd:$rt), (ins GPR64Opnd:$rs, uimm5:$pos, uimm5:$lenm1),
-         !strconcat(opstr, " $rt, $rs, $pos, $lenm1"),
-         [(set GPR64Opnd:$rt, (Op GPR64Opnd:$rs, imm:$pos, imm:$lenm1))],
+class ExtsCins<string opstr, InstrItinClass itin, RegisterOperand RO,
+               PatFrag PosImm, SDPatternOperator Op = null_frag>:
+  InstSE<(outs RO:$rt), (ins RO:$rs, uimm5:$pos, uimm5:$lenm1),
+         !strconcat(opstr, "\t$rt, $rs, $pos, $lenm1"),
+         [(set RO:$rt, (Op RO:$rs, PosImm:$pos, imm:$lenm1))],
          itin, FrmR, opstr> {
   let TwoOperandAliasConstraint = "$rt = $rs";
 }
 
 class SetCC64_R<string opstr, PatFrag cond_op> :
   InstSE<(outs GPR64Opnd:$rd), (ins GPR64Opnd:$rs, GPR64Opnd:$rt),
          !strconcat(opstr, "\t$rd, $rs, $rt"),
          [(set GPR64Opnd:$rd, (zext (cond_op GPR64Opnd:$rs,

 def BBIT132: CBranchBitNum<"bbit132", brtarget, setne, GPR64Opnd, uimm5_64,
                            0x100000000>, BBIT_FM<0x3e>, ASE_CNMIPS;
 
 // Multiply Doubleword to GPR
 def DMUL  : ArithLogicR<"dmul", GPR64Opnd, 1, II_DMUL, mul>,
             ADD_FM<0x1c, 0x03>, ASE_CNMIPS {
   let Defs = [HI0, LO0, P0, P1, P2];
 }
 
+let AdditionalPredicates = [NotInMicroMips] in {
   // Extract a signed bit field /+32
-def EXTS  : ExtsCins<"exts", II_EXT>, EXTS_FM<0x3a>, ASE_CNMIPS;
-def EXTS32: ExtsCins<"exts32", II_EXT>, EXTS_FM<0x3b>, ASE_CNMIPS;
+  def EXTS  : ExtsCins<"exts", II_EXT, GPR64Opnd, immZExt5>, EXTS_FM<0x3a>,
+              ASE_MIPS64_CNMIPS;
+  def EXTS32: ExtsCins<"exts32", II_EXT, GPR64Opnd, immZExt5Plus32>,
+              EXTS_FM<0x3b>, ASE_MIPS64_CNMIPS;
 
   // Clear and insert a bit field /+32
-def CINS  : ExtsCins<"cins", II_INS>, EXTS_FM<0x32>, ASE_CNMIPS;
-def CINS32: ExtsCins<"cins32", II_INS>, EXTS_FM<0x33>, ASE_CNMIPS;
+  def CINS  : ExtsCins<"cins", II_INS, GPR64Opnd, immZExt5, MipsCIns>,
+              EXTS_FM<0x32>, ASE_MIPS64_CNMIPS;
+  def CINS32: ExtsCins<"cins32", II_INS, GPR64Opnd, immZExt5Plus32, MipsCIns>,
+              EXTS_FM<0x33>, ASE_MIPS64_CNMIPS;
+  let isCodeGenOnly = 1 in {
+    def CINS_i32 : ExtsCins<"cins", II_INS, GPR32Opnd, immZExt5, MipsCIns>,
+                   EXTS_FM<0x32>, ASE_MIPS64_CNMIPS;
+    def CINS64_32 :InstSE<(outs GPR64Opnd:$rt),
+                          (ins GPR32Opnd:$rs, uimm5:$pos, uimm5:$lenm1),
+                          "cins\t$rt, $rs, $pos, $lenm1", [], II_INS, FrmR,
+                          "cins">,
+                   EXTS_FM<0x32>, ASE_MIPS64_CNMIPS;
+  }
+}
 
 // Move to multiplier/product register
 def MTM0   : MoveToLOHI<"mtm0", GPR64Opnd, [MPL0, P0, P1, P2]>, MTMR_FM<0x08>,
              ASE_CNMIPS;
 def MTM1   : MoveToLOHI<"mtm1", GPR64Opnd, [MPL1, P0, P1, P2]>, MTMR_FM<0x0c>,
              ASE_CNMIPS;
 def MTM2   : MoveToLOHI<"mtm2", GPR64Opnd, [MPL2, P0, P1, P2]>, MTMR_FM<0x0d>,
              ASE_CNMIPS;
 def MTP0   : MoveToLOHI<"mtp0", GPR64Opnd, [P0]>, MTMR_FM<0x09>, ASE_CNMIPS;

 }
 
 // 32-to-64-bit extension
 def : MipsPat<(i64 (anyext GPR32:$src)),
               (INSERT_SUBREG (i64 (IMPLICIT_DEF)), GPR32:$src, sub_32)>;
 def : MipsPat<(i64 (zext GPR32:$src)), (DSRL (DSLL64_32 GPR32:$src), 32)>;
 def : MipsPat<(i64 (sext GPR32:$src)), (SLL64_32 GPR32:$src)>;
 
+let AdditionalPredicates = [NotInMicroMips] in {
+  def : MipsPat<(i64 (zext GPR32:$src)), (DEXT64_32 GPR32:$src, 0, 32)>,
+        ISA_MIPS64R2;
+  def : MipsPat<(i64 (zext (i32 (shl GPR32:$rt, immZExt5:$imm)))),
+                (CINS64_32 GPR32:$rt, imm:$imm, (immZExt5To31 imm:$imm))>,
+        ASE_MIPS64_CNMIPS;
+}
+
 // Sign extend in register
 def : MipsPat<(i64 (sext_inreg GPR64:$src, i32)),
               (SLL64_64 GPR64:$src)>;
 
 // bswap MipsPattern
 def : MipsPat<(bswap GPR64:$rt), (DSHD (DSBH GPR64:$rt))>;
 
 // Carry pattern
def : MipsInstAlias<"bbit1 $rs, $p, $offset",
                     (BBIT132 GPR64Opnd:$rs, uimm5_plus32_normalize_64:$p,
                              brtarget:$offset), 0>,
       ASE_CNMIPS;
 
 // exts with $pos 32-63 in converted to exts32 with $pos 0-31
 def : MipsInstAlias<"exts $rt, $rs, $pos, $lenm1",
                     (EXTS32 GPR64Opnd:$rt, GPR64Opnd:$rs,
                             uimm5_plus32_normalize:$pos, uimm5:$lenm1), 0>,
-      ASE_CNMIPS;
+      ASE_MIPS64_CNMIPS;
 def : MipsInstAlias<"exts $rt, $pos, $lenm1",
                     (EXTS32 GPR64Opnd:$rt, GPR64Opnd:$rt,
                             uimm5_plus32_normalize:$pos, uimm5:$lenm1), 0>,
-      ASE_CNMIPS;
+      ASE_MIPS64_CNMIPS;
 
 // cins with $pos 32-63 in converted to cins32 with $pos 0-31
 def : MipsInstAlias<"cins $rt, $rs, $pos, $lenm1",
                     (CINS32 GPR64Opnd:$rt, GPR64Opnd:$rs,
                             uimm5_plus32_normalize:$pos, uimm5:$lenm1), 0>,
-      ASE_CNMIPS;
+      ASE_MIPS64_CNMIPS;
 def : MipsInstAlias<"cins $rt, $pos, $lenm1",
                     (CINS32 GPR64Opnd:$rt, GPR64Opnd:$rt,
                             uimm5_plus32_normalize:$pos, uimm5:$lenm1), 0>,
-      ASE_CNMIPS;
+      ASE_MIPS64_CNMIPS;
 
 //===----------------------------------------------------------------------===//
 // Assembler Pseudo Instructions
 //===----------------------------------------------------------------------===//
 
 class LoadImmediate64<string instr_asm, Operand Od, RegisterOperand RO> :
   MipsAsmPseudoInst<(outs RO:$rt), (ins Od:$imm64),
                      !strconcat(instr_asm, "\t$rt, $imm64")> ;

lib/Target/Mips/MipsISelLowering.h

enum NodeType : unsigned {
       Wrapper,
 
       DynAlloc,
 
       Sync,
 
       Ext,
       Ins,
+      CIns,
 
       // EXTR.W instrinsic nodes.
       EXTP,
       EXTPDP,
       EXTR_S_H,
       EXTR_W,
       EXTR_R_W,
       EXTR_RS_W,

lib/Target/Mips/MipsISelLowering.cpp

const char *MipsTargetLowering::getTargetNodeName(unsigned Opcode) const {
   case MipsISD::DivRemU16:         return "MipsISD::DivRemU16";
   case MipsISD::BuildPairF64:      return "MipsISD::BuildPairF64";
   case MipsISD::ExtractElementF64: return "MipsISD::ExtractElementF64";
   case MipsISD::Wrapper:           return "MipsISD::Wrapper";
   case MipsISD::DynAlloc:          return "MipsISD::DynAlloc";
   case MipsISD::Sync:              return "MipsISD::Sync";
   case MipsISD::Ext:               return "MipsISD::Ext";
   case MipsISD::Ins:               return "MipsISD::Ins";
+  case MipsISD::CIns:              return "MipsISD::CIns";
   case MipsISD::LWL:               return "MipsISD::LWL";
   case MipsISD::LWR:               return "MipsISD::LWR";
   case MipsISD::SWL:               return "MipsISD::SWL";
   case MipsISD::SWR:               return "MipsISD::SWR";
   case MipsISD::LDL:               return "MipsISD::LDL";
   case MipsISD::LDR:               return "MipsISD::LDR";
   case MipsISD::SDL:               return "MipsISD::SDL";
   case MipsISD::SDR:               return "MipsISD::SDR";
MipsTargetLowering::MipsTargetLowering(const MipsTargetMachine &TM,
 
   setTargetDAGCombine(ISD::SDIVREM);
   setTargetDAGCombine(ISD::UDIVREM);
   setTargetDAGCombine(ISD::SELECT);
   setTargetDAGCombine(ISD::AND);
   setTargetDAGCombine(ISD::OR);
   setTargetDAGCombine(ISD::ADD);
   setTargetDAGCombine(ISD::AssertZext);
+  setTargetDAGCombine(ISD::SHL);
 
   if (ABI.IsO32()) {
     // These libcalls are not available in 32-bit.
     setLibcallName(RTLIB::SHL_I128, nullptr);
     setLibcallName(RTLIB::SRL_I128, nullptr);
     setLibcallName(RTLIB::SRA_I128, nullptr);
   }
 
static SDValue performCMovFPCombine(SDNode *N, SelectionDAG &DAG,
   SDValue FCC = N->getOperand(1), Glue = N->getOperand(3);
   return DAG.getNode(Opc, SDLoc(N), ValueIfFalse.getValueType(),
                      ValueIfFalse, FCC, ValueIfTrue, Glue);
 }
 
 static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG,
                                  TargetLowering::DAGCombinerInfo &DCI,
                                  const MipsSubtarget &Subtarget) {
-  // Pattern match EXT.
-  //  $dst = and ((sra or srl) $src , pos), (2**size - 1)
-  //  => ext $dst, $src, size, pos
   if (DCI.isBeforeLegalizeOps() || !Subtarget.hasExtractInsert())
     return SDValue();
 
-  SDValue ShiftRight = N->getOperand(0), Mask = N->getOperand(1);
-  unsigned ShiftRightOpc = ShiftRight.getOpcode();
-
-  // Op's first operand must be a shift right.
-  if (ShiftRightOpc != ISD::SRA && ShiftRightOpc != ISD::SRL)
-    return SDValue();
+  SDValue FirstOperand = N->getOperand(0);
+  unsigned FirstOperandOpc = FirstOperand.getOpcode();
+  SDValue Mask = N->getOperand(1);
+  EVT ValTy = N->getValueType(0);
+  SDLoc DL(N);
 
-  // The second operand of the shift must be an immediate.
+  uint64_t Pos = 0, SMPos, SMSize;
   ConstantSDNode *CN;
-  if (!(CN = dyn_cast<ConstantSDNode>(ShiftRight.getOperand(1))))
-    return SDValue();
-
-  uint64_t Pos = CN->getZExtValue();
-  uint64_t SMPos, SMSize;
+  SDValue NewOperand;
+  unsigned Opc;
 
   // Op's second operand must be a shifted mask.
   if (!(CN = dyn_cast<ConstantSDNode>(Mask)) ||
       !isShiftedMask(CN->getZExtValue(), SMPos, SMSize))
     return SDValue();
 
+  if (FirstOperandOpc == ISD::SRA || FirstOperandOpc == ISD::SRL) {
+    // Pattern match EXT.
+    //  $dst = and ((sra or srl) $src , pos), (2**size - 1)
+    //  => ext $dst, $src, pos, size
+
+    // The second operand of the shift must be an immediate.
+    if (!(CN = dyn_cast<ConstantSDNode>(FirstOperand.getOperand(1))))
+      return SDValue();
+
+    Pos = CN->getZExtValue();
+
     // Return if the shifted mask does not start at bit 0 or the sum of its size
     // and Pos exceeds the word's size.
-  EVT ValTy = N->getValueType(0);
     if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits())
       return SDValue();
 
-  SDLoc DL(N);
-  return DAG.getNode(MipsISD::Ext, DL, ValTy,
-                     ShiftRight.getOperand(0),
+    Opc = MipsISD::Ext;
+    NewOperand = FirstOperand.getOperand(0);
+  } else if (FirstOperandOpc == ISD::SHL && Subtarget.hasCnMips()) {
+    // Pattern match CINS.
+    //  $dst = and (shl $src , pos), mask
+    //  => cins $dst, $src, pos, size
+    // mask is a shifted mask with consecutive 1's, pos = shift amount,
+    // size = population count.
+
+    // The second operand of the shift must be an immediate.
+    if (!(CN = dyn_cast<ConstantSDNode>(FirstOperand.getOperand(1))))
+      return SDValue();
+
+    Pos = CN->getZExtValue();
+
+    if (SMPos != Pos || Pos >= ValTy.getSizeInBits() || SMSize >= 32 ||
+        Pos + SMSize > ValTy.getSizeInBits())
+      return SDValue();
+
+    NewOperand = FirstOperand.getOperand(0);
+    // SMSize is 'location' (position) in this case, not size.
+    SMSize--;
+    Opc = MipsISD::CIns;
+  } else {
+    // Pattern match EXT.
+    //  $dst = and $src, (2**size - 1) , if size > 16
+    //  => ext $dst, $src, pos, size , pos = 0
+
+    // If the mask is <= 0xffff, andi can be used instead.
+    if (CN->getZExtValue() <= 0xffff)
+      return SDValue();
+
+    // Return if the mask doesn't start at position 0.
+    if (SMPos)
+      return SDValue();
+
+    Opc = MipsISD::Ext;
+    NewOperand = FirstOperand;
+  }
+  return DAG.getNode(Opc, DL, ValTy, NewOperand,
                      DAG.getConstant(Pos, DL, MVT::i32),
                      DAG.getConstant(SMSize, DL, MVT::i32));
 }
 
 static SDValue performORCombine(SDNode *N, SelectionDAG &DAG,
                                 TargetLowering::DAGCombinerInfo &DCI,
                                 const MipsSubtarget &Subtarget) {
   // Pattern match INS.
SDValue NewAssertZext = DAG.getNode(
         WiderAssertZext.getOperand(0), DAG.getValueType(NarrowerVT));
     return DAG.getNode(ISD::TRUNCATE, SDLoc(N), N->getValueType(0),
                        NewAssertZext);
   }
 
   return SDValue();
 }
 
+
+static SDValue performSHLCombine(SDNode *N, SelectionDAG &DAG,
+                                 TargetLowering::DAGCombinerInfo &DCI,
+                                 const MipsSubtarget &Subtarget) {
+  // Pattern match CINS.
+  //  $dst = shl (and $src , imm), pos
+  //  => cins $dst, $src, pos, size
+
+  if (DCI.isBeforeLegalizeOps() || !Subtarget.hasCnMips())
+    return SDValue();
+
+  SDValue FirstOperand = N->getOperand(0);
+  unsigned FirstOperandOpc = FirstOperand.getOpcode();
+  SDValue SecondOperand = N->getOperand(1);
+  EVT ValTy = N->getValueType(0);
+  SDLoc DL(N);
+
+  uint64_t Pos = 0, SMPos, SMSize;
+  ConstantSDNode *CN;
+  SDValue NewOperand;
+
+  // The second operand of the shift must be an immediate.
+  if (!(CN = dyn_cast<ConstantSDNode>(SecondOperand)))
+    return SDValue();
+
+  Pos = CN->getZExtValue();
+
+  if (Pos >= ValTy.getSizeInBits())
+    return SDValue();
+
+  if (FirstOperandOpc != ISD::AND)
+    return SDValue();
+
+  // AND's second operand must be a shifted mask.
+  if (!(CN = dyn_cast<ConstantSDNode>(FirstOperand.getOperand(1))) ||
+      !isShiftedMask(CN->getZExtValue(), SMPos, SMSize))
+    return SDValue();
+
+  // Return if the shifted mask does not start at bit 0 or the sum of its size
+  // and Pos exceeds the word's size.
+  if (SMPos != 0 || SMSize > 32 || Pos + SMSize > ValTy.getSizeInBits())
+    return SDValue();
+
+  NewOperand = FirstOperand.getOperand(0);
+  // SMSize is 'location' (position) in this case, not size.
+  SMSize--;
+
+  return DAG.getNode(MipsISD::CIns, DL, ValTy, NewOperand,
+                     DAG.getConstant(Pos, DL, MVT::i32),
+                     DAG.getConstant(SMSize, DL, MVT::i32));
+}
+
 SDValue  MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI)
   const {
   SelectionDAG &DAG = DCI.DAG;
   unsigned Opc = N->getOpcode();
 
   switch (Opc) {
   default: break;
   case ISD::SDIVREM:
   case ISD::UDIVREM:
     return performDivRemCombine(N, DAG, DCI, Subtarget);
   case ISD::SELECT:
     return performSELECTCombine(N, DAG, DCI, Subtarget);
   case MipsISD::CMovFP_F:
   case MipsISD::CMovFP_T:
     return performCMovFPCombine(N, DAG, DCI, Subtarget);
   case ISD::AND:
     return performANDCombine(N, DAG, DCI, Subtarget);
   case ISD::OR:
     return performORCombine(N, DAG, DCI, Subtarget);
   case ISD::ADD:
     return performADDCombine(N, DAG, DCI, Subtarget);
   case ISD::AssertZext:
     return performAssertZextCombine(N, DAG, DCI, Subtarget);
+  case ISD::SHL:
+    return performSHLCombine(N, DAG, DCI, Subtarget);
   }
 
   return SDValue();
 }
 
 bool MipsTargetLowering::isCheapToSpeculateCttz() const {
   return Subtarget.hasMips32();
 }

lib/Target/Mips/MipsInstrInfo.td

 // This instruction is illegal since movn can take only register operands.
 
 def MipsWrapper    : SDNode<"MipsISD::Wrapper", SDTIntBinOp>;
 
 def MipsSync : SDNode<"MipsISD::Sync", SDT_Sync, [SDNPHasChain,SDNPSideEffect]>;
 
 def MipsExt :  SDNode<"MipsISD::Ext", SDT_Ext>;
 def MipsIns :  SDNode<"MipsISD::Ins", SDT_Ins>;
+def MipsCIns : SDNode<"MipsISD::CIns", SDT_Ext>;
 
 def MipsLWL : SDNode<"MipsISD::LWL", SDTMipsLoadLR,
                      [SDNPHasChain, SDNPMayLoad, SDNPMemOperand]>;
 def MipsLWR : SDNode<"MipsISD::LWR", SDTMipsLoadLR,
                      [SDNPHasChain, SDNPMayLoad, SDNPMemOperand]>;
 def MipsSWL : SDNode<"MipsISD::SWL", SDTStore,
                      [SDNPHasChain, SDNPMayStore, SDNPMemOperand]>;
 def MipsSWR : SDNode<"MipsISD::SWR", SDTStore,

 }]>;
 def immZExt5Plus32 : PatLeaf<(imm), [{
   return isUInt<5>(N->getZExtValue() - 32);
 }]>;
 def immZExt5Plus33 : PatLeaf<(imm), [{
   return isUInt<5>(N->getZExtValue() - 33);
 }]>;
 
+def immZExt5To31 : SDNodeXForm<imm, [{
+  return getImm(N, 31 - N->getZExtValue());
+}]>;
+
 // True if (N + 1) fits in 16-bit field.
 def immSExt16Plus1 : PatLeaf<(imm), [{
   return isInt<17>(N->getSExtValue()) && isInt<16>(N->getSExtValue() + 1);
 }]>;
 
 // Mips Address Mode! SDNode frameindex could possibily be a match
 // since load and store instructions from stack used it.
 def addr :

lib/Target/Mips/MipsSEISelLowering.cpp

MipsSETargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
   case ISD::OR:
     Val = performORCombine(N, DAG, DCI, Subtarget);
     break;
   case ISD::SUBE:
     return performSUBECombine(N, DAG, DCI, Subtarget);
   case ISD::MUL:
     return performMULCombine(N, DAG, DCI, this);
   case ISD::SHL:
-    return performSHLCombine(N, DAG, DCI, Subtarget);
+    Val = performSHLCombine(N, DAG, DCI, Subtarget);
+    break;
   case ISD::SRA:
     return performSRACombine(N, DAG, DCI, Subtarget);
   case ISD::SRL:
     return performSRLCombine(N, DAG, DCI, Subtarget);
   case ISD::VSELECT:
     return performVSELECTCombine(N, DAG);
   case ISD::XOR:
     Val = performXORCombine(N, DAG, Subtarget);

test/CodeGen/Mips/cins.ll

+; RUN: llc -march=mips64 -mcpu=octeon -target-abi=n64 < %s -o - | FileCheck %s
+
+define i64 @cins_zext(i32 signext %n) {
+entry:
+  %shl = shl i32 %n, 5
+  %conv = zext i32 %shl to i64
+  ret i64 %conv
+
+; CHECK-LABEL: cins_zext:
+; CHECK:       cins $[[R0:[0-9]+]], $[[R1:[0-9]+]], 5, 26
+
+}
+
+define i64 @cins_and_shl(i64 zeroext %n) {
+entry:
+  %and = shl i64 %n, 8
+  %shl = and i64 %and, 16776960
+  ret i64 %shl
+
+; CHECK-LABEL: cins_and_shl:
+; CHECK:       cins $[[R0:[0-9]+]], $[[R1:[0-9]+]], 8, 15
+
+}
+
+define i64 @cins_and_shl32(i64 zeroext %n) {
+entry:
+  %and = shl i64 %n, 38
+  %shl = and i64 %and, 18014123631575040
+  ret i64 %shl
+
+; CHECK-LABEL: cins_and_shl32:
+; CHECK:       cins32 $[[R0:[0-9]+]], $[[R1:[0-9]+]], 6, 15
+
+}
+
+define zeroext i16 @cins_and_shl_16(i16 zeroext %n) {
+entry:
+  %0 = shl i16 %n, 2
+  %1 = and i16 %0, 60
+  ret i16 %1
+
+; CHECK-LABEL: cins_and_shl_16:
+; CHECK:       cins $[[R0:[0-9]+]], $[[R1:[0-9]+]], 2, 3
+
+}
+
+define zeroext i8 @cins_and_shl_8(i8 zeroext %n) {
+entry:
+  %0 = shl i8 %n, 2
+  %1 = and i8 %0, 12
+  ret i8 %1
+
+; CHECK-LABEL: cins_and_shl_8:
+; CHECK:       cins $[[R0:[0-9]+]], $[[R1:[0-9]+]], 2, 1
+
+}
+
+define i32 @cins_i32(i32 signext %a) {
+entry:
+  %and = shl i32 %a, 17
+  %shl = and i32 %and, 536739840
+  ret i32 %shl
+
+; CHECK-LABEL: cins_i32:
+; CHECK:       cins $[[R0:[0-9]+]], $[[R1:[0-9]+]], 17, 11
+
+}
+
+define i64 @cins_shl_and(i32 signext %n) {
+entry:
+  %and = and i32 %n, 65535
+  %conv = zext i32 %and to i64
+  %shl = shl nuw nsw i64 %conv, 31
+  ret i64 %shl
+
+; CHECK-LABEL: cins_shl_and:
+; CHECK:       cins $[[R0:[0-9]+]], $[[R1:[0-9]+]], 31, 15
+
+}
+
+
+define i64 @cins_shl_and32(i32 signext %n) {
+entry:
+  %and = and i32 %n, 65535
+  %conv = zext i32 %and to i64
+  %shl = shl nuw nsw i64 %conv, 47
+  ret i64 %shl
+
+; CHECK-LABEL: cins_shl_and32:
+; CHECK:       cins32 $[[R0:[0-9]+]], $[[R1:[0-9]+]], 15, 15
+
+}

test/CodeGen/Mips/dext.ll

+; RUN: llc -march=mips64 -mcpu=mips64r2 -target-abi=n64 < %s -o - | FileCheck %s
+
+define i64 @dext_add_zext(i32 signext %n) {
+entry:
+  %add = add i32 %n, 1
+  %res = zext i32 %add to i64
+  ret i64 %res
+
+; CHECK-LABEL: dext_add_zext:
+; CHECK:       dext $[[R0:[0-9]+]], $[[R0:[0-9]+]], 0, 32
+
+}
+
+define i32 @ext_and24(i32 signext %a) {
+entry:
+  %and = and i32 %a, 16777215
+  ret i32 %and
+
+; CHECK-LABEL: ext_and24:
+; CHECK:       ext $[[R0:[0-9]+]], $[[R1:[0-9]+]], 0, 24
+
+}
+
+define i64 @dext_and32(i64 zeroext %a) {
+entry:
+  %and = and i64 %a, 4294967295
+  ret i64 %and
+
+; CHECK-LABEL: dext_and32:
+; CHECK:       dext $[[R0:[0-9]+]], $[[R1:[0-9]+]], 0, 32
+
+}
+
+define i64 @dext_and35(i64 zeroext %a) {
+entry:
+  %and = and i64 %a, 34359738367
+  ret i64 %and
+
+; CHECK-LABEL: dext_and35:
+; CHECK:       dextm $[[R0:[0-9]+]], $[[R1:[0-9]+]], 0, 35
+
+}
+
+define i64 @dext_and20(i64 zeroext %a) {
+entry:
+  %and = and i64 %a, 1048575
+  ret i64 %and
+
+; CHECK-LABEL: dext_and20:
+; CHECK:       dext $[[R0:[0-9]+]], $[[R1:[0-9]+]], 0, 20
+
+}
+
+define i64 @dext_and16(i64 zeroext %a) {
+entry:
+  %and = and i64 %a, 65535
+  ret i64 %and
+
+; CHECK-LABEL: dext_and16:
+; CHECK:       andi $[[R0:[0-9]+]], $[[R1:[0-9]+]], 65535
+
+}
+
+define i64 @dext_lsr_and20(i64 zeroext %a) {
+entry:
+  %shr = lshr i64 %a, 5
+  %and = and i64 %shr, 1048575
+  ret i64 %and
+
+; CHECK-LABEL: dext_lsr_and20:
+; CHECK:       dext $[[R0:[0-9]+]], $[[R1:[0-9]+]], 5, 20
+
+}
+
+define i64 @dext_lsr_and8(i64 zeroext %a) {
+entry:
+  %shr = lshr i64 %a, 40
+  %and = and i64 %shr, 255
+  ret i64 %and
+
+; CHECK-LABEL: dext_lsr_and8:
+; CHECK:       dextu $[[R0:[0-9]+]], $[[R1:[0-9]+]], 40, 8
+
+}
+
+define i64 @dext_zext(i32 signext %a) {
+entry:
+  %conv = zext i32 %a to i64
+  ret i64 %conv
+
+; CHECK-LABEL: dext_zext:
+; CHECK:       dext $[[R0:[0-9]+]], $[[R1:[0-9]+]], 0, 32
+
+}
+
+define i64 @dext_and_lsr(i64 zeroext %n) {
+entry:
+  %and = lshr i64 %n, 8
+  %shr = and i64 %and, 4095
+  ret i64 %shr
+
+; CHECK-LABEL: dext_and_lsr:
+; CHECK:       dext $[[R0:[0-9]+]], $[[R1:[0-9]+]], 8, 12
+
+}

test/CodeGen/Mips/load-store-left-right.ll

 ; RUN: llc -march=mipsel   -mcpu=mips32              -relocation-model=pic < %s | FileCheck -check-prefixes=ALL,MIPS32,MIPS32-EL %s
 ; RUN: llc -march=mips     -mcpu=mips32              -relocation-model=pic < %s | FileCheck -check-prefixes=ALL,MIPS32,MIPS32-EB %s
 ; RUN: llc -march=mipsel   -mcpu=mips32r2            -relocation-model=pic < %s | FileCheck -check-prefixes=ALL,MIPS32,MIPS32-EL %s
 ; RUN: llc -march=mips     -mcpu=mips32r2            -relocation-model=pic < %s | FileCheck -check-prefixes=ALL,MIPS32,MIPS32-EB %s
 ; RUN: llc -march=mipsel   -mcpu=mips32r6            -relocation-model=pic < %s | FileCheck -check-prefixes=ALL,MIPS32R6,MIPS32R6-EL %s
 ; RUN: llc -march=mips     -mcpu=mips32r6            -relocation-model=pic < %s | FileCheck -check-prefixes=ALL,MIPS32R6,MIPS32R6-EB %s
 ; RUN: llc -march=mips64el -mcpu=mips4    -target-abi=n64 -relocation-model=pic < %s | FileCheck -check-prefixes=ALL,MIPS64,MIPS64-EL %s
 ; RUN: llc -march=mips64   -mcpu=mips4    -target-abi=n64 -relocation-model=pic < %s | FileCheck -check-prefixes=ALL,MIPS64,MIPS64-EB %s
 ; RUN: llc -march=mips64el -mcpu=mips64   -target-abi=n64 -relocation-model=pic < %s | FileCheck -check-prefixes=ALL,MIPS64,MIPS64-EL %s
 ; RUN: llc -march=mips64   -mcpu=mips64   -target-abi=n64 -relocation-model=pic < %s | FileCheck -check-prefixes=ALL,MIPS64,MIPS64-EB %s
-; RUN: llc -march=mips64el -mcpu=mips64r2 -target-abi=n64 -relocation-model=pic < %s | FileCheck -check-prefixes=ALL,MIPS64,MIPS64-EL %s
-; RUN: llc -march=mips64   -mcpu=mips64r2 -target-abi=n64 -relocation-model=pic < %s | FileCheck -check-prefixes=ALL,MIPS64,MIPS64-EB %s
+; RUN: llc -march=mips64el -mcpu=mips64r2 -target-abi=n64 -relocation-model=pic < %s | FileCheck -check-prefixes=ALL,MIPS64,MIPS64R2-EL %s
+; RUN: llc -march=mips64   -mcpu=mips64r2 -target-abi=n64 -relocation-model=pic < %s | FileCheck -check-prefixes=ALL,MIPS64,MIPS64R2-EB %s
 ; RUN: llc -march=mips64el -mcpu=mips64r6 -target-abi=n64 -relocation-model=pic < %s | FileCheck -check-prefixes=ALL,MIPS64R6,MIPS64R6-EL %s
 ; RUN: llc -march=mips64   -mcpu=mips64r6 -target-abi=n64 -relocation-model=pic < %s | FileCheck -check-prefixes=ALL,MIPS64R6,MIPS64R6-EB %s
 
 %struct.SLL = type { i64 }
 %struct.SI = type { i32 }
 %struct.SUI = type { i32 }
 
 @sll = common global %struct.SLL zeroinitializer, align 1

 ; MIPS32-EB:     lwr $[[R0]], 3($[[R1]])
 
 ; MIPS32R6:      lw $[[PTR:[0-9]+]], %got(si)(
 ; MIPS32R6:      lw $2, 0($[[PTR]])
 
 ; MIPS64-EL:     lwl $[[R0:[0-9]+]], 3($[[R1:[0-9]+]])
 ; MIPS64-EL:     lwr $[[R0]], 0($[[R1]])
 
+; MIPS64R2-EL:   lwl $[[R0:[0-9]+]], 3($[[R1:[0-9]+]])
+; MIPS64R2-EL:   lwr $[[R0]], 0($[[R1]])
+
 ; MIPS64-EB:     lwl $[[R0:[0-9]+]], 0($[[R1:[0-9]+]])
 ; MIPS64-EB:     lwr $[[R0]], 3($[[R1]])
 
+; MIPS64R2-EB:   lwl $[[R0:[0-9]+]], 0($[[R1:[0-9]+]])
+; MIPS64R2-EB:   lwr $[[R0]], 3($[[R1]])
+
 ; MIPS64R6:      ld $[[PTR:[0-9]+]], %got_disp(si)(
 ; MIPS64R6:      lw $2, 0($[[PTR]])
 
   %0 = load i32, i32* getelementptr inbounds (%struct.SI, %struct.SI* @si, i32 0, i32 0), align 1
   ret i32 %0
 }
 
 define void @store_SI(i32 signext %a) nounwind {
 entry:
 ; ALL-LABEL: store_SI:
 
 ; MIPS32-EL:     swl $[[R0:[0-9]+]], 3($[[R1:[0-9]+]])
 ; MIPS32-EL:     swr $[[R0]], 0($[[R1]])
 
 ; MIPS32-EB:     swl $[[R0:[0-9]+]], 0($[[R1:[0-9]+]])
 ; MIPS32-EB:     swr $[[R0]], 3($[[R1]])
 
 ; MIPS32R6:      lw $[[PTR:[0-9]+]], %got(si)(
 ; MIPS32R6:      sw $4, 0($[[PTR]])
 
 ; MIPS64-EL:     swl $[[R0:[0-9]+]], 3($[[R1:[0-9]+]])
 ; MIPS64-EL:     swr $[[R0]], 0($[[R1]])
 
+; MIPS64R2-EL:   swl $[[R0:[0-9]+]], 3($[[R1:[0-9]+]])
+; MIPS64R2-EL:   swr $[[R0]], 0($[[R1]])
+
 ; MIPS64-EB:     swl $[[R0:[0-9]+]], 0($[[R1:[0-9]+]])
 ; MIPS64-EB:     swr $[[R0]], 3($[[R1]])
 
+; MIPS64R2-EB:   swl $[[R0:[0-9]+]], 0($[[R1:[0-9]+]])
+; MIPS64R2-EB:   swr $[[R0]], 3($[[R1]])
+
 ; MIPS64R6:      ld $[[PTR:[0-9]+]], %got_disp(si)(
 ; MIPS64R6:      sw $4, 0($[[PTR]])
 
   store i32 %a, i32* getelementptr inbounds (%struct.SI, %struct.SI* @si, i32 0, i32 0), align 1
   ret void
 }
 
 define i64 @load_SLL() nounwind readonly {

 
 ; MIPS32R6:      lw $[[PTR:[0-9]+]], %got(sll)(
 ; MIPS32R6-DAG:  lw $2, 0($[[PTR]])
 ; MIPS32R6-DAG:  lw $3, 4($[[PTR]])
 
 ; MIPS64-EL:     ldl $[[R0:[0-9]+]], 7($[[R1:[0-9]+]])
 ; MIPS64-EL:     ldr $[[R0]], 0($[[R1]])
 
+; MIPS64R2-EL:   ldl $[[R0:[0-9]+]], 7($[[R1:[0-9]+]])
+; MIPS64R2-EL:   ldr $[[R0]], 0($[[R1]])
+
 ; MIPS64-EB:     ldl $[[R0:[0-9]+]], 0($[[R1:[0-9]+]])
 ; MIPS64-EB:     ldr $[[R0]], 7($[[R1]])
 
+; MIPS64R2-EB:   ldl $[[R0:[0-9]+]], 0($[[R1:[0-9]+]])
+; MIPS64R2-EB:   ldr $[[R0]], 7($[[R1]])
+
 ; MIPS64R6:      ld $[[PTR:[0-9]+]], %got_disp(sll)(
 ; MIPS64R6:      ld $2, 0($[[PTR]])
 
   %0 = load i64, i64* getelementptr inbounds (%struct.SLL, %struct.SLL* @sll, i64 0, i32 0), align 1
   ret i64 %0
 }
 
 define i64 @load_SI_sext_to_i64() nounwind readonly {

 ; MIPS32R6-EL:   lw $2, 0($[[PTR]])
 ; MIPS32R6-EL:   sra $3, $2, 31
 ; MIPS32R6-EB:   lw $3, 0($[[PTR]])
 ; MIPS32R6-EB:   sra $2, $3, 31
 
 ; MIPS64-EL:     lwl $[[R0:[0-9]+]], 3($[[R1:[0-9]+]])
 ; MIPS64-EL:     lwr $[[R0]], 0($[[R1]])
 
+; MIPS64R2-EL:   lwl $[[R0:[0-9]+]], 3($[[R1:[0-9]+]])
+; MIPS64R2-EL:   lwr $[[R0]], 0($[[R1]])
+
 ; MIPS64-EB:     lwl $[[R0:[0-9]+]], 0($[[R1:[0-9]+]])
 ; MIPS64-EB:     lwr $[[R0]], 3($[[R1]])
 
+; MIPS64R2-EB:   lwl $[[R0:[0-9]+]], 0($[[R1:[0-9]+]])
+; MIPS64R2-EB:   lwr $[[R0]], 3($[[R1]])
+
 ; MIPS64R6:      ld $[[PTR:[0-9]+]], %got_disp(si)(
 ; MIPS64R6:      lw $2, 0($[[PTR]])
 
   %0 = load i32, i32* getelementptr inbounds (%struct.SI, %struct.SI* @si, i64 0, i32 0), align 1
   %conv = sext i32 %0 to i64
   ret i64 %conv
 }
 

 
 ; MIPS64-EL-DAG: lwl $[[R0:[0-9]+]], 3($[[R1:[0-9]+]])
 ; MIPS64-EL-DAG: lwr $[[R0]], 0($[[R1]])
 ; MIPS64-EL-DAG: daddiu $[[R2:[0-9]+]], $zero, 1
 ; MIPS64-EL-DAG: dsll   $[[R3:[0-9]+]], $[[R2]], 32
 ; MIPS64-EL-DAG: daddiu $[[R4:[0-9]+]], $[[R3]], -1
 ; MIPS64-EL-DAG: and    ${{[0-9]+}}, $[[R0]], $[[R4]]
 
+; MIPS64R2-EL: lwl $[[R0:[0-9]+]], 3($[[R1:[0-9]+]])
+; MIPS64R2-EL: lwr $[[R0]], 0($[[R1]])
+; MIPS64R2-EL: dext $[[R0]], $[[R0]], 0, 32
+
 ; MIPS64-EB:     lwl $[[R0:[0-9]+]], 0($[[R1:[0-9]+]])
 ; MIPS64-EB:     lwr $[[R0]], 3($[[R1]])
 
+; MIPS64R2-EB-DAG:   lwl $[[R0:[0-9]+]], 0($[[R1:[0-9]+]])
+; MIPS64R2-EB-DAG:   lwr $[[R0]], 3($[[R1]])
+; MIPS64R2-EB-DAG:   dext $[[R0]], $[[R0]], 0, 32
+
 ; MIPS64R6:      ld $[[PTR:[0-9]+]], %got_disp(sui)(
 ; MIPS64R6:      lwu $2, 0($[[PTR]])
 
   %0 = load i32, i32* getelementptr inbounds (%struct.SUI, %struct.SUI* @sui, i64 0, i32 0), align 1
   %conv = zext i32 %0 to i64
   ret i64 %conv
 }
 

 
 ; MIPS32R6-DAG:  lw $[[PTR:[0-9]+]], %got(sll)(
 ; MIPS32R6-DAG:  sw $4, 0($[[PTR]])
 ; MIPS32R6-DAG:  sw $5, 4($[[PTR]])
 
 ; MIPS64-EL:     sdl $[[R0:[0-9]+]], 7($[[R1:[0-9]+]])
 ; MIPS64-EL:     sdr $[[R0]], 0($[[R1]])
 
+; MIPS64R2-EL:   sdl $[[R0:[0-9]+]], 7($[[R1:[0-9]+]])
+; MIPS64R2-EL:   sdr $[[R0]], 0($[[R1]])
+
 ; MIPS64-EB:     sdl $[[R0:[0-9]+]], 0($[[R1:[0-9]+]])
 ; MIPS64-EB:     sdr $[[R0]], 7($[[R1]])
 
+; MIPS64R2-EB:   sdl $[[R0:[0-9]+]], 0($[[R1:[0-9]+]])
+; MIPS64R2-EB:   sdr $[[R0]], 7($[[R1]])
+
 ; MIPS64R6:      ld $[[PTR:[0-9]+]], %got_disp(sll)(
 ; MIPS64R6:      sd $4, 0($[[PTR]])
 
   store i64 %a, i64* getelementptr inbounds (%struct.SLL, %struct.SLL* @sll, i64 0, i32 0), align 1
   ret void
 }
 
 define void @store_SI_trunc_from_i64(i32 signext %a) nounwind {
 entry:
 ; ALL-LABEL: store_SI_trunc_from_i64:
 
 ; MIPS32-EL:     swl $[[R0:[0-9]+]], 3($[[R1:[0-9]+]])
 ; MIPS32-EL:     swr $[[R0]], 0($[[R1]])
 
 ; MIPS32-EB:     swl $[[R0:[0-9]+]], 0($[[R1:[0-9]+]])
 ; MIPS32-EB:     swr $[[R0]], 3($[[R1]])
 
 ; MIPS32R6:      lw $[[PTR:[0-9]+]], %got(si)(
 ; MIPS32R6:      sw $4, 0($[[PTR]])
 
 ; MIPS64-EL:     swl $[[R0:[0-9]+]], 3($[[R1:[0-9]+]])
 ; MIPS64-EL:     swr $[[R0]], 0($[[R1]])
 
+; MIPS64R2-EL:   swl $[[R0:[0-9]+]], 3($[[R1:[0-9]+]])
+; MIPS64R2-EL:   swr $[[R0]], 0($[[R1]])
+
 ; MIPS64-EB:     swl $[[R0:[0-9]+]], 0($[[R1:[0-9]+]])
 ; MIPS64-EB:     swr $[[R0]], 3($[[R1]])
 
+; MIPS64R2-EB:   swl $[[R0:[0-9]+]], 0($[[R1:[0-9]+]])
+; MIPS64R2-EB:   swr $[[R0]], 3($[[R1]])
+
 ; MIPS64R6:      ld $[[PTR:[0-9]+]], %got_disp(si)(
 ; MIPS64R6:      sw $4, 0($[[PTR]])
 
   store i32 %a, i32* getelementptr inbounds (%struct.SI, %struct.SI* @si, i64 0, i32 0), align 1
   ret void
 }
 
 ;

 define void @copy_struct_S0() nounwind {
 entry:
 ; ALL-LABEL: copy_struct_S0:
 
 ; MIPS32-EL:     lw $[[PTR:[0-9]+]], %got(struct_s0)(
 ; MIPS32-EB:     lw $[[PTR:[0-9]+]], %got(struct_s0)(
 ; MIPS32R6:      lw $[[PTR:[0-9]+]], %got(struct_s0)(
 ; MIPS64-EL:     ld $[[PTR:[0-9]+]], %got_disp(struct_s0)(
+; MIPS64R2-EL:   ld $[[PTR:[0-9]+]], %got_disp(struct_s0)(
 ; MIPS64-EB:     ld $[[PTR:[0-9]+]], %got_disp(struct_s0)(
+; MIPS64R2-EB:   ld $[[PTR:[0-9]+]], %got_disp(struct_s0)(
 ; MIPS64R6:      ld $[[PTR:[0-9]+]], %got_disp(struct_s0)(
 
 ; FIXME: We should be able to do better than this on MIPS32r6/MIPS64r6 since
 ;        we have unaligned halfword load/store available
 ; ALL-DAG:       lbu $[[R1:[0-9]+]], 0($[[PTR]])
 ; ALL-DAG:       sb $[[R1]], 2($[[PTR]])
 ; ALL-DAG:       lbu $[[R1:[0-9]+]], 1($[[PTR]])
 ; ALL-DAG:       sb $[[R1]], 3($[[PTR]])

 
 ; MIPS32R6:      lw $[[PTR:[0-9]+]], %got(struct_s1)(
 ; MIPS32R6-DAG:  lhu $[[R1:[0-9]+]], 0($[[PTR]])
 ; MIPS32R6-DAG:  sh $[[R1]], 4($[[PTR]])
 ; MIPS32R6-DAG:  lhu $[[R1:[0-9]+]], 2($[[PTR]])
 ; MIPS32R6-DAG:  sh $[[R1]], 6($[[PTR]])
 
 ; MIPS64-EL:     ld $[[PTR:[0-9]+]], %got_disp(struct_s1)(
+; MIPS64R2-EL:   ld $[[PTR:[0-9]+]], %got_disp(struct_s1)(
 ; MIPS64-EB:     ld $[[PTR:[0-9]+]], %got_disp(struct_s1)(
+; MIPS64R2-EB:   ld $[[PTR:[0-9]+]], %got_disp(struct_s1)(
 ; MIPS64-DAG:    lbu $[[R1:[0-9]+]], 0($[[PTR]])
 ; MIPS64-DAG:    sb $[[R1]], 4($[[PTR]])
 ; MIPS64-DAG:    lbu $[[R1:[0-9]+]], 1($[[PTR]])
 ; MIPS64-DAG:    sb $[[R1]], 5($[[PTR]])
 ; MIPS64-DAG:    lbu $[[R1:[0-9]+]], 2($[[PTR]])
 ; MIPS64-DAG:    sb $[[R1]], 6($[[PTR]])
 ; MIPS64-DAG:    lbu $[[R1:[0-9]+]], 3($[[PTR]])
 ; MIPS64-DAG:    sb $[[R1]], 7($[[PTR]])

 ; MIPS64-EL-DAG: lwr $[[R1]],        0($[[PTR]])
 ; MIPS64-EL-DAG: swl $[[R1]],       11($[[PTR]])
 ; MIPS64-EL-DAG: swr $[[R1]],        8($[[PTR]])
 ; MIPS64-EL-DAG: lwl $[[R1:[0-9]+]], 7($[[PTR]])
 ; MIPS64-EL-DAG: lwr $[[R1]],        4($[[PTR]])
 ; MIPS64-EL-DAG: swl $[[R1]],       15($[[PTR]])
 ; MIPS64-EL-DAG: swr $[[R1]],       12($[[PTR]])
 
+; MIPS64R2-EL:     ld $[[PTR:[0-9]+]], %got_disp(struct_s2)(
+; MIPS64R2-EL-DAG: lwl $[[R1:[0-9]+]], 3($[[PTR]])
+; MIPS64R2-EL-DAG: lwr $[[R1]],        0($[[PTR]])
+; MIPS64R2-EL-DAG: swl $[[R1]],       11($[[PTR]])
+; MIPS64R2-EL-DAG: swr $[[R1]],        8($[[PTR]])
+; MIPS64R2-EL-DAG: lwl $[[R1:[0-9]+]], 7($[[PTR]])
+; MIPS64R2-EL-DAG: lwr $[[R1]],        4($[[PTR]])
+; MIPS64R2-EL-DAG: swl $[[R1]],       15($[[PTR]])
+; MIPS64R2-EL-DAG: swr $[[R1]],       12($[[PTR]])
+
 ; MIPS64-EB:     ld $[[PTR:[0-9]+]], %got_disp(struct_s2)(
 ; MIPS64-EB-DAG: lwl $[[R1:[0-9]+]], 0($[[PTR]])
 ; MIPS64-EB-DAG: lwr $[[R1]],        3($[[PTR]])
 ; MIPS64-EB-DAG: swl $[[R1]],        8($[[PTR]])
 ; MIPS64-EB-DAG: swr $[[R1]],       11($[[PTR]])
 ; MIPS64-EB-DAG: lwl $[[R1:[0-9]+]], 4($[[PTR]])
 ; MIPS64-EB-DAG: lwr $[[R1]],        7($[[PTR]])
 ; MIPS64-EB-DAG: swl $[[R1]],       12($[[PTR]])
 ; MIPS64-EB-DAG: swr $[[R1]],       15($[[PTR]])
 
+; MIPS64R2-EB:     ld $[[PTR:[0-9]+]], %got_disp(struct_s2)(
+; MIPS64R2-EB-DAG: lwl $[[R1:[0-9]+]], 0($[[PTR]])
+; MIPS64R2-EB-DAG: lwr $[[R1]],        3($[[PTR]])
+; MIPS64R2-EB-DAG: swl $[[R1]],        8($[[PTR]])
+; MIPS64R2-EB-DAG: swr $[[R1]],       11($[[PTR]])
+; MIPS64R2-EB-DAG: lwl $[[R1:[0-9]+]], 4($[[PTR]])
+; MIPS64R2-EB-DAG: lwr $[[R1]],        7($[[PTR]])
+; MIPS64R2-EB-DAG: swl $[[R1]],       12($[[PTR]])
+; MIPS64R2-EB-DAG: swr $[[R1]],       15($[[PTR]])
+
 ; MIPS64R6:      ld $[[PTR:[0-9]+]], %got_disp(struct_s2)(
 ; MIPS64R6-DAG:  lw $[[R1:[0-9]+]], 0($[[PTR]])
 ; MIPS64R6-DAG:  sw $[[R1]],        8($[[PTR]])
 ; MIPS64R6-DAG:  lw $[[R1:[0-9]+]], 4($[[PTR]])
 ; MIPS64R6-DAG:  sw $[[R1]],       12($[[PTR]])
 
   %0 = load %struct.S2, %struct.S2* getelementptr inbounds (%struct.S2, %struct.S2* @struct_s2, i32 0), align 1
   store %struct.S2 %0, %struct.S2* getelementptr inbounds (%struct.S2, %struct.S2* @struct_s2, i32 1), align 1

 ; MIPS32R6-EB-DAG: lbu $[[R3:[0-9]+]], 6($[[PTR]])
 ; MIPS32R6-EB-DAG: sll $[[T0:[0-9]+]], $[[R2]], 16
 ; MIPS32R6-EB-DAG: or  $5, $[[T0]], $[[R3]]
 
 ; MIPS64-EL:     ld $[[SPTR:[0-9]+]], %got_disp(arr)(
 ; MIPS64-EL-DAG: lwl $[[R1:[0-9]+]], 3($[[PTR]])
 ; MIPS64-EL-DAG: lwr $[[R1]],   0($[[PTR]])
 
+; MIPS64R2-EL:     ld $[[SPTR:[0-9]+]], %got_disp(arr)(
+; MIPS64R2-EL-DAG: lwl $[[R1:[0-9]+]], 3($[[PTR]])
+; MIPS64R2-EL-DAG: lwr $[[R1]],   0($[[PTR]])
+
 ; MIPS64-EB:     ld $[[SPTR:[0-9]+]], %got_disp(arr)(
 ; MIPS64-EB-DAG: lwl  $[[R1:[0-9]+]], 0($[[PTR]])
 ; MIPS64-EB-DAG: lwr  $[[R1]],   3($[[PTR]])
 ; MIPS64-EB-DAG: dsll $[[R1]], $[[R1]], 32
 ; MIPS64-EB-DAG: lbu  $[[R2:[0-9]+]], 5($[[PTR]])
 ; MIPS64-EB-DAG: lbu  $[[R3:[0-9]+]], 4($[[PTR]])
 ; MIPS64-EB-DAG: dsll $[[T0:[0-9]+]], $[[R3]], 8
 ; MIPS64-EB-DAG: or   $[[T1:[0-9]+]], $[[T0]], $[[R2]]
 ; MIPS64-EB-DAG: dsll $[[T1]], $[[T1]], 16
 ; MIPS64-EB-DAG: or   $[[T3:[0-9]+]], $[[R1]], $[[T1]]
 ; MIPS64-EB-DAG: lbu  $[[R4:[0-9]+]], 6($[[PTR]])
 ; MIPS64-EB-DAG: dsll $[[T4:[0-9]+]], $[[R4]], 8
 ; MIPS64-EB-DAG: or   $4, $[[T3]], $[[T4]]
 
+; MIPS64R2-EB: ld   $[[SPTR:[0-9]+]], %got_disp(arr)(
+; MIPS64R2-EB: lbu  $[[R1:[0-9]+]], 5($[[PTR]])
+; MIPS64R2-EB: lbu  $[[R2:[0-9]+]], 4($[[PTR]])
+; MIPS64R2-EB: dsll $[[T0:[0-9]+]], $[[R2]], 8
+; MIPS64R2-EB: or   $[[T1:[0-9]+]], $[[T0]], $[[R1]]
+; MIPS64R2-EB: dsll $[[T1]], $[[T1]], 16
+; MIPS64R2-EB: lwl  $[[R3:[0-9]+]], 0($[[PTR]])
+; MIPS64R2-EB: lwr  $[[R3]], 3($[[PTR]])
+; MIPS64R2-EB: dext $[[R3]], $[[R3]], 0, 32
+; MIPS64R2-EB: dsll $[[R3]], $[[R3]], 32
+; MIPS64R2-EB: or   $[[T2:[0-9]+]], $[[R3]], $[[T1]]
+; MIPS64R2-EB: lbu  $[[R4:[0-9]+]], 6($[[PTR]])
+; MIPS64R2-EB: dsll $[[T3:[0-9]+]], $[[R4]], 8
+; MIPS64R2-EB: or   $4, $[[T2]], $[[T3]]
+
 ; MIPS64R6:      ld $[[SPTR:[0-9]+]], %got_disp(arr)(
 
   tail call void @extern_func([7 x i8]* byval @arr) nounwind
   ret void
 }
 
 declare void @extern_func([7 x i8]* byval)

test/CodeGen/Mips/mips64-f128.ll

 ; RUN: llc -mtriple=mips64el-unknown-unknown -mcpu=mips4 -mattr=+soft-float -O1 \
 ; RUN:     -disable-mips-delay-filler -relocation-model=pic < %s | FileCheck \
-; RUN:     %s -check-prefixes=ALL,C_CC_FMT,PRER6
+; RUN:     %s -check-prefixes=ALL,C_CC_FMT,PRER6,NOT-R2R6
 ; RUN: llc -mtriple=mips64el-unknown-unknown -mcpu=mips64 -mattr=+soft-float -O1 \
 ; RUN:     -disable-mips-delay-filler -relocation-model=pic < %s | FileCheck \
-; RUN:     %s -check-prefixes=ALL,C_CC_FMT,PRER6
+; RUN:     %s -check-prefixes=ALL,C_CC_FMT,PRER6,NOT-R2R6
 ; RUN: llc -mtriple=mips64el-unknown-unknown -mcpu=mips64r2 -mattr=+soft-float \
 ; RUN:     -O1 -disable-mips-delay-filler -relocation-model=pic < %s | FileCheck \
-; RUN:     %s -check-prefixes=ALL,C_CC_FMT,PRER6
+; RUN:     %s -check-prefixes=ALL,C_CC_FMT,PRER6,R2R6
 ; RUN: llc -mtriple=mips64el-unknown-unknown -mcpu=mips64r6 -mattr=+soft-float \
 ; RUN:     -O1 -disable-mips-delay-filler -relocation-model=pic < %s | FileCheck \
-; RUN:     %s -check-prefixes=ALL,CMP_CC_FMT,R6
+; RUN:     %s -check-prefixes=ALL,CMP_CC_FMT,R6,R2R6
 
 @gld0 = external global fp128
 @gld1 = external global fp128
 @gld2 = external global fp128
 @gf1 = external global float
 @gd1 = external global double
 
 ; ALL-LABEL: addLD:

 
 define double @conv_double_LD(fp128 %a) {
 entry:
   %conv = fptrunc fp128 %a to double
   ret double %conv
 }
 
 ; ALL-LABEL:             libcall1_fabsl:
-; ALL-DAG: ld      $[[R0:[0-9]+]], 8($[[R4:[0-9]+]])
-; ALL-DAG: daddiu  $[[R1:[0-9]+]], $zero, 1
-; ALL-DAG: dsll    $[[R2:[0-9]+]], $[[R1]], 63
-; ALL-DAG: daddiu  $[[R3:[0-9]+]], $[[R2]], -1
-; ALL-DAG: and     $4, $[[R0]], $[[R3]]
-; ALL-DAG: ld      $2, 0($[[R4]])
+; NOT-R2R6-DAG: ld      $[[R0:[0-9]+]], 8($[[R4:[0-9]+]])
+; NOT-R2R6-DAG: daddiu  $[[R1:[0-9]+]], $zero, 1
+; NOT-R2R6-DAG: dsll    $[[R2:[0-9]+]], $[[R1]], 63
+; NOT-R2R6-DAG: daddiu  $[[R3:[0-9]+]], $[[R2]], -1
+; NOT-R2R6-DAG: and     $4, $[[R0]], $[[R3]]
+; NOT-R2R6-DAG: ld      $2, 0($[[R4]])
+
+; R2R6-DAG: ld     $[[R0:[0-9]+]], 0($[[R3:[0-9]+]])
+; R2R6-DAG: ld     $[[R1:[0-9]+]], 8($[[R3]])
+; R2R6-DAG: dextm  $[[R2:[0-9]+]], $[[R1]], 0, 63
 
 define fp128 @libcall1_fabsl() {
 entry:
   %0 = load fp128, fp128* @gld0, align 16
   %call = tail call fp128 @fabsl(fp128 %0) nounwind readnone
   ret fp128 %call
 }
 

 entry:
   %0 = tail call fp128 @llvm.powi.f128(fp128 %a, i32 %b)
   ret fp128 %0
 }
 
 declare fp128 @llvm.powi.f128(fp128, i32) #3
 
 ; ALL-LABEL:     libcall2_copysignl:
-; ALL-DAG: daddiu $[[R2:[0-9]+]], $zero, 1
-; ALL-DAG: dsll   $[[R3:[0-9]+]], $[[R2]], 63
-; ALL-DAG: ld     $[[R0:[0-9]+]], %got_disp(gld1)
-; ALL-DAG: ld     $[[R1:[0-9]+]], 8($[[R0]])
-; ALL-DAG: and    $[[R4:[0-9]+]], $[[R1]], $[[R3]]
-; ALL-DAG: ld     $[[R5:[0-9]+]], %got_disp(gld0)
-; ALL-DAG: ld     $[[R6:[0-9]+]], 8($[[R5]])
-; ALL-DAG: daddiu $[[R7:[0-9]+]], $[[R3]], -1
-; ALL-DAG: and    $[[R8:[0-9]+]], $[[R6]], $[[R7]]
-; ALL-DAG: or     $4, $[[R8]], $[[R4]]
-; ALL-DAG: ld     $2, 0($[[R5]])
+; ALL-DAG:      daddiu $[[R2:[0-9]+]], $zero, 1
+; ALL-DAG:      dsll   $[[R3:[0-9]+]], $[[R2]], 63
+; ALL-DAG:      ld     $[[R0:[0-9]+]], %got_disp(gld1)
+; ALL-DAG:      ld     $[[R1:[0-9]+]], 8($[[R0]])
+; ALL-DAG:      and    $[[R4:[0-9]+]], $[[R1]], $[[R3]]
+; ALL-DAG:      ld     $[[R5:[0-9]+]], %got_disp(gld0)
+; ALL-DAG:      ld     $[[R6:[0-9]+]], 8($[[R5]])
+; NOT-R2R6-DAG: daddiu $[[R7:[0-9]+]], $[[R3]], -1
+; NOT-R2R6-DAG: and    $[[R8:[0-9]+]], $[[R6]], $[[R7]]
+; NOT-R2R6-DAG: or     $4, $[[R8]], $[[R4]]
+; R2R6-DAG:     dextm  $[[R7:[0-9]+]], $[[R6]], 0, 63
+; R2R6-DAG:     or     $4, $[[R7]], $[[R4]]
+; ALL-DAG:      ld     $2, 0($[[R5]])
 
 define fp128 @libcall2_copysignl() {
 entry:
   %0 = load fp128, fp128* @gld0, align 16
   %1 = load fp128, fp128* @gld1, align 16
   %call = tail call fp128 @copysignl(fp128 %0, fp128 %1) nounwind readnone
   ret fp128 %call
 }

test/MC/Mips/sext_64_32.ll

 ; RUN: llc -march=mips64el -filetype=obj -mcpu=mips64r2 %s -o - | \
 ; RUN:   llvm-objdump -disassemble - | FileCheck %s
 
 ; Sign extend from 32 to 64 was creating nonsense opcodes
 
 ; CHECK: sll ${{[a-z0-9]+}}, ${{[a-z0-9]+}}, 0
 
 define i64 @foo(i32 %ival) nounwind readnone {
 entry:
   %conv = sext i32 %ival to i64
   ret i64 %conv
 }
 
-; CHECK: dsll32 ${{[a-z0-9]+}}, ${{[a-z0-9]+}}, 0
+; CHECK-LABEL: foo_2:
+; CHECK: dext ${{[a-z0-9]+}}, ${{[a-z0-9]+}}, 0, 32
 
 define i64 @foo_2(i32 %ival_2) nounwind readnone {
 entry:
   %conv_2 = zext i32 %ival_2 to i64
   ret i64 %conv_2
 }