[mips][msa] Added support for matching bins[lr]i.[bhwd] from normal IR (i.e. not intrinsics)

This required correcting the definition of the bins[lr]i intrinsics because
the result is also the first operand.

It also required removing the (arbitrary) check for 32-bit immediates in
MipsSEDAGToDAGISel::selectVSplat().

Currently using binsli.d with 2 bits set in the mask doesn't select binsli.d
because the constant is legalized into a ConstantPool. Similar things can
happen with binsri.d with more than 10 bits set in the mask. The resulting
code when this happens is correct but not optimal.



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@193687 91177308-0d34-0410-b5e6-96231b3b80d8

7 files changed, 163 insertions, 27 deletions

diff --git a/lib/Target/Mips/MSA.txt b/lib/Target/Mips/MSA.txt
index f7c42c06ce..802217a886 100644
--- a/lib/Target/Mips/MSA.txt
+++ b/lib/Target/Mips/MSA.txt
@@ -43,7 +43,14 @@ splati.w:
         It is not possible to emit splati.w since shf.w covers the same cases.
         shf.w will be emitted instead.
 
-copy_s.w
+copy_s.w:
         On MIPS32, the copy_u.d intrinsic will emit this instruction instead of
         copy_u.w. This is semantically equivalent since the general-purpose
         register file is 32-bits wide.
+
+binsri.[bhwd],  binsli.[bhwd]:
+        These two operations are equivalent to each other with the operands
+        swapped and condition inverted. The compiler may use either one as
+        appropriate.
+        Furthermore, the compiler may use bsel.[bhwd] for some masks that do
+        not survive the legalization process (this is a bug and will be fixed).
diff --git a/lib/Target/Mips/MipsISelDAGToDAG.cpp b/lib/Target/Mips/MipsISelDAGToDAG.cpp
index 0d239fd46d..d0a41e755e 100644
--- a/lib/Target/Mips/MipsISelDAGToDAG.cpp
+++ b/lib/Target/Mips/MipsISelDAGToDAG.cpp
@@ -149,6 +149,16 @@ bool MipsDAGToDAGISel::selectVSplatUimmPow2(SDValue N, SDValue &Imm) const {
   return false;
 }
 
+bool MipsDAGToDAGISel::selectVSplatMaskL(SDValue N, SDValue &Imm) const {
+  llvm_unreachable("Unimplemented function.");
+  return false;
+}
+
+bool MipsDAGToDAGISel::selectVSplatMaskR(SDValue N, SDValue &Imm) const {
+  llvm_unreachable("Unimplemented function.");
+  return false;
+}
+
 /// Select instructions not customized! Used for
 /// expanded, promoted and normal instructions
 SDNode* MipsDAGToDAGISel::Select(SDNode *Node) {
diff --git a/lib/Target/Mips/MipsISelDAGToDAG.h b/lib/Target/Mips/MipsISelDAGToDAG.h
index e5695c41c4..6bc96446d3 100644
--- a/lib/Target/Mips/MipsISelDAGToDAG.h
+++ b/lib/Target/Mips/MipsISelDAGToDAG.h
@@ -96,6 +96,12 @@ private:
   virtual bool selectVSplatSimm5(SDValue N, SDValue &Imm) const;
   /// \brief Select constant vector splats whose value is a power of 2.
   virtual bool selectVSplatUimmPow2(SDValue N, SDValue &Imm) const;
+  /// \brief Select constant vector splats whose value is a run of set bits
+  /// ending at the most significant bit
+  virtual bool selectVSplatMaskL(SDValue N, SDValue &Imm) const;
+  /// \brief Select constant vector splats whose value is a run of set bits
+  /// starting at bit zero.
+  virtual bool selectVSplatMaskR(SDValue N, SDValue &Imm) const;
 
   virtual SDNode *Select(SDNode *N);
 
diff --git a/lib/Target/Mips/MipsMSAInstrInfo.td b/lib/Target/Mips/MipsMSAInstrInfo.td
index 65d85e34ee..a393fdc45e 100644
--- a/lib/Target/Mips/MipsMSAInstrInfo.td
+++ b/lib/Target/Mips/MipsMSAInstrInfo.td
@@ -315,6 +315,18 @@ def vsplati64_simm5 : SplatComplexPattern<vsplat_simm5, v2i64, 1,
 def vsplat_uimm_pow2 : ComplexPattern<vAny, 1, "selectVSplatUimmPow2",
                                       [build_vector, bitconvert]>;
 
+// Any build_vector that is a constant splat with only a consecutive sequence
+// of left-most bits set.
+def vsplat_maskl_bits : SplatComplexPattern<vsplat_uimm8, vAny, 1,
+                                            "selectVSplatMaskL",
+                                            [build_vector, bitconvert]>;
+
+// Any build_vector that is a constant splat with only a consecutive sequence
+// of right-most bits set.
+def vsplat_maskr_bits : SplatComplexPattern<vsplat_uimm8, vAny, 1,
+                                            "selectVSplatMaskR",
+                                            [build_vector, bitconvert]>;
+
 def fms : PatFrag<(ops node:$wd, node:$ws, node:$wt),
                   (fsub node:$wd, (fmul node:$ws, node:$wt))>;
 
@@ -1079,6 +1091,31 @@ class MSA_BIT_D_DESC_BASE<string instr_asm, SDPatternOperator OpNode,
   InstrItinClass Itinerary = itin;
 }
 
+class MSA_BIT_BINSXI_DESC_BASE<string instr_asm, ValueType Ty,
+                               ComplexPattern Mask, RegisterOperand ROWD,
+                               RegisterOperand ROWS = ROWD,
+                               InstrItinClass itin = NoItinerary> {
+  dag OutOperandList = (outs ROWD:$wd);
+  dag InOperandList = (ins ROWD:$wd_in, ROWS:$ws, vsplat_uimm8:$m);
+  string AsmString = !strconcat(instr_asm, "\t$wd, $ws, $m");
+  list<dag> Pattern = [(set ROWD:$wd, (vselect (Ty Mask:$m), (Ty ROWD:$wd_in),
+                                               ROWS:$ws))];
+  InstrItinClass Itinerary = itin;
+  string Constraints = "$wd = $wd_in";
+}
+
+class MSA_BIT_BINSLI_DESC_BASE<string instr_asm, ValueType Ty,
+                               RegisterOperand ROWD,
+                               RegisterOperand ROWS = ROWD,
+                               InstrItinClass itin = NoItinerary> :
+  MSA_BIT_BINSXI_DESC_BASE<instr_asm, Ty, vsplat_maskl_bits, ROWD, ROWS, itin>;
+
+class MSA_BIT_BINSRI_DESC_BASE<string instr_asm, ValueType Ty,
+                               RegisterOperand ROWD,
+                               RegisterOperand ROWS = ROWD,
+                               InstrItinClass itin = NoItinerary> :
+  MSA_BIT_BINSXI_DESC_BASE<instr_asm, Ty, vsplat_maskr_bits, ROWD, ROWS, itin>;
+
 class MSA_BIT_SPLAT_DESC_BASE<string instr_asm, SDPatternOperator OpNode,
                               SplatComplexPattern SplatImm,
                               RegisterOperand ROWD, RegisterOperand ROWS = ROWD,
@@ -1486,28 +1523,20 @@ class BINSL_H_DESC : MSA_3R_DESC_BASE<"binsl.h", int_mips_binsl_h, MSA128HOpnd>;
 class BINSL_W_DESC : MSA_3R_DESC_BASE<"binsl.w", int_mips_binsl_w, MSA128WOpnd>;
 class BINSL_D_DESC : MSA_3R_DESC_BASE<"binsl.d", int_mips_binsl_d, MSA128DOpnd>;
 
-class BINSLI_B_DESC : MSA_BIT_B_DESC_BASE<"binsli.b", int_mips_binsli_b,
-                                          MSA128BOpnd>;
-class BINSLI_H_DESC : MSA_BIT_H_DESC_BASE<"binsli.h", int_mips_binsli_h,
-                                          MSA128HOpnd>;
-class BINSLI_W_DESC : MSA_BIT_W_DESC_BASE<"binsli.w", int_mips_binsli_w,
-                                          MSA128WOpnd>;
-class BINSLI_D_DESC : MSA_BIT_D_DESC_BASE<"binsli.d", int_mips_binsli_d,
-                                          MSA128DOpnd>;
+class BINSLI_B_DESC : MSA_BIT_BINSLI_DESC_BASE<"binsli.b", v16i8, MSA128BOpnd>;
+class BINSLI_H_DESC : MSA_BIT_BINSLI_DESC_BASE<"binsli.h", v8i16, MSA128HOpnd>;
+class BINSLI_W_DESC : MSA_BIT_BINSLI_DESC_BASE<"binsli.w", v4i32, MSA128WOpnd>;
+class BINSLI_D_DESC : MSA_BIT_BINSLI_DESC_BASE<"binsli.d", v2i64, MSA128DOpnd>;
 
 class BINSR_B_DESC : MSA_3R_DESC_BASE<"binsr.b", int_mips_binsr_b, MSA128BOpnd>;
 class BINSR_H_DESC : MSA_3R_DESC_BASE<"binsr.h", int_mips_binsr_h, MSA128HOpnd>;
 class BINSR_W_DESC : MSA_3R_DESC_BASE<"binsr.w", int_mips_binsr_w, MSA128WOpnd>;
 class BINSR_D_DESC : MSA_3R_DESC_BASE<"binsr.d", int_mips_binsr_d, MSA128DOpnd>;
 
-class BINSRI_B_DESC : MSA_BIT_B_DESC_BASE<"binsri.b", int_mips_binsri_b,
-                                          MSA128BOpnd>;
-class BINSRI_H_DESC : MSA_BIT_H_DESC_BASE<"binsri.h", int_mips_binsri_h,
-                                          MSA128HOpnd>;
-class BINSRI_W_DESC : MSA_BIT_W_DESC_BASE<"binsri.w", int_mips_binsri_w,
-                                          MSA128WOpnd>;
-class BINSRI_D_DESC : MSA_BIT_D_DESC_BASE<"binsri.d", int_mips_binsri_d,
-                                          MSA128DOpnd>;
+class BINSRI_B_DESC : MSA_BIT_BINSRI_DESC_BASE<"binsri.b", v16i8, MSA128BOpnd>;
+class BINSRI_H_DESC : MSA_BIT_BINSRI_DESC_BASE<"binsri.h", v8i16, MSA128HOpnd>;
+class BINSRI_W_DESC : MSA_BIT_BINSRI_DESC_BASE<"binsri.w", v4i32, MSA128WOpnd>;
+class BINSRI_D_DESC : MSA_BIT_BINSRI_DESC_BASE<"binsri.d", v2i64, MSA128DOpnd>;
 
 class BMNZ_V_DESC : MSA_VEC_DESC_BASE<"bmnz.v", int_mips_bmnz_v, MSA128BOpnd>;
 
diff --git a/lib/Target/Mips/MipsSEISelDAGToDAG.cpp b/lib/Target/Mips/MipsSEISelDAGToDAG.cpp
index 0a82a3a83f..93e7bdf38a 100644
--- a/lib/Target/Mips/MipsSEISelDAGToDAG.cpp
+++ b/lib/Target/Mips/MipsSEISelDAGToDAG.cpp
@@ -379,12 +379,6 @@ bool MipsSEDAGToDAGISel::selectIntAddrMM(SDValue Addr, SDValue &Base,
 // Returns true and sets Imm if:
 // * MSA is enabled
 // * N is a ISD::BUILD_VECTOR representing a constant splat
-// * The splat value fits in a signed 32-bit value.
-//
-// That last requirement isn't strictly a requirement of the instruction set
-// but it simplifies the callers by allowing them to assume they don't have to
-// handle 64-bit values. The callers will also be placing stricter requirements
-// on the immediates so this doesn't prohibit selection of legal immediates.
 bool MipsSEDAGToDAGISel::selectVSplat(SDNode *N, APInt &Imm) const {
   if (!Subtarget.hasMSA())
     return false;
@@ -403,10 +397,6 @@ bool MipsSEDAGToDAGISel::selectVSplat(SDNode *N, APInt &Imm) const {
                              !Subtarget.isLittle()))
     return false;
 
-  // None of the immediate forms can handle more than 32 bits
-  if (!SplatValue.isIntN(32))
-    return false;
-
   Imm = SplatValue;
 
   return true;
@@ -525,6 +515,70 @@ bool MipsSEDAGToDAGISel::selectVSplatUimmPow2(SDValue N, SDValue &Imm) const {
   return false;
 }
 
+// Select constant vector splats whose value only has a consecutive sequence
+// of left-most bits set (e.g. 0b11...1100...00).
+//
+// In addition to the requirements of selectVSplat(), this function returns
+// true and sets Imm if:
+// * The splat value is the same width as the elements of the vector
+// * The splat value is a consecutive sequence of left-most bits.
+//
+// This function looks through ISD::BITCAST nodes.
+// TODO: This might not be appropriate for big-endian MSA since BITCAST is
+//       sometimes a shuffle in big-endian mode.
+bool MipsSEDAGToDAGISel::selectVSplatMaskL(SDValue N, SDValue &Imm) const {
+  APInt ImmValue;
+  EVT EltTy = N->getValueType(0).getVectorElementType();
+
+  if (N->getOpcode() == ISD::BITCAST)
+    N = N->getOperand(0);
+
+  if (selectVSplat(N.getNode(), ImmValue) &&
+      ImmValue.getBitWidth() == EltTy.getSizeInBits()) {
+    // Extract the run of set bits starting with bit zero from the bitwise
+    // inverse of ImmValue, and test that the inverse of this is the same
+    // as the original value.
+    if (ImmValue == ~(~ImmValue & ~(~ImmValue + 1))) {
+
+      Imm = CurDAG->getTargetConstant(ImmValue.countPopulation(), EltTy);
+      return true;
+    }
+  }
+
+  return false;
+}
+
+// Select constant vector splats whose value only has a consecutive sequence
+// of right-most bits set (e.g. 0b00...0011...11).
+//
+// In addition to the requirements of selectVSplat(), this function returns
+// true and sets Imm if:
+// * The splat value is the same width as the elements of the vector
+// * The splat value is a consecutive sequence of right-most bits.
+//
+// This function looks through ISD::BITCAST nodes.
+// TODO: This might not be appropriate for big-endian MSA since BITCAST is
+//       sometimes a shuffle in big-endian mode.
+bool MipsSEDAGToDAGISel::selectVSplatMaskR(SDValue N, SDValue &Imm) const {
+  APInt ImmValue;
+  EVT EltTy = N->getValueType(0).getVectorElementType();
+
+  if (N->getOpcode() == ISD::BITCAST)
+    N = N->getOperand(0);
+
+  if (selectVSplat(N.getNode(), ImmValue) &&
+      ImmValue.getBitWidth() == EltTy.getSizeInBits()) {
+    // Extract the run of set bits starting with bit zero, and test that the
+    // result is the same as the original value
+    if (ImmValue == (ImmValue & ~(ImmValue + 1))) {
+      Imm = CurDAG->getTargetConstant(ImmValue.countPopulation(), EltTy);
+      return true;
+    }
+  }
+
+  return false;
+}
+
 std::pair<bool, SDNode*> MipsSEDAGToDAGISel::selectNode(SDNode *Node) {
   unsigned Opcode = Node->getOpcode();
   SDLoc DL(Node);
diff --git a/lib/Target/Mips/MipsSEISelDAGToDAG.h b/lib/Target/Mips/MipsSEISelDAGToDAG.h
index 759d3afc37..851fba0e19 100644
--- a/lib/Target/Mips/MipsSEISelDAGToDAG.h
+++ b/lib/Target/Mips/MipsSEISelDAGToDAG.h
@@ -81,6 +81,12 @@ private:
   virtual bool selectVSplatSimm5(SDValue N, SDValue &Imm) const;
   /// \brief Select constant vector splats whose value is a power of 2.
   virtual bool selectVSplatUimmPow2(SDValue N, SDValue &Imm) const;
+  /// \brief Select constant vector splats whose value is a run of set bits
+  /// ending at the most significant bit
+  virtual bool selectVSplatMaskL(SDValue N, SDValue &Imm) const;
+  /// \brief Select constant vector splats whose value is a run of set bits
+  /// starting at bit zero.
+  virtual bool selectVSplatMaskR(SDValue N, SDValue &Imm) const;
 
   virtual std::pair<bool, SDNode*> selectNode(SDNode *Node);
 
diff --git a/lib/Target/Mips/MipsSEISelLowering.cpp b/lib/Target/Mips/MipsSEISelLowering.cpp
index d858e2aac5..9c543330cc 100644
--- a/lib/Target/Mips/MipsSEISelLowering.cpp
+++ b/lib/Target/Mips/MipsSEISelLowering.cpp
@@ -1260,6 +1260,30 @@ SDValue MipsSETargetLowering::lowerINTRINSIC_WO_CHAIN(SDValue Op,
   case Intrinsic::mips_andi_b:
     return DAG.getNode(ISD::AND, DL, Op->getValueType(0), Op->getOperand(1),
                        lowerMSASplatImm(Op, 2, DAG));
+  case Intrinsic::mips_binsli_b:
+  case Intrinsic::mips_binsli_h:
+  case Intrinsic::mips_binsli_w:
+  case Intrinsic::mips_binsli_d: {
+    EVT VecTy = Op->getValueType(0);
+    EVT EltTy = VecTy.getVectorElementType();
+    APInt Mask = APInt::getHighBitsSet(EltTy.getSizeInBits(),
+                                       Op->getConstantOperandVal(3));
+    return DAG.getNode(ISD::VSELECT, DL, VecTy,
+                       DAG.getConstant(Mask, VecTy, true), Op->getOperand(1),
+                       Op->getOperand(2));
+  }
+  case Intrinsic::mips_binsri_b:
+  case Intrinsic::mips_binsri_h:
+  case Intrinsic::mips_binsri_w:
+  case Intrinsic::mips_binsri_d: {
+    EVT VecTy = Op->getValueType(0);
+    EVT EltTy = VecTy.getVectorElementType();
+    APInt Mask = APInt::getLowBitsSet(EltTy.getSizeInBits(),
+                                      Op->getConstantOperandVal(3));
+    return DAG.getNode(ISD::VSELECT, DL, VecTy,
+                       DAG.getConstant(Mask, VecTy, true), Op->getOperand(1),
+                       Op->getOperand(2));
+  }
   case Intrinsic::mips_bnz_b:
   case Intrinsic::mips_bnz_h:
   case Intrinsic::mips_bnz_w: