[RISCV] Add matching of codegen patterns to RISCV Bit Manipulation Zbb asm instructions

This patch provides optimization of bit manipulation operations by
enabling the +experimental-b target feature.
It adds matching of single block patterns of instructions to specific
bit-manip instructions from the base subset (zbb subextension) of the
experimental B extension of RISC-V.
It adds also the correspondent codegen tests.

This patch is based on Claire Wolf's proposal for the bit manipulation
extension of RISCV:
https://github.com/riscv/riscv-bitmanip/blob/master/bitmanip-0.92.pdf

Differential Revision: https://reviews.llvm.org/D79870

lib/Target/RISCV/RISCVISelDAGToDAG.cpp

@@ -184,6 +184,196 @@ bool RISCVDAGToDAGISel::SelectAddrFI(SDValue Addr, SDValue &Base) {
   return false;
 }
 
+// Check that it is a SLOI (Shift Left Ones Immediate). We first check that
+// it is the right node tree:
+//
+//  (OR (SHL RS1, VC2), VC1)
+//
+// and then we check that VC1, the mask used to fill with ones, is compatible
+// with VC2, the shamt:
+//
+//  VC1 == maskTrailingOnes<uint64_t>(VC2)
+
+bool RISCVDAGToDAGISel::SelectSLOI(SDValue N, SDValue &RS1, SDValue &Shamt) {
+  MVT XLenVT = Subtarget->getXLenVT();
+  if (N.getOpcode() == ISD::OR) {
+    SDValue Or = N;
+    if (Or.getOperand(0).getOpcode() == ISD::SHL) {
+      SDValue Shl = Or.getOperand(0);
+      if (isa<ConstantSDNode>(Shl.getOperand(1)) &&
+          isa<ConstantSDNode>(Or.getOperand(1))) {
+        if (XLenVT == MVT::i64) {
+          uint64_t VC1 = Or.getConstantOperandVal(1);
+          uint64_t VC2 = Shl.getConstantOperandVal(1);
+          if (VC1 == maskTrailingOnes<uint64_t>(VC2)) {
+            RS1 = Shl.getOperand(0);
+            Shamt = CurDAG->getTargetConstant(VC2, SDLoc(N),
+                           Shl.getOperand(1).getValueType());
+            return true;
+          }
+        }
+        if (XLenVT == MVT::i32) {
+          uint32_t VC1 = Or.getConstantOperandVal(1);
+          uint32_t VC2 = Shl.getConstantOperandVal(1);
+          if (VC1 == maskTrailingOnes<uint32_t>(VC2)) {
+            RS1 = Shl.getOperand(0);
+            Shamt = CurDAG->getTargetConstant(VC2, SDLoc(N),
+                           Shl.getOperand(1).getValueType());
+            return true;
+          }
+        }
+      }
+    }
+  }
+  return false;
+}
+
+// Check that it is a SROI (Shift Right Ones Immediate). We first check that
+// it is the right node tree:
+//
+//  (OR (SRL RS1, VC2), VC1)
+//
+// and then we check that VC1, the mask used to fill with ones, is compatible
+// with VC2, the shamt:
+//
+//  VC1 == maskLeadingOnes<uint64_t>(VC2)
+
+bool RISCVDAGToDAGISel::SelectSROI(SDValue N, SDValue &RS1, SDValue &Shamt) {
+  MVT XLenVT = Subtarget->getXLenVT();
+  if (N.getOpcode() == ISD::OR) {
+    SDValue Or = N;
+    if (Or.getOperand(0).getOpcode() == ISD::SRL) {
+      SDValue Srl = Or.getOperand(0);
+      if (isa<ConstantSDNode>(Srl.getOperand(1)) &&
+          isa<ConstantSDNode>(Or.getOperand(1))) {
+        if (XLenVT == MVT::i64) {
+          uint64_t VC1 = Or.getConstantOperandVal(1);
+          uint64_t VC2 = Srl.getConstantOperandVal(1);
+          if (VC1 == maskLeadingOnes<uint64_t>(VC2)) {
+            RS1 = Srl.getOperand(0);
+            Shamt = CurDAG->getTargetConstant(VC2, SDLoc(N),
+                           Srl.getOperand(1).getValueType());
+            return true;
+          }
+        }
+        if (XLenVT == MVT::i32) {
+          uint32_t VC1 = Or.getConstantOperandVal(1);
+          uint32_t VC2 = Srl.getConstantOperandVal(1);
+          if (VC1 == maskLeadingOnes<uint32_t>(VC2)) {
+            RS1 = Srl.getOperand(0);
+            Shamt = CurDAG->getTargetConstant(VC2, SDLoc(N),
+                           Srl.getOperand(1).getValueType());
+            return true;
+          }
+        }
+      }
+    }
+  }
+  return false;
+}
+
+// Check that it is a SLLIUW (Shift Logical Left Immediate Unsigned i32
+// on RV64).
+// SLLIUW is the same as SLLI except for the fact that it clears the bits
+// XLEN-1:32 of the input RS1 before shifting.
+// We first check that it is the right node tree:
+//
+//  (AND (SHL RS1, VC2), VC1)
+//
+// We check that VC2, the shamt is less than 32, otherwise the pattern is
+// exactly the same as SLLI and we give priority to that.
+// Eventually we check that that VC1, the mask used to clear the upper 32 bits
+// of RS1, is correct:
+//
+//  VC1 == (0xFFFFFFFF << VC2)
+
+bool RISCVDAGToDAGISel::SelectSLLIUW(SDValue N, SDValue &RS1, SDValue &Shamt) {
+  if (N.getOpcode() == ISD::AND && Subtarget->getXLenVT() == MVT::i64) {
+    SDValue And = N;
+    if (And.getOperand(0).getOpcode() == ISD::SHL) {
+      SDValue Shl = And.getOperand(0);
+      if (isa<ConstantSDNode>(Shl.getOperand(1)) &&
+          isa<ConstantSDNode>(And.getOperand(1))) {
+        uint64_t VC1 = And.getConstantOperandVal(1);
+        uint64_t VC2 = Shl.getConstantOperandVal(1);
+        if (VC2 < 32 && VC1 == ((uint64_t)0xFFFFFFFF << VC2)) {
+          RS1 = Shl.getOperand(0);
+          Shamt = CurDAG->getTargetConstant(VC2, SDLoc(N),
+                                            Shl.getOperand(1).getValueType());
+          return true;
+        }
+      }
+    }
+  }
+  return false;
+}
+
+// Check that it is a SLOIW (Shift Left Ones Immediate i32 on RV64).
+// We first check that it is the right node tree:
+//
+//  (SIGN_EXTEND_INREG (OR (SHL RS1, VC2), VC1))
+//
+// and then we check that VC1, the mask used to fill with ones, is compatible
+// with VC2, the shamt:
+//
+//  VC1 == maskTrailingOnes<uint32_t>(VC2)
+
+bool RISCVDAGToDAGISel::SelectSLOIW(SDValue N, SDValue &RS1, SDValue &Shamt) {
+  if (Subtarget->getXLenVT() == MVT::i64 &&
+      N.getOpcode() == ISD::SIGN_EXTEND_INREG &&
+      cast<VTSDNode>(N.getOperand(1))->getVT() == MVT::i32) {
+    if (N.getOperand(0).getOpcode() == ISD::OR) {
+      SDValue Or = N.getOperand(0);
+      if (Or.getOperand(0).getOpcode() == ISD::SHL) {
+        SDValue Shl = Or.getOperand(0);
+        if (isa<ConstantSDNode>(Shl.getOperand(1)) &&
+            isa<ConstantSDNode>(Or.getOperand(1))) {
+          uint32_t VC1 = Or.getConstantOperandVal(1);
+          uint32_t VC2 = Shl.getConstantOperandVal(1);
+          if (VC1 == maskTrailingOnes<uint32_t>(VC2)) {
+            RS1 = Shl.getOperand(0);
+            Shamt = CurDAG->getTargetConstant(VC2, SDLoc(N),
+                                              Shl.getOperand(1).getValueType());
+            return true;
+          }
+        }
+      }
+    }
+  }
+  return false;
+}
+
+// Check that it is a SROIW (Shift Right Ones Immediate i32 on RV64).
+// We first check that it is the right node tree:
+//
+//  (OR (SHL RS1, VC2), VC1)
+//
+// and then we check that VC1, the mask used to fill with ones, is compatible
+// with VC2, the shamt:
+//
+//  VC1 == maskLeadingOnes<uint32_t>(VC2)
+
+bool RISCVDAGToDAGISel::SelectSROIW(SDValue N, SDValue &RS1, SDValue &Shamt) {
+  if (N.getOpcode() == ISD::OR && Subtarget->getXLenVT() == MVT::i64) {
+    SDValue Or = N;
+    if (Or.getOperand(0).getOpcode() == ISD::SRL) {
+      SDValue Srl = Or.getOperand(0);
+      if (isa<ConstantSDNode>(Srl.getOperand(1)) &&
+          isa<ConstantSDNode>(Or.getOperand(1))) {
+        uint32_t VC1 = Or.getConstantOperandVal(1);
+        uint32_t VC2 = Srl.getConstantOperandVal(1);
+        if (VC1 == maskLeadingOnes<uint32_t>(VC2)) {
+          RS1 = Srl.getOperand(0);
+          Shamt = CurDAG->getTargetConstant(VC2, SDLoc(N),
+                                            Srl.getOperand(1).getValueType());
+          return true;
+        }
+      }
+    }
+  }
+  return false;
+}
+
 // Merge an ADDI into the offset of a load/store instruction where possible.
 // (load (addi base, off1), off2) -> (load base, off1+off2)
 // (store val, (addi base, off1), off2) -> (store val, base, off1+off2)

lib/Target/RISCV/RISCVISelDAGToDAG.h

@@ -45,6 +45,12 @@ public:
 
   bool SelectAddrFI(SDValue Addr, SDValue &Base);
 
+  bool SelectSLOI(SDValue N, SDValue &RS1, SDValue &Shamt);
+  bool SelectSROI(SDValue N, SDValue &RS1, SDValue &Shamt);
+  bool SelectSLLIUW(SDValue N, SDValue &RS1, SDValue &Shamt);
+  bool SelectSLOIW(SDValue N, SDValue &RS1, SDValue &Shamt);
+  bool SelectSROIW(SDValue N, SDValue &RS1, SDValue &Shamt);
+
 // Include the pieces autogenerated from the target description.
 #include "RISCVGenDAGISel.inc"
 

lib/Target/RISCV/RISCVISelLowering.cpp

@@ -152,9 +152,12 @@ RISCVTargetLowering::RISCVTargetLowering(const TargetMachine &TM,
   setOperationAction(ISD::ROTL, XLenVT, Expand);
   setOperationAction(ISD::ROTR, XLenVT, Expand);
   setOperationAction(ISD::BSWAP, XLenVT, Expand);
+
+  if (!Subtarget.hasStdExtZbb()) {
     setOperationAction(ISD::CTTZ, XLenVT, Expand);
     setOperationAction(ISD::CTLZ, XLenVT, Expand);
     setOperationAction(ISD::CTPOP, XLenVT, Expand);
+  }
 
   ISD::CondCode FPCCToExtend[] = {
       ISD::SETOGT, ISD::SETOGE, ISD::SETONE, ISD::SETUEQ, ISD::SETUGT,

lib/Target/RISCV/RISCVInstrInfoB.td

@@ -632,3 +632,79 @@ let Predicates = [HasStdExtZbproposedc, HasStdExtZbbOrZbp, HasStdExtC, IsRV64] i
 def : CompressPat<(PACK GPRC:$rs1, GPRC:$rs1, X0),
                   (C_ZEXTW GPRC:$rs1)>;
 } // Predicates = [HasStdExtZbproposedc, HasStdExtC, IsRV64]
+
+//===----------------------------------------------------------------------===//
+// Codegen patterns
+//===----------------------------------------------------------------------===//
+def SLOIPat   : ComplexPattern<XLenVT, 2, "SelectSLOI", [or]>;
+def SROIPat   : ComplexPattern<XLenVT, 2, "SelectSROI", [or]>;
+def SLLIUWPat : ComplexPattern<i64, 2, "SelectSLLIUW", [and]>;
+def SLOIWPat  : ComplexPattern<i64, 2, "SelectSLOIW", [sext_inreg]>;
+def SROIWPat  : ComplexPattern<i64, 2, "SelectSROIW", [or]>;
+
+let Predicates = [HasStdExtZbb] in {
+def : Pat<(xor (shl (xor GPR:$rs1, -1), GPR:$rs2), -1),
+          (SLO GPR:$rs1, GPR:$rs2)>;
+def : Pat<(xor (srl (xor GPR:$rs1, -1), GPR:$rs2), -1),
+          (SRO GPR:$rs1, GPR:$rs2)>;
+def : Pat<(SLOIPat GPR:$rs1, uimmlog2xlen:$shamt),
+          (SLOI GPR:$rs1, uimmlog2xlen:$shamt)>;
+def : Pat<(SROIPat GPR:$rs1, uimmlog2xlen:$shamt),
+          (SROI GPR:$rs1, uimmlog2xlen:$shamt)>;
+def : Pat<(ctlz GPR:$rs1), (CLZ GPR:$rs1)>;
+def : Pat<(cttz GPR:$rs1), (CTZ GPR:$rs1)>;
+def : Pat<(ctpop GPR:$rs1), (PCNT GPR:$rs1)>;
+} // Predicates = [HasStdExtZbb]
+
+let Predicates = [HasStdExtZbb, IsRV32] in
+def : Pat<(sra (shl GPR:$rs1, (i32 24)), (i32 24)), (SEXTB GPR:$rs1)>;
+let Predicates = [HasStdExtZbb, IsRV64] in
+def : Pat<(sra (shl GPR:$rs1, (i64 56)), (i64 56)), (SEXTB GPR:$rs1)>;
+
+let Predicates = [HasStdExtZbb, IsRV32] in
+def : Pat<(sra (shl GPR:$rs1, (i32 16)), (i32 16)), (SEXTH GPR:$rs1)>;
+let Predicates = [HasStdExtZbb, IsRV64] in
+def : Pat<(sra (shl GPR:$rs1, (i64 48)), (i64 48)), (SEXTH GPR:$rs1)>;
+
+let Predicates = [HasStdExtZbb] in {
+def : Pat<(smin GPR:$rs1, GPR:$rs2), (MIN  GPR:$rs1, GPR:$rs2)>;
+def : Pat<(riscv_selectcc GPR:$rs1, GPR:$rs2, (XLenVT 20), GPR:$rs1, GPR:$rs2),
+          (MIN  GPR:$rs1, GPR:$rs2)>;
+def : Pat<(smax GPR:$rs1, GPR:$rs2), (MAX  GPR:$rs1, GPR:$rs2)>;
+def : Pat<(riscv_selectcc GPR:$rs2, GPR:$rs1, (XLenVT 20), GPR:$rs1, GPR:$rs2),
+          (MAX  GPR:$rs1, GPR:$rs2)>;
+def : Pat<(umin GPR:$rs1, GPR:$rs2), (MINU GPR:$rs1, GPR:$rs2)>;
+def : Pat<(riscv_selectcc GPR:$rs1, GPR:$rs2, (XLenVT 12), GPR:$rs1, GPR:$rs2),
+          (MINU  GPR:$rs1, GPR:$rs2)>;
+def : Pat<(umax GPR:$rs1, GPR:$rs2), (MAXU GPR:$rs1, GPR:$rs2)>;
+def : Pat<(riscv_selectcc GPR:$rs2, GPR:$rs1, (XLenVT 12), GPR:$rs1, GPR:$rs2),
+          (MAXU  GPR:$rs1, GPR:$rs2)>;
+} // Predicates = [HasStdExtZbb]
+
+let Predicates = [HasStdExtZbb, IsRV64] in {
+def : Pat<(and (add GPR:$rs, simm12:$simm12), (i64 0xFFFFFFFF)),
+          (ADDIWU GPR:$rs, simm12:$simm12)>;
+def : Pat<(SLLIUWPat GPR:$rs1, uimmlog2xlen:$shamt),
+          (SLLIUW GPR:$rs1, uimmlog2xlen:$shamt)>;
+def : Pat<(and (add GPR:$rs1, GPR:$rs2), (i64 0xFFFFFFFF)),
+          (ADDWU GPR:$rs1, GPR:$rs2)>;
+def : Pat<(and (sub GPR:$rs1, GPR:$rs2), (i64 0xFFFFFFFF)),
+          (SUBWU GPR:$rs1, GPR:$rs2)>;
+def : Pat<(add GPR:$rs1, (and GPR:$rs2, (i64 0xFFFFFFFF))),
+          (ADDUW GPR:$rs1, GPR:$rs2)>;
+def : Pat<(sub GPR:$rs1, (and GPR:$rs2, (i64 0xFFFFFFFF))),
+          (SUBUW GPR:$rs1, GPR:$rs2)>;
+def : Pat<(xor (riscv_sllw (xor GPR:$rs1, -1), GPR:$rs2), -1),
+          (SLOW GPR:$rs1, GPR:$rs2)>;
+def : Pat<(xor (riscv_srlw (xor GPR:$rs1, -1), GPR:$rs2), -1),
+          (SROW GPR:$rs1, GPR:$rs2)>;
+def : Pat<(SLOIWPat GPR:$rs1, uimmlog2xlen:$shamt),
+          (SLOIW GPR:$rs1, uimmlog2xlen:$shamt)>;
+def : Pat<(SROIWPat GPR:$rs1, uimmlog2xlen:$shamt),
+          (SROIW GPR:$rs1, uimmlog2xlen:$shamt)>;
+def : Pat<(add (ctlz (and GPR:$rs1, (i64 0xFFFFFFFF))), (i64 -32)),
+          (CLZW GPR:$rs1)>;
+// We don't pattern-match CTZW here as it has the same pattern and result as
+// RV64 CTZ
+def : Pat<(ctpop (and GPR:$rs1, (i64 0xFFFFFFFF))), (PCNTW GPR:$rs1)>;
+} // Predicates = [HasStdExtZbb, IsRV64]