[llvm-commits] CVS: llvm/lib/Target/PowerPC/PPC32ISelSimple.cpp PowerPCInstrInfo.td

Nate Begeman natebegeman at mac.com
Fri Aug 20 02:56:33 PDT 2004



Changes in directory llvm/lib/Target/PowerPC:

PPC32ISelSimple.cpp updated: 1.71 -> 1.72
PowerPCInstrInfo.td updated: 1.25 -> 1.26
---
Log message:

Implement code to convert SetCC into straight line code where appropriate.  Add necessary instructions for this transformation to the .td file.

---
Diffs of the changes:  (+173 -22)

Index: llvm/lib/Target/PowerPC/PPC32ISelSimple.cpp
diff -u llvm/lib/Target/PowerPC/PPC32ISelSimple.cpp:1.71 llvm/lib/Target/PowerPC/PPC32ISelSimple.cpp:1.72
--- llvm/lib/Target/PowerPC/PPC32ISelSimple.cpp:1.71	Thu Aug 19 16:00:12 2004
+++ llvm/lib/Target/PowerPC/PPC32ISelSimple.cpp	Fri Aug 20 04:56:22 2004
@@ -33,6 +33,7 @@
 
 namespace {
   Statistic<> GEPFolds("ppc-codegen", "Number of GEPs folded");
+  Statistic<> NumSetCC("ppc-codegen", "Number of SetCC straight-lined");
 
   /// TypeClass - Used by the PowerPC backend to group LLVM types by their basic
   /// PPC Representation.
@@ -280,6 +281,10 @@
       abort();
     }
 
+    unsigned ExtendOrClear(MachineBasicBlock *MBB,
+                           MachineBasicBlock::iterator IP,
+                           unsigned Reg, const Type *CompTy);
+
     /// promote32 - Make a value 32-bits wide, and put it somewhere.
     ///
     void promote32(unsigned targetReg, const ValueRecord &VR);
@@ -959,36 +964,43 @@
     BuildMI(*MBB, IP, PPC::FCMPU, 2, PPC::CR0).addReg(LHS).addReg(RHS);
 }
 
-/// EmitComparison - emits a comparison of the two operands, returning the
-/// extended setcc code to use.  The result is in CR0.
-///
-unsigned ISel::EmitComparison(unsigned OpNum, Value *Op0, Value *Op1,
-                              MachineBasicBlock *MBB,
-                              MachineBasicBlock::iterator IP) {
-  // The arguments are already supposed to be of the same type.
-  const Type *CompTy = Op0->getType();
+unsigned ISel::ExtendOrClear(MachineBasicBlock *MBB,
+                             MachineBasicBlock::iterator IP,
+                             unsigned Reg, const Type *CompTy) {
   unsigned Class = getClassB(CompTy);
-  unsigned Op0r = getReg(Op0, MBB, IP);
 
-  // Before we do a comparison, we have to make sure that we're truncating our
-  // registers appropriately.
+  // Before we do a comparison or SetCC, we have to make sure that we truncate
+  // the source registers appropriately.
   if (Class == cByte) {
     unsigned TmpReg = makeAnotherReg(CompTy);
     if (CompTy->isSigned())
-      BuildMI(*MBB, IP, PPC::EXTSB, 1, TmpReg).addReg(Op0r);
+      BuildMI(*MBB, IP, PPC::EXTSB, 1, TmpReg).addReg(Reg);
     else
-      BuildMI(*MBB, IP, PPC::RLWINM, 4, TmpReg).addReg(Op0r).addImm(0)
+      BuildMI(*MBB, IP, PPC::RLWINM, 4, TmpReg).addReg(Reg).addImm(0)
         .addImm(24).addImm(31);
-    Op0r = TmpReg;
+    Reg = TmpReg;
   } else if (Class == cShort) {
     unsigned TmpReg = makeAnotherReg(CompTy);
     if (CompTy->isSigned())
-      BuildMI(*MBB, IP, PPC::EXTSH, 1, TmpReg).addReg(Op0r);
+      BuildMI(*MBB, IP, PPC::EXTSH, 1, TmpReg).addReg(Reg);
     else
-      BuildMI(*MBB, IP, PPC::RLWINM, 4, TmpReg).addReg(Op0r).addImm(0)
+      BuildMI(*MBB, IP, PPC::RLWINM, 4, TmpReg).addReg(Reg).addImm(0)
         .addImm(16).addImm(31);
-    Op0r = TmpReg;
+    Reg = TmpReg;
   }
+  return Reg;
+}
+
+/// EmitComparison - emits a comparison of the two operands, returning the
+/// extended setcc code to use.  The result is in CR0.
+///
+unsigned ISel::EmitComparison(unsigned OpNum, Value *Op0, Value *Op1,
+                              MachineBasicBlock *MBB,
+                              MachineBasicBlock::iterator IP) {
+  // The arguments are already supposed to be of the same type.
+  const Type *CompTy = Op0->getType();
+  unsigned Class = getClassB(CompTy);
+  unsigned Op0r = ExtendOrClear(MBB, IP, getReg(Op0, MBB, IP), CompTy);
   
   // Use crand for lt, gt and crandc for le, ge
   unsigned CROpcode = (OpNum == 2 || OpNum == 4) ? PPC::CRAND : PPC::CRANDC;
@@ -1098,13 +1110,148 @@
   if (canFoldSetCCIntoBranchOrSelect(&I))
     return;
 
+  Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
+  const Type *Ty = Op0->getType();
+  unsigned Class = getClassB(Ty);
+  unsigned Opcode = I.getOpcode();
   unsigned DestReg = getReg(I);
-  unsigned OpNum = I.getOpcode();
-  const Type *Ty = I.getOperand (0)->getType();
 
-  EmitComparison(OpNum, I.getOperand(0), I.getOperand(1), BB, BB->end());
+  // If the comparison type is byte, short, or int, then we can emit a
+  // branchless version of the SetCC that puts 0 (false) or 1 (true) in the
+  // destination register.
+  if (Class <= cInt) {
+    ++NumSetCC;
+    MachineBasicBlock::iterator MI = BB->end();
+    unsigned OpNum = getSetCCNumber(Opcode);
+    unsigned Op0Reg = getReg(Op0, BB, MI);
+    Op0Reg = ExtendOrClear(BB, MI, Op0Reg, Ty);
+    
+    // comparisons against constant zero often have shorter sequences than the
+    // general case, handled below.
+    ConstantInt *CI = dyn_cast<ConstantInt>(Op1);
+    if (CI && CI->getRawValue() == 0) {
+      switch(OpNum) {
+      case 0: { // eq0
+        unsigned TempReg = makeAnotherReg(Type::IntTy);
+        BuildMI(*BB, MI, PPC::CNTLZW, 1, TempReg).addReg(Op0Reg);
+        BuildMI(*BB, MI, PPC::SRWI, 2, DestReg).addReg(TempReg).addImm(5);
+        break;
+        } 
+      case 1: { // ne0
+        unsigned TempReg = makeAnotherReg(Type::IntTy);
+        BuildMI(*BB, MI, PPC::ADDIC, 2, TempReg).addReg(Op0Reg).addSImm(-1);
+        BuildMI(*BB, MI, PPC::SUBFE, 2, DestReg).addReg(TempReg).addReg(Op0Reg);
+        break;
+        } 
+      case 2: { // lt0, always false if unsigned
+        if (Ty->isSigned())
+          BuildMI(*BB, MI, PPC::SRWI, 2, DestReg).addReg(Op0Reg).addImm(31);
+        else
+          BuildMI(*BB, MI, PPC::LI, 1, DestReg).addSImm(0);
+        break;
+        }
+      case 3: { // ge0, always true if unsigned
+        if (Ty->isSigned()) { 
+          unsigned TempReg = makeAnotherReg(Type::IntTy);
+          BuildMI(*BB, MI, PPC::SRWI, 2, TempReg).addReg(Op0Reg).addImm(31);
+          BuildMI(*BB, MI, PPC::XORI, 2, DestReg).addReg(TempReg).addImm(1);
+        } else {
+          BuildMI(*BB, MI, PPC::LI, 1, DestReg).addSImm(1);
+        }
+        break;
+        }
+      case 4: { // gt0, equivalent to ne0 if unsigned
+        unsigned Temp1 = makeAnotherReg(Type::IntTy);
+        unsigned Temp2 = makeAnotherReg(Type::IntTy);
+        if (Ty->isSigned()) { 
+          BuildMI(*BB, MI, PPC::NEG, 2, Temp1).addReg(Op0Reg);
+          BuildMI(*BB, MI, PPC::ANDC, 2, Temp2).addReg(Temp1).addReg(Op0Reg);
+          BuildMI(*BB, MI, PPC::SRWI, 2, DestReg).addReg(Temp2).addImm(31);
+        } else {
+          BuildMI(*BB, MI, PPC::ADDIC, 2, Temp1).addReg(Op0Reg).addSImm(-1);
+          BuildMI(*BB, MI, PPC::SUBFE, 2, DestReg).addReg(Temp1).addReg(Op0Reg);
+        }
+        break;
+        }
+      case 5: { // le0, equivalent to eq0 if unsigned
+        unsigned Temp1 = makeAnotherReg(Type::IntTy);
+        unsigned Temp2 = makeAnotherReg(Type::IntTy);
+        if (Ty->isSigned()) { 
+          BuildMI(*BB, MI, PPC::NEG, 2, Temp1).addReg(Op0Reg);
+          BuildMI(*BB, MI, PPC::ORC, 2, Temp2).addReg(Op0Reg).addReg(Temp1);
+          BuildMI(*BB, MI, PPC::SRWI, 2, DestReg).addReg(Temp2).addImm(31);
+        } else {
+          BuildMI(*BB, MI, PPC::CNTLZW, 1, Temp1).addReg(Op0Reg);
+          BuildMI(*BB, MI, PPC::SRWI, 2, DestReg).addReg(Temp1).addImm(5);
+        }
+        break;
+        }
+      } // switch
+      return;
+    }
+    unsigned Op1Reg = getReg(Op1, BB, MI);
+    switch(OpNum) {
+    case 0: { // eq
+      unsigned Temp1 = makeAnotherReg(Type::IntTy);
+      unsigned Temp2 = makeAnotherReg(Type::IntTy);
+      BuildMI(*BB, MI, PPC::SUBF, 2, Temp1).addReg(Op0Reg).addReg(Op1Reg);
+      BuildMI(*BB, MI, PPC::CNTLZW, 1, Temp2).addReg(Temp1);
+      BuildMI(*BB, MI, PPC::SRWI, 2, DestReg).addReg(Temp2).addImm(5);
+      break;
+      } 
+    case 1: { // ne
+      unsigned Temp1 = makeAnotherReg(Type::IntTy);
+      unsigned Temp2 = makeAnotherReg(Type::IntTy);
+      BuildMI(*BB, MI, PPC::SUBF, 2, Temp1).addReg(Op0Reg).addReg(Op1Reg);
+      BuildMI(*BB, MI, PPC::ADDIC, 2, Temp2).addReg(Temp1).addSImm(-1);
+      BuildMI(*BB, MI, PPC::SUBFE, 2, DestReg).addReg(Temp2).addReg(Temp1);
+      break;
+      } 
+    case 2: 
+    case 4: { // lt, gt
+      unsigned Temp1 = makeAnotherReg(Type::IntTy);
+      unsigned Temp2 = makeAnotherReg(Type::IntTy);
+      unsigned Temp3 = makeAnotherReg(Type::IntTy);
+      unsigned Temp4 = makeAnotherReg(Type::IntTy);
+      if (OpNum == 4) std::swap(Op0Reg, Op1Reg);
+      if (Ty->isSigned()) { 
+        BuildMI(*BB, MI, PPC::SUBFC, 2, Temp1).addReg(Op1Reg).addReg(Op0Reg);
+        BuildMI(*BB, MI, PPC::EQV, 2, Temp2).addReg(Op1Reg).addReg(Op0Reg);
+        BuildMI(*BB, MI, PPC::SRWI, 2, Temp3).addReg(Temp2).addImm(31);
+        BuildMI(*BB, MI, PPC::ADDZE, 1, Temp4).addReg(Temp3);
+        BuildMI(*BB, MI, PPC::RLWINM, 4, DestReg).addReg(Temp4).addImm(0)
+          .addImm(31).addImm(31);
+      } else {
+        BuildMI(*BB, MI, PPC::SUBFC, 2, Temp1).addReg(Op1Reg).addReg(Op0Reg);
+        BuildMI(*BB, MI, PPC::SUBFE, 2, Temp2).addReg(Temp1).addReg(Temp1);
+        BuildMI(*BB, MI, PPC::NEG, 2, DestReg).addReg(Temp2);
+      }
+      break;
+      }
+    case 3: 
+    case 5: { // le, ge
+      unsigned Temp1 = makeAnotherReg(Type::IntTy);
+      unsigned Temp2 = makeAnotherReg(Type::IntTy);
+      unsigned Temp3 = makeAnotherReg(Type::IntTy);
+      if (OpNum == 3) std::swap(Op0Reg, Op1Reg);
+      if (Ty->isSigned()) {
+        BuildMI(*BB, MI, PPC::SRWI, 2, Temp1).addReg(Op0Reg).addImm(31);
+        BuildMI(*BB, MI, PPC::SRAWI, 2, Temp2).addReg(Op1Reg).addImm(31);
+        BuildMI(*BB, MI, PPC::SUBFC, 2, Temp3).addReg(Op0Reg).addReg(Op1Reg);
+        BuildMI(*BB, MI, PPC::ADDE, 2, DestReg).addReg(Temp2).addReg(Temp1);
+      } else {
+        BuildMI(*BB, MI, PPC::LI, 1, Temp2).addSImm(-1);
+        BuildMI(*BB, MI, PPC::SUBFC, 2, Temp1).addReg(Op0Reg).addReg(Op1Reg);
+        BuildMI(*BB, MI, PPC::SUBFZE, 1, DestReg).addReg(Temp2);
+      }
+      break;
+      }
+    } // switch
+    return;
+  }
+  EmitComparison(Opcode, Op0, Op1, BB, BB->end());
   
-  unsigned Opcode = getPPCOpcodeForSetCCNumber(OpNum);
+  unsigned PPCOpcode = getPPCOpcodeForSetCCNumber(Opcode);
   MachineBasicBlock *thisMBB = BB;
   const BasicBlock *LLVM_BB = BB->getBasicBlock();
   ilist<MachineBasicBlock>::iterator It = BB;
@@ -1121,7 +1268,7 @@
   // bCC. But MBB->getFirstTerminator() can't understand this.
   MachineBasicBlock *copy1MBB = new MachineBasicBlock(LLVM_BB);
   F->getBasicBlockList().insert(It, copy1MBB);
-  BuildMI(BB, Opcode, 2).addReg(PPC::CR0).addMBB(copy1MBB);
+  BuildMI(BB, PPCOpcode, 2).addReg(PPC::CR0).addMBB(copy1MBB);
   MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
   F->getBasicBlockList().insert(It, copy0MBB);
   BuildMI(BB, PPC::B, 1).addMBB(copy0MBB);


Index: llvm/lib/Target/PowerPC/PowerPCInstrInfo.td
diff -u llvm/lib/Target/PowerPC/PowerPCInstrInfo.td:1.25 llvm/lib/Target/PowerPC/PowerPCInstrInfo.td:1.26
--- llvm/lib/Target/PowerPC/PowerPCInstrInfo.td:1.25	Thu Aug 19 00:20:54 2004
+++ llvm/lib/Target/PowerPC/PowerPCInstrInfo.td	Fri Aug 20 04:56:22 2004
@@ -96,12 +96,14 @@
 def CMPL : XForm_16<"cmpl", 31, 32, 0, 0>;
 def CMPLW : XForm_16_ext<"cmplw", 31, 32, 0, 0>;
 def CMPLD : XForm_16_ext<"cmpld", 31, 32, 1, 0>;
+def CNTLZW : XForm_11<"cntlzw", 31, 26, 0, 0, 0>;
 def CRAND : XLForm_1<"crand", 19, 257, 0, 0>;
 def CRANDC : XLForm_1<"crandc", 19, 129, 0, 0>;
 def CRNOR : XLForm_1<"crnor", 19, 33, 0, 0>;
 def CROR : XLForm_1<"cror", 19, 449, 0, 0>;
 def DIVW : XOForm_1<"divw", 31, 491, 0, 0, 0, 0>;
 def DIVWU : XOForm_1<"divwu", 31, 459, 0, 0, 0, 0>;
+def EQV  : XForm_6<"eqv",  31, 284, 0, 0, 0>;
 def EXTSB : XForm_11<"extsb", 31, 954, 0, 0, 0>;
 def EXTSH : XForm_11<"extsh", 31, 922, 0, 0, 0>;
 def FADD : AForm_2<"fadd", 63, 21, 0, 0, 0>;
@@ -150,6 +152,7 @@
 def NEG : XOForm_3<"neg", 31, 104, 0, 0, 0, 0>;
 def NOR  : XForm_6<"nor",  31, 124, 0, 0, 0>;
 def NOP : DForm_4_zero<"nop", 24, 0, 0, (ops), "nop">;
+def ORC  : XForm_6<"orc",  31, 412, 0, 0, 0>;
 def ORI  : DForm_4<24, 0, 0,
                    (ops GPRC:$dst, GPRC:$src1, u16imm:$src2),
                    "ori $dst, $src1, $src2">;
@@ -171,6 +174,7 @@
 def SRAWI  : XForm_10<"srawi",  31, 824, 0, 0, 0>;
 def SRAD  : XForm_6<"srad",  31, 794, 0, 1, 0>;
 def SRAW  : XForm_6<"sraw",  31, 792, 0, 0, 0>;
+def SRWI : MForm_2<"srwi", 21, 0, 0, 0>;
 def STB : DForm_3<"stb", 38, 0, 0>;
 def STBU : DForm_3<"stbu", 39, 0, 0>;
 def STBX : XForm_8<"stbx", 31, 215, 0, 0>;






More information about the llvm-commits mailing list