Use predication instead of pseudo-opcodes when folding into MOVCC.

Now that it is possible to dynamically tie MachineInstr operands,
predicated instructions are possible in SSA form:

  %vreg3<def> = SUBri %vreg1, -2147483647, pred:14, pred:%noreg, %opt:%noreg
  %vreg4<def,tied1> = MOVCCr %vreg3<tied0>, %vreg1, %pred:12, pred:%CPSR

Becomes a predicated SUBri with a tied imp-use:

  SUBri %vreg1, -2147483647, pred:13, pred:%CPSR, opt:%noreg, %vreg1<imp-use,tied0>

This means that any instruction that is safe to move can be folded into
a MOVCC, and the *CC pseudo-instructions are no longer needed.

The test case changes reflect that Thumb2SizeReduce recognizes the
predicated instructions. It didn't understand the pseudos.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@163274 91177308-0d34-0410-b5e6-96231b3b80d8

4 files changed, 41 insertions, 66 deletions

diff --git a/lib/Target/ARM/ARMBaseInstrInfo.cpp b/lib/Target/ARM/ARMBaseInstrInfo.cpp
index f8e554fcbf..10c41ee6c4 100644
--- a/lib/Target/ARM/ARMBaseInstrInfo.cpp
+++ b/lib/Target/ARM/ARMBaseInstrInfo.cpp
@@ -1580,16 +1580,20 @@ ARMBaseInstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
 }
 
 /// Identify instructions that can be folded into a MOVCC instruction, and
-/// return the corresponding opcode for the predicated pseudo-instruction.
-static unsigned canFoldIntoMOVCC(unsigned Reg, MachineInstr *&MI,
-                                 const MachineRegisterInfo &MRI) {
+/// return the defining instruction.
+static MachineInstr *canFoldIntoMOVCC(unsigned Reg,
+                                      const MachineRegisterInfo &MRI,
+                                      const TargetInstrInfo *TII) {
   if (!TargetRegisterInfo::isVirtualRegister(Reg))
     return 0;
   if (!MRI.hasOneNonDBGUse(Reg))
     return 0;
-  MI = MRI.getVRegDef(Reg);
+  MachineInstr *MI = MRI.getVRegDef(Reg);
   if (!MI)
     return 0;
+  // MI is folded into the MOVCC by predicating it.
+  if (!MI->isPredicable())
+    return 0;
   // Check if MI has any non-dead defs or physreg uses. This also detects
   // predicated instructions which will be reading CPSR.
   for (unsigned i = 1, e = MI->getNumOperands(); i != e; ++i) {
@@ -1599,55 +1603,18 @@ static unsigned canFoldIntoMOVCC(unsigned Reg, MachineInstr *&MI,
       return 0;
     if (!MO.isReg())
       continue;
+    // MI can't have any tied operands, that would conflict with predication.
+    if (MO.isTied())
+      return 0;
     if (TargetRegisterInfo::isPhysicalRegister(MO.getReg()))
       return 0;
     if (MO.isDef() && !MO.isDead())
       return 0;
   }
-  switch (MI->getOpcode()) {
-  default: return 0;
-  case ARM::ANDri:   return ARM::ANDCCri;
-  case ARM::ANDrr:   return ARM::ANDCCrr;
-  case ARM::ANDrsi:  return ARM::ANDCCrsi;
-  case ARM::ANDrsr:  return ARM::ANDCCrsr;
-  case ARM::t2ANDri: return ARM::t2ANDCCri;
-  case ARM::t2ANDrr: return ARM::t2ANDCCrr;
-  case ARM::t2ANDrs: return ARM::t2ANDCCrs;
-  case ARM::EORri:   return ARM::EORCCri;
-  case ARM::EORrr:   return ARM::EORCCrr;
-  case ARM::EORrsi:  return ARM::EORCCrsi;
-  case ARM::EORrsr:  return ARM::EORCCrsr;
-  case ARM::t2EORri: return ARM::t2EORCCri;
-  case ARM::t2EORrr: return ARM::t2EORCCrr;
-  case ARM::t2EORrs: return ARM::t2EORCCrs;
-  case ARM::ORRri:   return ARM::ORRCCri;
-  case ARM::ORRrr:   return ARM::ORRCCrr;
-  case ARM::ORRrsi:  return ARM::ORRCCrsi;
-  case ARM::ORRrsr:  return ARM::ORRCCrsr;
-  case ARM::t2ORRri: return ARM::t2ORRCCri;
-  case ARM::t2ORRrr: return ARM::t2ORRCCrr;
-  case ARM::t2ORRrs: return ARM::t2ORRCCrs;
-
-  // ARM ADD/SUB
-  case ARM::ADDri:   return ARM::ADDCCri;
-  case ARM::ADDrr:   return ARM::ADDCCrr;
-  case ARM::ADDrsi:  return ARM::ADDCCrsi;
-  case ARM::ADDrsr:  return ARM::ADDCCrsr;
-  case ARM::SUBri:   return ARM::SUBCCri;
-  case ARM::SUBrr:   return ARM::SUBCCrr;
-  case ARM::SUBrsi:  return ARM::SUBCCrsi;
-  case ARM::SUBrsr:  return ARM::SUBCCrsr;
-
-  // Thumb2 ADD/SUB
-  case ARM::t2ADDri:   return ARM::t2ADDCCri;
-  case ARM::t2ADDri12: return ARM::t2ADDCCri12;
-  case ARM::t2ADDrr:   return ARM::t2ADDCCrr;
-  case ARM::t2ADDrs:   return ARM::t2ADDCCrs;
-  case ARM::t2SUBri:   return ARM::t2SUBCCri;
-  case ARM::t2SUBri12: return ARM::t2SUBCCri12;
-  case ARM::t2SUBrr:   return ARM::t2SUBCCrr;
-  case ARM::t2SUBrs:   return ARM::t2SUBCCrs;
-  }
+  bool DontMoveAcrossStores = true;
+  if (!MI->isSafeToMove(TII, /* AliasAnalysis = */ 0, DontMoveAcrossStores))
+    return 0;
+  return MI;
 }
 
 bool ARMBaseInstrInfo::analyzeSelect(const MachineInstr *MI,
@@ -1676,19 +1643,18 @@ MachineInstr *ARMBaseInstrInfo::optimizeSelect(MachineInstr *MI,
   assert((MI->getOpcode() == ARM::MOVCCr || MI->getOpcode() == ARM::t2MOVCCr) &&
          "Unknown select instruction");
   const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
-  MachineInstr *DefMI = 0;
-  unsigned Opc = canFoldIntoMOVCC(MI->getOperand(2).getReg(), DefMI, MRI);
-  bool Invert = !Opc;
-  if (!Opc)
-    Opc = canFoldIntoMOVCC(MI->getOperand(1).getReg(), DefMI, MRI);
-  if (!Opc)
+  MachineInstr *DefMI = canFoldIntoMOVCC(MI->getOperand(2).getReg(), MRI, this);
+  bool Invert = !DefMI;
+  if (!DefMI)
+    DefMI = canFoldIntoMOVCC(MI->getOperand(1).getReg(), MRI, this);
+  if (!DefMI)
     return 0;
 
   // Create a new predicated version of DefMI.
   // Rfalse is the first use.
   MachineInstrBuilder NewMI = BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
-                                      get(Opc), MI->getOperand(0).getReg())
-    .addOperand(MI->getOperand(Invert ? 2 : 1));
+                                      DefMI->getDesc(),
+                                      MI->getOperand(0).getReg());
 
   // Copy all the DefMI operands, excluding its (null) predicate.
   const MCInstrDesc &DefDesc = DefMI->getDesc();
@@ -1707,6 +1673,15 @@ MachineInstr *ARMBaseInstrInfo::optimizeSelect(MachineInstr *MI,
   if (NewMI->hasOptionalDef())
     AddDefaultCC(NewMI);
 
+  // The output register value when the predicate is false is an implicit
+  // register operand tied to the first def.
+  // The tie makes the register allocator ensure the FalseReg is allocated the
+  // same register as operand 0.
+  MachineOperand FalseReg = MI->getOperand(Invert ? 2 : 1);
+  FalseReg.setImplicit();
+  NewMI->addOperand(FalseReg);
+  NewMI->tieOperands(0, NewMI->getNumOperands() - 1);
+
   // The caller will erase MI, but not DefMI.
   DefMI->eraseFromParent();
   return NewMI;
diff --git a/test/CodeGen/ARM/select.ll b/test/CodeGen/ARM/select.ll
index 5575566628..62708ed53d 100644
--- a/test/CodeGen/ARM/select.ll
+++ b/test/CodeGen/ARM/select.ll
@@ -80,7 +80,7 @@ define double @f7(double %a, double %b) {
 ; CHECK-NEON:      adr     [[R2:r[0-9]+]], LCPI7_0
 ; CHECK-NEON-NEXT: cmp     r0, [[R3]]
 ; CHECK-NEON-NEXT: it      eq
-; CHECK-NEON-NEXT: addeq.w {{r.*}}, [[R2]]
+; CHECK-NEON-NEXT: addeq{{.*}} [[R2]], #4
 ; CHECK-NEON-NEXT: ldr
 ; CHECK-NEON:      bx
 
diff --git a/test/CodeGen/ARM/select_xform.ll b/test/CodeGen/ARM/select_xform.ll
index cfc0e70120..7507808912 100644
--- a/test/CodeGen/ARM/select_xform.ll
+++ b/test/CodeGen/ARM/select_xform.ll
@@ -9,7 +9,7 @@ define i32 @t1(i32 %a, i32 %b, i32 %c) nounwind {
 
 ; T2: t1:
 ; T2: mvn r0, #-2147483648
-; T2: addle.w r1, r1
+; T2: addle r1, r0
 ; T2: mov r0, r1
   %tmp1 = icmp sgt i32 %c, 10
   %tmp2 = select i1 %tmp1, i32 0, i32 2147483647
@@ -23,7 +23,7 @@ define i32 @t2(i32 %a, i32 %b, i32 %c, i32 %d) nounwind {
 ; ARM: mov r0, r1
 
 ; T2: t2:
-; T2: suble.w r1, r1, #10
+; T2: suble r1, #10
 ; T2: mov r0, r1
   %tmp1 = icmp sgt i32 %c, 10
   %tmp2 = select i1 %tmp1, i32 0, i32 10
@@ -37,7 +37,7 @@ define i32 @t3(i32 %a, i32 %b, i32 %x, i32 %y) nounwind {
 ; ARM: mov r0, r3
 
 ; T2: t3:
-; T2: andge.w r3, r3, r2
+; T2: andge r3, r2
 ; T2: mov r0, r3
   %cond = icmp slt i32 %a, %b
   %z = select i1 %cond, i32 -1, i32 %x
@@ -51,7 +51,7 @@ define i32 @t4(i32 %a, i32 %b, i32 %x, i32 %y) nounwind {
 ; ARM: mov r0, r3
 
 ; T2: t4:
-; T2: orrge.w r3, r3, r2
+; T2: orrge r3, r2
 ; T2: mov r0, r3
   %cond = icmp slt i32 %a, %b
   %z = select i1 %cond, i32 0, i32 %x
@@ -81,7 +81,7 @@ define i32 @t6(i32 %a, i32 %b, i32 %c, i32 %d) nounwind {
 
 ; T2: t6:
 ; T2-NOT: movge
-; T2: eorlt.w r3, r3, r2
+; T2: eorlt r3, r2
   %cond = icmp slt i32 %a, %b
   %tmp1 = select i1 %cond, i32 %c, i32 0
   %tmp2 = xor i32 %tmp1, %d
@@ -200,7 +200,7 @@ entry:
 
 ; T2: t13
 ; T2: cmp r1, #10
-; T2: addgt.w r0, r0, #1
+; T2: addgt r0, #1
   %cmp = icmp sgt i32 %a, 10
   %conv = zext i1 %cmp to i32
   %add = add i32 %conv, %c
@@ -216,7 +216,7 @@ entry:
 
 ; T2: t14
 ; T2: cmp r1, #10
-; T2: subgt.w r0, r0, #1
+; T2: subgt r0, #1
   %cmp = icmp sgt i32 %a, 10
   %conv = sext i1 %cmp to i32
   %add = add i32 %conv, %c
diff --git a/test/CodeGen/Thumb2/thumb2-select_xform.ll b/test/CodeGen/Thumb2/thumb2-select_xform.ll
index ead198f216..ed4d26d746 100644
--- a/test/CodeGen/Thumb2/thumb2-select_xform.ll
+++ b/test/CodeGen/Thumb2/thumb2-select_xform.ll
@@ -5,7 +5,7 @@ define i32 @t1(i32 %a, i32 %b, i32 %c) nounwind {
 ; CHECK: mvn r0, #-2147483648
 ; CHECK: cmp r2, #10
 ; CHECK: it  le
-; CHECK: addle.w r1, r1, r0
+; CHECK: addle r1, r0
 ; CHECK: mov r0, r1
         %tmp1 = icmp sgt i32 %c, 10
         %tmp2 = select i1 %tmp1, i32 0, i32 2147483647
@@ -30,7 +30,7 @@ define i32 @t3(i32 %a, i32 %b, i32 %c, i32 %d) nounwind {
 ; CHECK: t3
 ; CHECK: cmp r2, #10
 ; CHECK: it  le
-; CHECK: suble.w r1, r1, #10
+; CHECK: suble r1, #10
 ; CHECK: mov r0, r1
         %tmp1 = icmp sgt i32 %c, 10
         %tmp2 = select i1 %tmp1, i32 0, i32 10