X86: Use push-pop for materializing 8-bit immediates for minsize (take 2)

[hans]

Summary

This is the same as http://reviews.llvm.org/D15549, with added logic for avoiding clobbering of the red zone (PR26023).

Please take a look!

[majnemer]

test/CodeGen/X86/powi.ll

32–34 ↗

(On Diff #50946)

Why did the constant in this test change?

[hans]

test/CodeGen/X86/powi.ll

32–34 ↗

(On Diff #50946)

Because with the old patch we'd materialize 15 with push/pop, and I figured just tweaking the constant was more clear and less intrusive than updating the expected instructions.

However, in this new patch we conservatively don't use push/pop here because maybe the function could write to the redzone. I'll restore the test.

[hans]

Restore constant in test/CodeGen/X86/powi.ll

[DavidKreitzer]

FWIW, here is a link to the change set for the original commit of the push/pop optimization: http://reviews.llvm.org/D15549. I just diffed that patch against this new one for the purpose of easier review.

lib/Target/X86/X86InstrInfo.cpp

5369 ↗

(On Diff #51336)

This code runs after the red-zone decision has been made, right? Would it be better to capture that decision as part of the MFI and do a more accurate "uses red zone" check here?

If you agree but don't want to make that change now, a FIXME comment would be helpful.

[hans]

In D18246#382637, @DavidKreitzer wrote:

FWIW, here is a link to the change set for the original commit of the push/pop optimization: http://reviews.llvm.org/D15549. I just diffed that patch against this new one for the purpose of easier review.

Oops, it seems like I linked to the wrong one in the description of this patch.

lib/Target/X86/X86InstrInfo.cpp

5369 ↗

(On Diff #51336)

I wasn't sure how to do this, but I think I've found a way. I'll upload a new patch.

[hans]

Accurately check whether the red zone is used or not.

[DavidKreitzer]

Please fix the >80 char line. Otherwise, LGTM.

lib/Target/X86/X86InstrInfo.cpp

5409 ↗

(On Diff #51573)

It looks like this line is too long.

[DavidKreitzer]

And thanks for making the more precise fix, Hans!

[majnemer]

lib/Target/X86/X86InstrCompiler.td

286 ↗

(On Diff #51573)

Why isn't this NotWin64WithoutFP ?

[hans]

lib/Target/X86/X86InstrCompiler.td

286 ↗

(On Diff #51573)

Looks like I forgot to add it. I'll fix this and add a test when committing.

llvm/trunk/lib/Target/X86/X86FrameLowering.cpp

if (Is64Bit && !Fn->hasFnAttribute(Attribute::NoRedZone) &&
       !TRI->needsStackRealignment(MF) &&
       !MFI->hasVarSizedObjects() &&             // No dynamic alloca.
       !MFI->adjustsStack() &&                   // No calls.
       !IsWin64CC &&                             // Win64 has no Red Zone
       !MFI->hasCopyImplyingStackAdjustment() && // Don't push and pop.
       !MF.shouldSplitStack()) {                 // Regular stack
     uint64_t MinSize = X86FI->getCalleeSavedFrameSize();
     if (HasFP) MinSize += SlotSize;
+    X86FI->setUsesRedZone(MinSize > 0 || StackSize > 0);
     StackSize = std::max(MinSize, StackSize > 128 ? StackSize - 128 : 0);
     MFI->setStackSize(StackSize);
   }
 
   // Insert stack pointer adjustment for later moving of return addr.  Only
   // applies to tail call optimized functions where the callee argument stack
   // size is bigger than the callers.
   if (TailCallReturnAddrDelta < 0) {

llvm/trunk/lib/Target/X86/X86ISelDAGToDAG.cpp

class X86DAGToDAGISel final : public SelectionDAGISel {
     /// Keep a pointer to the X86Subtarget around so that we can
     /// make the right decision when generating code for different targets.
     const X86Subtarget *Subtarget;
 
     /// If true, selector should try to optimize for code size instead of
     /// performance.
     bool OptForSize;
 
+    /// If true, selector should try to optimize for minimum code size.
+    bool OptForMinSize;
+
   public:
     explicit X86DAGToDAGISel(X86TargetMachine &tm, CodeGenOpt::Level OptLevel)
-        : SelectionDAGISel(tm, OptLevel), OptForSize(false) {}
+        : SelectionDAGISel(tm, OptLevel), OptForSize(false),
+          OptForMinSize(false) {}
 
     const char *getPassName() const override {
       return "X86 DAG->DAG Instruction Selection";
     }
 
     bool runOnMachineFunction(MachineFunction &MF) override {
       // Reset the subtarget each time through.
       Subtarget = &MF.getSubtarget<X86Subtarget>();
static bool isCalleeLoad(SDValue Callee, SDValue &Chain, bool HasCallSeq) {
   if (Chain.getOperand(0).getOpcode() == ISD::TokenFactor &&
       Callee.getValue(1).isOperandOf(Chain.getOperand(0).getNode()) &&
       Callee.getValue(1).hasOneUse())
     return true;
   return false;
 }
 
 void X86DAGToDAGISel::PreprocessISelDAG() {
-  // OptForSize is used in pattern predicates that isel is matching.
+  // OptFor[Min]Size are used in pattern predicates that isel is matching.
   OptForSize = MF->getFunction()->optForSize();
+  OptForMinSize = MF->getFunction()->optForMinSize();
+  assert((!OptForMinSize || OptForSize) && "OptForMinSize implies OptForSize");
 
   for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
        E = CurDAG->allnodes_end(); I != E; ) {
     SDNode *N = &*I++; // Preincrement iterator to avoid invalidation issues.
 
     if (OptLevel != CodeGenOpt::None &&
         // Only does this when target favors doesn't favor register indirect
         // call.

llvm/trunk/lib/Target/X86/X86InstrCompiler.td

 
 //===----------------------------------------------------------------------===//
 // Alias Instructions
 //===----------------------------------------------------------------------===//
 
 // Alias instruction mapping movr0 to xor.
 // FIXME: remove when we can teach regalloc that xor reg, reg is ok.
 let Defs = [EFLAGS], isReMaterializable = 1, isAsCheapAsAMove = 1,
-    isPseudo = 1 in
+    isPseudo = 1, AddedComplexity = 20 in
 def MOV32r0  : I<0, Pseudo, (outs GR32:$dst), (ins), "",
                  [(set GR32:$dst, 0)], IIC_ALU_NONMEM>, Sched<[WriteZero]>;
 
 // Other widths can also make use of the 32-bit xor, which may have a smaller
 // encoding and avoid partial register updates.
 def : Pat<(i8 0), (EXTRACT_SUBREG (MOV32r0), sub_8bit)>;
 def : Pat<(i16 0), (EXTRACT_SUBREG (MOV32r0), sub_16bit)>;
 def : Pat<(i64 0), (SUBREG_TO_REG (i64 0), (MOV32r0), sub_32bit)> {
   let AddedComplexity = 20;
 }
 
 let Predicates = [OptForSize, NotSlowIncDec, Not64BitMode],
-    AddedComplexity = 1 in {
+    AddedComplexity = 15 in {
   // Pseudo instructions for materializing 1 and -1 using XOR+INC/DEC,
   // which only require 3 bytes compared to MOV32ri which requires 5.
   let Defs = [EFLAGS], isReMaterializable = 1, isPseudo = 1 in {
     def MOV32r1 : I<0, Pseudo, (outs GR32:$dst), (ins), "",
                         [(set GR32:$dst, 1)]>;
     def MOV32r_1 : I<0, Pseudo, (outs GR32:$dst), (ins), "",
                         [(set GR32:$dst, -1)]>;
   }
 
   // MOV16ri is 4 bytes, so the instructions above are smaller.
   def : Pat<(i16 1), (EXTRACT_SUBREG (MOV32r1), sub_16bit)>;
   def : Pat<(i16 -1), (EXTRACT_SUBREG (MOV32r_1), sub_16bit)>;
 }
 
+let isReMaterializable = 1, isPseudo = 1, AddedComplexity = 10 in {
+// AddedComplexity higher than MOV64ri but lower than MOV32r0 and MOV32r1.
+// FIXME: Add itinerary class and Schedule.
+def MOV32ImmSExti8 : I<0, Pseudo, (outs GR32:$dst), (ins i32i8imm:$src), "",
+                       [(set GR32:$dst, i32immSExt8:$src)]>,
+                     Requires<[OptForMinSize, NotWin64WithoutFP]>;
+def MOV64ImmSExti8 : I<0, Pseudo, (outs GR64:$dst), (ins i64i8imm:$src), "",
+                       [(set GR64:$dst, i64immSExt8:$src)]>,
+                     Requires<[OptForMinSize, NotWin64WithoutFP]>;
+}
+
 // Materialize i64 constant where top 32-bits are zero. This could theoretically
 // use MOV32ri with a SUBREG_TO_REG to represent the zero-extension, however
 // that would make it more difficult to rematerialize.
 let isReMaterializable = 1, isAsCheapAsAMove = 1,
     isPseudo = 1, hasSideEffects = 0 in
 def MOV32ri64 : I<0, Pseudo, (outs GR32:$dst), (ins i64i32imm:$src), "", []>;
 
 // This 64-bit pseudo-move can be used for both a 64-bit constant that is

llvm/trunk/lib/Target/X86/X86InstrInfo.h

 #include "X86RegisterInfo.h"
 #include "llvm/ADT/DenseMap.h"
 #include "llvm/Target/TargetInstrInfo.h"
 
 #define GET_INSTRINFO_HEADER
 #include "X86GenInstrInfo.inc"
 
 namespace llvm {
+  class MachineInstrBuilder;
   class X86RegisterInfo;
   class X86Subtarget;
 
 namespace X86 {
   // X86 specific condition code. These correspond to X86_*_COND in
   // X86InstrInfo.td. They must be kept in synch.
 enum CondCode {
   COND_A = 0,
MachineInstr *foldMemoryOperandCustom(MachineFunction &MF, MachineInstr *MI,
                                         ArrayRef<MachineOperand> MOs,
                                         MachineBasicBlock::iterator InsertPt,
                                         unsigned Size, unsigned Align) const;
 
   /// isFrameOperand - Return true and the FrameIndex if the specified
   /// operand and follow operands form a reference to the stack frame.
   bool isFrameOperand(const MachineInstr *MI, unsigned int Op,
                       int &FrameIndex) const;
+
+  /// Expand the MOVImmSExti8 pseudo-instructions.
+  bool ExpandMOVImmSExti8(MachineInstrBuilder &MIB) const;
 };
 
 } // End llvm namespace
 
 #endif

llvm/trunk/lib/Target/X86/X86InstrInfo.cpp

 #include "X86Subtarget.h"
 #include "X86TargetMachine.h"
 #include "llvm/ADT/STLExtras.h"
 #include "llvm/CodeGen/LiveVariables.h"
 #include "llvm/CodeGen/MachineConstantPool.h"
 #include "llvm/CodeGen/MachineDominators.h"
 #include "llvm/CodeGen/MachineFrameInfo.h"
 #include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineModuleInfo.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
 #include "llvm/CodeGen/StackMaps.h"
 #include "llvm/IR/DerivedTypes.h"
 #include "llvm/IR/Function.h"
 #include "llvm/IR/LLVMContext.h"
 #include "llvm/MC/MCAsmInfo.h"
 #include "llvm/MC/MCExpr.h"
 #include "llvm/MC/MCInst.h"
static bool expandMOV32r1(MachineInstrBuilder &MIB, const TargetInstrInfo &TII,
 
   // Turn the pseudo into an INC or DEC.
   MIB->setDesc(TII.get(MinusOne ? X86::DEC32r : X86::INC32r));
   MIB.addReg(Reg);
 
   return true;
 }
 
+bool X86InstrInfo::ExpandMOVImmSExti8(MachineInstrBuilder &MIB) const {
+  MachineBasicBlock &MBB = *MIB->getParent();
+  DebugLoc DL = MIB->getDebugLoc();
+  int64_t Imm = MIB->getOperand(1).getImm();
+  assert(Imm != 0 && "Using push/pop for 0 is not efficient.");
+  MachineBasicBlock::iterator I = MIB.getInstr();
+
+  int StackAdjustment;
+
+  if (Subtarget.is64Bit()) {
+    assert(MIB->getOpcode() == X86::MOV64ImmSExti8 ||
+           MIB->getOpcode() == X86::MOV32ImmSExti8);
+
+    // Can't use push/pop lowering if the function might write to the red zone.
+    X86MachineFunctionInfo *X86FI =
+        MBB.getParent()->getInfo<X86MachineFunctionInfo>();
+    if (X86FI->getUsesRedZone()) {
+      MIB->setDesc(get(MIB->getOpcode() == X86::MOV32ImmSExti8 ? X86::MOV32ri
+                                                               : X86::MOV64ri));
+      return true;
+    }
+
+    // 64-bit mode doesn't have 32-bit push/pop, so use 64-bit operations and
+    // widen the register if necessary.
+    StackAdjustment = 8;
+    BuildMI(MBB, I, DL, get(X86::PUSH64i8)).addImm(Imm);
+    MIB->setDesc(get(X86::POP64r));
+    MIB->getOperand(0)
+        .setReg(getX86SubSuperRegister(MIB->getOperand(0).getReg(), 64));
+  } else {
+    assert(MIB->getOpcode() == X86::MOV32ImmSExti8);
+    StackAdjustment = 4;
+    BuildMI(MBB, I, DL, get(X86::PUSH32i8)).addImm(Imm);
+    MIB->setDesc(get(X86::POP32r));
+  }
+
+  // Build CFI if necessary.
+  MachineFunction &MF = *MBB.getParent();
+  const X86FrameLowering *TFL = Subtarget.getFrameLowering();
+  bool IsWin64Prologue = MF.getTarget().getMCAsmInfo()->usesWindowsCFI();
+  bool NeedsDwarfCFI =
+      !IsWin64Prologue &&
+      (MF.getMMI().hasDebugInfo() || MF.getFunction()->needsUnwindTableEntry());
+  bool EmitCFI = !TFL->hasFP(MF) && NeedsDwarfCFI;
+  if (EmitCFI) {
+    TFL->BuildCFI(MBB, I, DL,
+        MCCFIInstruction::createAdjustCfaOffset(nullptr, StackAdjustment));
+    TFL->BuildCFI(MBB, std::next(I), DL,
+        MCCFIInstruction::createAdjustCfaOffset(nullptr, -StackAdjustment));
+  }
+
+  return true;
+}
+
 // LoadStackGuard has so far only been implemented for 64-bit MachO. Different
 // code sequence is needed for other targets.
 static void expandLoadStackGuard(MachineInstrBuilder &MIB,
                                  const TargetInstrInfo &TII) {
   MachineBasicBlock &MBB = *MIB->getParent();
   DebugLoc DL = MIB->getDebugLoc();
   unsigned Reg = MIB->getOperand(0).getReg();
   const GlobalValue *GV =
bool X86InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
   MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
   switch (MI->getOpcode()) {
   case X86::MOV32r0:
     return Expand2AddrUndef(MIB, get(X86::XOR32rr));
   case X86::MOV32r1:
     return expandMOV32r1(MIB, *this, /*MinusOne=*/ false);
   case X86::MOV32r_1:
     return expandMOV32r1(MIB, *this, /*MinusOne=*/ true);
+  case X86::MOV32ImmSExti8:
+  case X86::MOV64ImmSExti8:
+    return ExpandMOVImmSExti8(MIB);
   case X86::SETB_C8r:
     return Expand2AddrUndef(MIB, get(X86::SBB8rr));
   case X86::SETB_C16r:
     return Expand2AddrUndef(MIB, get(X86::SBB16rr));
   case X86::SETB_C32r:
     return Expand2AddrUndef(MIB, get(X86::SBB32rr));
   case X86::SETB_C64r:
     return Expand2AddrUndef(MIB, get(X86::SBB64rr));

llvm/trunk/lib/Target/X86/X86InstrInfo.td

 def In16BitMode  : Predicate<"Subtarget->is16Bit()">,
                              AssemblerPredicate<"Mode16Bit", "16-bit mode">;
 def Not16BitMode : Predicate<"!Subtarget->is16Bit()">,
                              AssemblerPredicate<"!Mode16Bit", "Not 16-bit mode">;
 def In32BitMode  : Predicate<"Subtarget->is32Bit()">,
                              AssemblerPredicate<"Mode32Bit", "32-bit mode">;
 def IsWin64      : Predicate<"Subtarget->isTargetWin64()">;
 def NotWin64     : Predicate<"!Subtarget->isTargetWin64()">;
+def NotWin64WithoutFP : Predicate<"!Subtarget->isTargetWin64() ||"
+                                  "Subtarget->getFrameLowering()->hasFP(*MF)">;
 def IsPS4        : Predicate<"Subtarget->isTargetPS4()">;
 def NotPS4       : Predicate<"!Subtarget->isTargetPS4()">;
 def IsNaCl       : Predicate<"Subtarget->isTargetNaCl()">;
 def NotNaCl      : Predicate<"!Subtarget->isTargetNaCl()">;
 def SmallCode    : Predicate<"TM.getCodeModel() == CodeModel::Small">;
 def KernelCode   : Predicate<"TM.getCodeModel() == CodeModel::Kernel">;
 def FarData      : Predicate<"TM.getCodeModel() != CodeModel::Small &&"
                              "TM.getCodeModel() != CodeModel::Kernel">;
 def NearData     : Predicate<"TM.getCodeModel() == CodeModel::Small ||"
                              "TM.getCodeModel() == CodeModel::Kernel">;
 def IsStatic     : Predicate<"TM.getRelocationModel() == Reloc::Static">;
 def IsNotPIC     : Predicate<"TM.getRelocationModel() != Reloc::PIC_">;
 def OptForSize   : Predicate<"OptForSize">;
+def OptForMinSize : Predicate<"OptForMinSize">;
 def OptForSpeed  : Predicate<"!OptForSize">;
 def FastBTMem    : Predicate<"!Subtarget->isBTMemSlow()">;
 def CallImmAddr  : Predicate<"Subtarget->IsLegalToCallImmediateAddr(TM)">;
 def FavorMemIndirectCall  : Predicate<"!Subtarget->callRegIndirect()">;
 def NotSlowIncDec : Predicate<"!Subtarget->slowIncDec()">;
 def HasFastMem32 : Predicate<"!Subtarget->isUnalignedMem32Slow()">;
 def HasMFence    : Predicate<"Subtarget->hasMFence()">;
 

llvm/trunk/lib/Target/X86/X86MachineFunctionInfo.h

class X86MachineFunctionInfo : public MachineFunctionInfo {
   /// The frame index of a stack object containing the original frame pointer
   /// used to address arguments in a function using a base pointer.
   int SEHFramePtrSaveIndex = 0;
 
   /// True if this function has a subset of CSRs that is handled explicitly via
   /// copies.
   bool IsSplitCSR = false;
 
+  /// True if this function uses the red zone.
+  bool UsesRedZone = false;
+
 private:
   /// ForwardedMustTailRegParms - A list of virtual and physical registers
   /// that must be forwarded to every musttail call.
   SmallVector<ForwardedRegister, 1> ForwardedMustTailRegParms;
 
 public:
   X86MachineFunctionInfo() = default;
 
public:
   void setSEHFramePtrSaveIndex(int Index) { SEHFramePtrSaveIndex = Index; }
 
   SmallVectorImpl<ForwardedRegister> &getForwardedMustTailRegParms() {
     return ForwardedMustTailRegParms;
   }
 
   bool isSplitCSR() const { return IsSplitCSR; }
   void setIsSplitCSR(bool s) { IsSplitCSR = s; }
+
+  bool getUsesRedZone() const { return UsesRedZone; }
+  void setUsesRedZone(bool V) { UsesRedZone = V; }
 };
 
 } // End llvm namespace
 
 #endif

llvm/trunk/test/CodeGen/X86/materialize-one.ll

-; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=+cmov %s -o - | FileCheck %s --check-prefix=CHECK32
-; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+cmov %s -o - | FileCheck %s --check-prefix=CHECK64
-
-define i32 @one32() optsize {
-entry:
-  ret i32 1
-
-; CHECK32-LABEL: one32
-; CHECK32:       xorl %eax, %eax
-; CHECK32-NEXT:  incl %eax
-; CHECK32-NEXT:  ret
-
-; FIXME: Figure out the best approach in 64-bit mode.
-; CHECK64-LABEL: one32
-; CHECK64:       movl $1, %eax
-; CHECK64-NEXT:  retq
-}
-
-define i32 @minus_one32() optsize {
-entry:
-  ret i32 -1
-
-; CHECK32-LABEL: minus_one32
-; CHECK32:       xorl %eax, %eax
-; CHECK32-NEXT:  decl %eax
-; CHECK32-NEXT:  ret
-}
-
-define i16 @one16() optsize {
-entry:
-  ret i16 1
-
-; CHECK32-LABEL: one16
-; CHECK32:       xorl %eax, %eax
-; CHECK32-NEXT:  incl %eax
-; CHECK32-NEXT:  retl
-}
-
-define i16 @minus_one16() optsize {
-entry:
-  ret i16 -1
-
-; CHECK32-LABEL: minus_one16
-; CHECK32:       xorl %eax, %eax
-; CHECK32-NEXT:  decl %eax
-; CHECK32-NEXT:  retl
-}
-
-define i32 @test_rematerialization() optsize {
-entry:
-  ; Materialize -1 (thiscall forces it into %ecx).
-  tail call x86_thiscallcc void @f(i32 -1)
-
-  ; Clobber all registers except %esp, leaving nowhere to store the -1 besides
-  ; spilling it to the stack.
-  tail call void asm sideeffect "", "~{eax},~{ebx},~{ecx},~{edx},~{edi},~{esi},~{ebp},~{dirflag},~{fpsr},~{flags}"()
-
-  ; -1 should be re-materialized here instead of getting spilled above.
-  ret i32 -1
-
-; CHECK32-LABEL: test_rematerialization
-; CHECK32:       xorl %ecx, %ecx
-; CHECK32-NEXT:  decl %ecx
-; CHECK32:       calll
-; CHECK32:       xorl %eax, %eax
-; CHECK32-NEXT:  decl %eax
-; CHECK32-NOT:   %eax
-; CHECK32:       retl
-}
-
-define i32 @test_rematerialization2(i32 %x) optsize {
-entry:
-  ; Materialize -1 (thiscall forces it into %ecx).
-  tail call x86_thiscallcc void @f(i32 -1)
-
-  ; Clobber all registers except %esp, leaving nowhere to store the -1 besides
-  ; spilling it to the stack.
-  tail call void asm sideeffect "", "~{eax},~{ebx},~{ecx},~{edx},~{edi},~{esi},~{ebp},~{dirflag},~{fpsr},~{flags}"()
-
-  ; Define eflags.
-  %a = icmp ne i32 %x, 123
-  %b = zext i1 %a to i32
-  ; Cause -1 to be rematerialized right in front of the cmov, which needs eflags.
-  ; It must therefore not use the xor-dec lowering.
-  %c = select i1 %a, i32 %b, i32 -1
-  ret i32 %c
-
-; CHECK32-LABEL: test_rematerialization2
-; CHECK32:       xorl %ecx, %ecx
-; CHECK32-NEXT:  decl %ecx
-; CHECK32:       calll
-; CHECK32:       cmpl
-; CHECK32:       setne
-; CHECK32-NOT:   xorl
-; CHECK32:       movl $-1
-; CHECK32:       cmov
-; CHECK32:       retl
-}
-
-declare x86_thiscallcc void @f(i32)

llvm/trunk/test/CodeGen/X86/materialize.ll

+; RUN: llc -mtriple=i686-unknown-linux-gnu -mattr=+cmov %s -o - | FileCheck %s --check-prefix=CHECK32
+; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+cmov %s -o - | FileCheck %s --check-prefix=CHECK64
+; RUN: llc -mtriple=x86_64-pc-win32 -mattr=+cmov %s -o - | FileCheck %s --check-prefix=CHECKWIN64
+
+define i32 @one32_nooptsize() {
+entry:
+  ret i32 1
+
+; When not optimizing for size, use mov.
+; CHECK32-LABEL: one32_nooptsize:
+; CHECK32:       movl $1, %eax
+; CHECK32-NEXT:  retl
+; CHECK64-LABEL: one32_nooptsize:
+; CHECK64:       movl $1, %eax
+; CHECK64-NEXT:  retq
+}
+
+define i32 @one32() optsize {
+entry:
+  ret i32 1
+
+; CHECK32-LABEL: one32:
+; CHECK32:       xorl %eax, %eax
+; CHECK32-NEXT:  incl %eax
+; CHECK32-NEXT:  retl
+
+; FIXME: Figure out the best approach in 64-bit mode.
+; CHECK64-LABEL: one32:
+; CHECK64:       movl $1, %eax
+; CHECK64-NEXT:  retq
+}
+
+define i32 @one32_minsize() minsize {
+entry:
+  ret i32 1
+
+; On 32-bit, xor-inc is preferred over push-pop.
+; CHECK32-LABEL: one32_minsize:
+; CHECK32:       xorl %eax, %eax
+; CHECK32-NEXT:  incl %eax
+; CHECK32-NEXT:  retl
+
+; On 64-bit we don't do xor-inc yet, so push-pop it is. Note that we have to
+; pop into a 64-bit register even when we just need 32 bits.
+; CHECK64-LABEL: one32_minsize:
+; CHECK64:       pushq $1
+; CHECK64:       .cfi_adjust_cfa_offset 8
+; CHECK64:       popq %rax
+; CHECK64:       .cfi_adjust_cfa_offset -8
+; CHECK64-NEXT:  retq
+
+; On Win64 we can't adjust the stack unless there's a frame pointer.
+; CHECKWIN64-LABEL: one32_minsize:
+; CHECKWIN64:       movl $1, %eax
+; CHECKWIN64-NEXT:  retq
+}
+
+define i32 @pr26023() minsize {
+entry:
+  %x = alloca [120 x i8]
+  %0 = getelementptr inbounds [120 x i8], [120 x i8]* %x, i64 0, i64 0
+  call void asm sideeffect "", "imr,~{memory},~{dirflag},~{fpsr},~{flags}"(i8* %0)
+  %arrayidx = getelementptr inbounds [120 x i8], [120 x i8]* %x, i64 0, i64 119
+  store volatile i8 -2, i8* %arrayidx
+  call void asm sideeffect "", "r,~{dirflag},~{fpsr},~{flags}"(i32 5)
+  %1 = load volatile i8, i8* %arrayidx
+  %conv = sext i8 %1 to i32
+  ret i32 %conv
+
+; The function writes to the redzone, so push/pop cannot be used.
+; CHECK64-LABEL: pr26023:
+; CHECK64:       movl $5, %ecx
+; CHECK64:       retq
+
+; 32-bit X86 doesn't have a redzone.
+; CHECK32-LABEL: pr26023:
+; CHECK32:       pushl $5
+; CHECK32:       popl %ecx
+; CHECK32:       retl
+}
+
+
+define i64 @one64_minsize() minsize {
+entry:
+  ret i64 1
+; On 64-bit we don't do xor-inc yet, so push-pop it is.
+; CHECK64-LABEL: one64_minsize:
+; CHECK64:       pushq $1
+; CHECK64:       .cfi_adjust_cfa_offset 8
+; CHECK64:       popq %rax
+; CHECK64:       .cfi_adjust_cfa_offset -8
+; CHECK64-NEXT:  retq
+
+; On Win64 we can't adjust the stack unless there's a frame pointer.
+; CHECKWIN64-LABEL: one64_minsize:
+; CHECKWIN64:       movl $1, %eax
+; CHECKWIN64-NEXT:  retq
+}
+
+define i32 @minus_one32() optsize {
+entry:
+  ret i32 -1
+
+; CHECK32-LABEL: minus_one32:
+; CHECK32:       xorl %eax, %eax
+; CHECK32-NEXT:  decl %eax
+; CHECK32-NEXT:  retl
+}
+
+define i32 @minus_one32_minsize() minsize {
+entry:
+  ret i32 -1
+
+; xor-dec is preferred over push-pop.
+; CHECK32-LABEL: minus_one32_minsize:
+; CHECK32:       xorl %eax, %eax
+; CHECK32-NEXT:  decl %eax
+; CHECK32-NEXT:  retl
+}
+
+define i16 @one16() optsize {
+entry:
+  ret i16 1
+
+; CHECK32-LABEL: one16:
+; CHECK32:       xorl %eax, %eax
+; CHECK32-NEXT:  incl %eax
+; CHECK32-NEXT:  retl
+}
+
+define i16 @minus_one16() optsize {
+entry:
+  ret i16 -1
+
+; CHECK32-LABEL: minus_one16:
+; CHECK32:       xorl %eax, %eax
+; CHECK32-NEXT:  decl %eax
+; CHECK32-NEXT:  retl
+}
+
+define i32 @minus_five32() minsize {
+entry:
+  ret i32 -5
+
+; CHECK32-LABEL: minus_five32:
+; CHECK32: pushl $-5
+; CHECK32: popl %eax
+; CHECK32: retl
+}
+
+define i64 @minus_five64() minsize {
+entry:
+  ret i64 -5
+
+; CHECK64-LABEL: minus_five64:
+; CHECK64: pushq $-5
+; CHECK64:       .cfi_adjust_cfa_offset 8
+; CHECK64: popq %rax
+; CHECK64:       .cfi_adjust_cfa_offset -8
+; CHECK64: retq
+}
+
+define i32 @rematerialize_minus_one() optsize {
+entry:
+  ; Materialize -1 (thiscall forces it into %ecx).
+  tail call x86_thiscallcc void @f(i32 -1)
+
+  ; Clobber all registers except %esp, leaving nowhere to store the -1 besides
+  ; spilling it to the stack.
+  tail call void asm sideeffect "", "~{eax},~{ebx},~{ecx},~{edx},~{edi},~{esi},~{ebp},~{dirflag},~{fpsr},~{flags}"()
+
+  ; -1 should be re-materialized here instead of getting spilled above.
+  ret i32 -1
+
+; CHECK32-LABEL: rematerialize_minus_one
+; CHECK32:       xorl %ecx, %ecx
+; CHECK32-NEXT:  decl %ecx
+; CHECK32:       calll
+; CHECK32:       xorl %eax, %eax
+; CHECK32-NEXT:  decl %eax
+; CHECK32-NOT:   %eax
+; CHECK32:       retl
+}
+
+define i32 @rematerialize_minus_one_eflags(i32 %x) optsize {
+entry:
+  ; Materialize -1 (thiscall forces it into %ecx).
+  tail call x86_thiscallcc void @f(i32 -1)
+
+  ; Clobber all registers except %esp, leaving nowhere to store the -1 besides
+  ; spilling it to the stack.
+  tail call void asm sideeffect "", "~{eax},~{ebx},~{ecx},~{edx},~{edi},~{esi},~{ebp},~{dirflag},~{fpsr},~{flags}"()
+
+  ; Define eflags.
+  %a = icmp ne i32 %x, 123
+  %b = zext i1 %a to i32
+  ; Cause -1 to be rematerialized right in front of the cmov, which needs eflags.
+  ; It must therefore not use the xor-dec lowering.
+  %c = select i1 %a, i32 %b, i32 -1
+  ret i32 %c
+
+; CHECK32-LABEL: rematerialize_minus_one_eflags
+; CHECK32:       xorl %ecx, %ecx
+; CHECK32-NEXT:  decl %ecx
+; CHECK32:       calll
+; CHECK32:       cmpl
+; CHECK32:       setne
+; CHECK32-NOT:   xorl
+; CHECK32:       movl $-1
+; CHECK32:       cmov
+; CHECK32:       retl
+}
+
+declare x86_thiscallcc void @f(i32)

llvm/trunk/test/CodeGen/X86/powi.ll

 ; CHECK-NEXT:    jmp
   %ret = tail call double @llvm.powi.f64(double %a, i32 15) nounwind ; <double> [#uses=1]
   ret double %ret
 }
 
 define double @pow_wrapper_minsize(double %a) minsize {
 ; CHECK-LABEL: pow_wrapper_minsize:
 ; CHECK:       # BB#0:
-; CHECK-NEXT:    movl  $15, %edi
-; CHECK-NEXT:    jmp
+; CHECK-NEXT:    pushq $15
+; CHECK:         popq %rdi
+; CHECK:         jmp
   %ret = tail call double @llvm.powi.f64(double %a, i32 15) nounwind ; <double> [#uses=1]
   ret double %ret
 }
 
 declare double @llvm.powi.f64(double, i32) nounwind readonly