Eliminate stack adjustment for float-returning functions

src/IceInstX86BaseImpl.h

 //===- subzero/src/IceInstX86BaseImpl.h - Generic X86 instructions -*- C++ -*=//
 //
 //                        The Subzero Code Generator
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 ///
 /// \file
@@ skipping 39 matching lines @@
                                         InstArithmetic::OpKind Op,
                                         Variable *Beacon)
     : InstX86Base<Machine>(Func, InstX86Base<Machine>::FakeRMW, 3, nullptr),
       Op(Op) {
   this->addSource(Data);
   this->addSource(Addr);
   this->addSource(Beacon);
 }
 
 template <class Machine>
-InstX86AdjustStack<Machine>::InstX86AdjustStack(Cfg *Func, SizeT Amount,
+InstX86AdjustStack<Machine>::InstX86AdjustStack(Cfg *Func, int32_t Amount,
                                                 Variable *Esp)
     : InstX86Base<Machine>(Func, InstX86Base<Machine>::Adjuststack, 1, Esp),
       Amount(Amount) {
   this->addSource(Esp);
 }
 
 template <class Machine>
 InstX86Mul<Machine>::InstX86Mul(Cfg *Func, Variable *Dest, Variable *Source1,
                                 Operand *Source2)
     : InstX86Base<Machine>(Func, InstX86Base<Machine>::Mul, 2, Dest) {
@@ skipping 503 matching lines @@
   if (const auto *CI = llvm::dyn_cast<ConstantInteger32>(CallTarget)) {
     // Emit without a leading '$'.
     Str << CI->getValue();
   } else if (const auto DirectCallTarget =
                  llvm::dyn_cast<ConstantRelocatable>(CallTarget)) {
     DirectCallTarget->emitWithoutPrefix(Target);
   } else {
     Str << "*";
     CallTarget->emit(Func);
   }
-  Target->resetStackAdjustment();
 }
 
 template <class Machine>
 void InstX86Call<Machine>::emitIAS(const Cfg *Func) const {
   typename InstX86Base<Machine>::Traits::Assembler *Asm =
       Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
   Operand *CallTarget = getCallTarget();
   auto *Target = InstX86Base<Machine>::getTarget(Func);
   if (const auto *Var = llvm::dyn_cast<Variable>(CallTarget)) {
     if (Var->hasReg()) {
       Asm->call(InstX86Base<Machine>::Traits::getEncodedGPR(Var->getRegNum()));
     } else {
       Asm->call(Target->stackVarToAsmOperand(Var));
     }
   } else if (const auto *Mem = llvm::dyn_cast<
                  typename InstX86Base<Machine>::Traits::X86OperandMem>(
                  CallTarget)) {
     assert(Mem->getSegmentRegister() ==
            InstX86Base<Machine>::Traits::X86OperandMem::DefaultSegment);
     Asm->call(Mem->toAsmAddress(Asm, Target));
   } else if (const auto *CR = llvm::dyn_cast<ConstantRelocatable>(CallTarget)) {
     assert(CR->getOffset() == 0 && "We only support calling a function");
     Asm->call(CR);
   } else if (const auto *Imm = llvm::dyn_cast<ConstantInteger32>(CallTarget)) {
     Asm->call(Immediate(Imm->getValue()));
   } else {
     llvm_unreachable("Unexpected operand type");
   }
-  Target->resetStackAdjustment();
 }
 
 template <class Machine>
 void InstX86Call<Machine>::dump(const Cfg *Func) const {
   if (!BuildDefs::dump())
     return;
   Ostream &Str = Func->getContext()->getStrDump();
   if (this->getDest()) {
     this->dumpDest(Func);
     Str << " = ";
@@ skipping 2004 matching lines @@
   Ostream &Str = Func->getContext()->getStrEmit();
   assert(this->getSrcSize() == 0);
   // TODO(jvoung,stichnot): Utilize this by setting Dest to nullptr to
   // "partially" delete the fstp if the Dest is unused. Even if Dest is unused,
   // the fstp should be kept for the SideEffects of popping the stack.
   if (!this->getDest()) {
     Str << "\tfstp\tst(0)";
     return;
   }
   Type Ty = this->getDest()->getType();
-  size_t Width = typeWidthInBytes(Ty);
   if (!this->getDest()->hasReg()) {
     Str << "\tfstp" << this->getFldString(Ty) << "\t";
     this->getDest()->emit(Func);
     return;
   }
   // Dest is a physical (xmm) register, so st(0) needs to go through memory.
-  // Hack this by creating a temporary stack slot, spilling st(0) there,
-  // loading it into the xmm register, and deallocating the stack slot.
-  Str << "\tsubl\t$" << Width << ", %esp\n";
+  // Hack this by caller-reserved memory at the top of stack, spilling st(0)

[Jim Stichnoth, 2015/11/13 23:30:33]

by using

[sehr, 2015/11/14 00:30:56]

Done.

+  // there, and loading it into the xmm register.
   Str << "\tfstp" << this->getFldString(Ty) << "\t"
       << "(%esp)\n";
   Str << "\tmov" << InstX86Base<Machine>::Traits::TypeAttributes[Ty].SdSsString
       << "\t"
       << "(%esp), ";
   this->getDest()->emit(Func);
   Str << "\n";
-  Str << "\taddl\t$" << Width << ", %esp";
 }
 
 template <class Machine>
 void InstX86Fstp<Machine>::emitIAS(const Cfg *Func) const {
   typename InstX86Base<Machine>::Traits::Assembler *Asm =
       Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
   assert(this->getSrcSize() == 0);
   const Variable *Dest = this->getDest();
   // TODO(jvoung,stichnot): Utilize this by setting Dest to nullptr to
   // "partially" delete the fstp if the Dest is unused. Even if Dest is unused,
   // the fstp should be kept for the SideEffects of popping the stack.
   if (!Dest) {
     Asm->fstp(InstX86Base<Machine>::Traits::RegisterSet::getEncodedSTReg(0));
     return;
   }
   auto *Target = InstX86Base<Machine>::getTarget(Func);
   Type Ty = Dest->getType();
   if (!Dest->hasReg()) {
     typename InstX86Base<Machine>::Traits::Address StackAddr(
         Target->stackVarToAsmOperand(Dest));
     Asm->fstp(Ty, StackAddr);
   } else {
     // Dest is a physical (xmm) register, so st(0) needs to go through memory.
-    // Hack this by creating a temporary stack slot, spilling st(0) there,
-    // loading it into the xmm register, and deallocating the stack slot.
-    Immediate Width(typeWidthInBytes(Ty));
-    Asm->sub(IceType_i32,
-             InstX86Base<Machine>::Traits::RegisterSet::Encoded_Reg_esp, Width);
+    // Hack this by using caller-reserved memory at the top of stack, spilling
+    // st(0) there, and loading it into the xmm register.
     typename InstX86Base<Machine>::Traits::Address StackSlot =
         typename InstX86Base<Machine>::Traits::Address(
             InstX86Base<Machine>::Traits::RegisterSet::Encoded_Reg_esp, 0,
             AssemblerFixup::NoFixup);
     Asm->fstp(Ty, StackSlot);
     Asm->movss(Ty,
                InstX86Base<Machine>::Traits::getEncodedXmm(Dest->getRegNum()),
                StackSlot);
-    Asm->add(IceType_i32,
-             InstX86Base<Machine>::Traits::RegisterSet::Encoded_Reg_esp, Width);
   }
 }
 
 template <class Machine>
 void InstX86Fstp<Machine>::dump(const Cfg *Func) const {
   if (!BuildDefs::dump())
     return;
   Ostream &Str = Func->getContext()->getStrDump();
   this->dumpDest(Func);
   Str << " = fstp." << this->getDest()->getType() << ", st(0)";
@@ skipping 221 matching lines @@
   Ostream &Str = Func->getContext()->getStrDump();
   this->dumpDest(Func);
   Str << " = pop." << this->getDest()->getType() << " ";
 }
 
 template <class Machine>
 void InstX86AdjustStack<Machine>::emit(const Cfg *Func) const {
   if (!BuildDefs::dump())
     return;
   Ostream &Str = Func->getContext()->getStrEmit();
-  Str << "\tsubl\t$" << Amount << ", %esp";
+  if (Amount > 0)
+    Str << "\tsubl\t$" << Amount << ", %esp";
+  else
+    Str << "\taddl\t$" << -Amount << ", %esp";
   auto *Target = InstX86Base<Machine>::getTarget(Func);
   Target->updateStackAdjustment(Amount);
 }
 
 template <class Machine>
 void InstX86AdjustStack<Machine>::emitIAS(const Cfg *Func) const {
   typename InstX86Base<Machine>::Traits::Assembler *Asm =
       Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
-  Asm->sub(IceType_i32,
-           InstX86Base<Machine>::Traits::RegisterSet::Encoded_Reg_esp,
-           Immediate(Amount));
+  if (Amount > 0)
+    Asm->sub(IceType_i32,
+             InstX86Base<Machine>::Traits::RegisterSet::Encoded_Reg_esp,
+             Immediate(Amount));
+  else
+    Asm->add(IceType_i32,
+             InstX86Base<Machine>::Traits::RegisterSet::Encoded_Reg_esp,
+             Immediate(-Amount));
   auto *Target = InstX86Base<Machine>::getTarget(Func);
   Target->updateStackAdjustment(Amount);
 }
 
 template <class Machine>
 void InstX86AdjustStack<Machine>::dump(const Cfg *Func) const {
   if (!BuildDefs::dump())
     return;
   Ostream &Str = Func->getContext()->getStrDump();
-  Str << "esp = sub.i32 esp, " << Amount;
+  if (Amount > 0)
+    Str << "esp = sub.i32 esp, " << Amount;
+  else
+    Str << "esp = add.i32 esp, " << -Amount;
 }
 
 template <class Machine>
 void InstX86Push<Machine>::emit(const Cfg *Func) const {
   if (!BuildDefs::dump())
     return;
   Ostream &Str = Func->getContext()->getStrEmit();
   assert(this->getSrcSize() == 1);
   // Push is currently only used for saving GPRs.
   const auto *Var = llvm::cast<Variable>(this->getSrc(0));
@@ skipping 277 matching lines @@
     return;
   Ostream &Str = Func->getContext()->getStrDump();
   Str << "IACA_END";
 }
 
 } // end of namespace X86Internal
 
 } // end of namespace Ice
 
 #endif // SUBZERO_SRC_ICEINSTX86BASEIMPL_H