Unify alloca, outgoing arg, and prolog construction

src/IceTargetLoweringX8664.cpp

 //===- subzero/src/IceTargetLoweringX8664.cpp - x86-64 lowering -----------===//
 //
 //                        The Subzero Code Generator
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 ///
 /// \file
@@ skipping 152 matching lines @@
   // This intends to match the section "Function Calling Sequence" of the
   // document "System V Application Binary Interface."
   NeedsStackAlignment = true;
 
   using OperandList =
       llvm::SmallVector<Operand *, constexprMax(Traits::X86_MAX_XMM_ARGS,
                                                 Traits::X86_MAX_GPR_ARGS)>;
   OperandList XmmArgs;
   OperandList GprArgs;
   OperandList StackArgs, StackArgLocations;
-  uint32_t ParameterAreaSizeBytes = 0;
+  int32_t ParameterAreaSizeBytes = 0;
 
   // Classify each argument operand according to the location where the
   // argument is passed.
   for (SizeT i = 0, NumArgs = Instr->getNumArgs(); i < NumArgs; ++i) {
     Operand *Arg = Instr->getArg(i);
     Type Ty = Arg->getType();
     // The PNaCl ABI requires the width of arguments to be at least 32 bits.
     assert(typeWidthInBytes(Ty) >= 4);
     if (isVectorType(Ty) && XmmArgs.size() < Traits::X86_MAX_XMM_ARGS) {
       XmmArgs.push_back(Arg);
@@ skipping 13 matching lines @@
       Constant *Loc = Ctx->getConstantInt32(ParameterAreaSizeBytes);
       StackArgLocations.push_back(
           Traits::X86OperandMem::create(Func, Ty, esp, Loc));
       ParameterAreaSizeBytes += typeWidthInBytesOnStack(Arg->getType());
     }
   }
 
   // Adjust the parameter area so that the stack is aligned. It is assumed that
   // the stack is already aligned at the start of the calling sequence.
   ParameterAreaSizeBytes = Traits::applyStackAlignment(ParameterAreaSizeBytes);
-
-  // Subtract the appropriate amount for the argument area. This also takes
-  // care of setting the stack adjustment during emission.
-  //
-  // TODO: If for some reason the call instruction gets dead-code eliminated
-  // after lowering, we would need to ensure that the pre-call and the
-  // post-call esp adjustment get eliminated as well.
-  if (ParameterAreaSizeBytes) {
-    _adjust_stack(ParameterAreaSizeBytes);
-  }
+  assert(static_cast<uint32_t>(ParameterAreaSizeBytes) <=
+         maxOutArgsSizeBytes());
 
   // Copy arguments that are passed on the stack to the appropriate stack
   // locations.
   for (SizeT i = 0, e = StackArgs.size(); i < e; ++i) {
     lowerStore(InstStore::create(Func, StackArgs[i], StackArgLocations[i]));
   }
 
   // Copy arguments to be passed in registers to the appropriate registers.
   // TODO: Investigate the impact of lowering arguments passed in registers
   // after lowering stack arguments as opposed to the other way around.
@@ skipping 175 matching lines @@
   // | 4. global spill area   |
   // +------------------------+
   // | 5. padding             |
   // +------------------------+
   // | 6. local spill area    |
   // +------------------------+
   // | 7. padding             |
   // +------------------------+
   // | 8. allocas             |
   // +------------------------+
+  // | 9. padding             |
+  // +------------------------+
+  // | 10. out args           |
+  // +------------------------+ <--- StackPointer
   //
   // The following variables record the size in bytes of the given areas:
   //  * X86_RET_IP_SIZE_BYTES:  area 1
   //  * PreservedRegsSizeBytes: area 2
   //  * SpillAreaPaddingBytes:  area 3
   //  * GlobalsSize:            area 4
   //  * GlobalsAndSubsequentPaddingSize: areas 4 - 5
   //  * LocalsSpillAreaSize:    area 6
-  //  * SpillAreaSizeBytes:     areas 3 - 7
+  //  * SpillAreaSizeBytes:     areas 3 - 10
+  //  * maxOutArgsSizeBytes():  area 10
 
   // Determine stack frame offsets for each Variable without a register
   // assignment. This can be done as one variable per stack slot. Or, do
   // coalescing by running the register allocator again with an infinite set of
   // registers (as a side effect, this gives variables a second chance at
   // physical register assignment).
   //
   // A middle ground approach is to leverage sparsity and allocate one block of
   // space on the frame for globals (variables with multi-block lifetime), and
   // one block to share for locals (single-block lifetime).
@@ skipping 71 matching lines @@
   uint32_t LocalsSlotsPaddingBytes = 0;
   alignStackSpillAreas(Traits::X86_RET_IP_SIZE_BYTES + PreservedRegsSizeBytes,
                        SpillAreaAlignmentBytes, GlobalsSize,
                        LocalsSlotsAlignmentBytes, &SpillAreaPaddingBytes,
                        &LocalsSlotsPaddingBytes);
   SpillAreaSizeBytes += SpillAreaPaddingBytes + LocalsSlotsPaddingBytes;
   uint32_t GlobalsAndSubsequentPaddingSize =
       GlobalsSize + LocalsSlotsPaddingBytes;
 
   // Align esp if necessary.
-  if (NeedsStackAlignment) {
+  if (!NeedsStackAlignment) {
+    SpillAreaSizeBytes += maxOutArgsSizeBytes();
+  } else {
     uint32_t StackOffset =
         Traits::X86_RET_IP_SIZE_BYTES + PreservedRegsSizeBytes;
     uint32_t StackSize =
         Traits::applyStackAlignment(StackOffset + SpillAreaSizeBytes);
+    StackSize = Traits::applyStackAlignment(StackSize + maxOutArgsSizeBytes());
     SpillAreaSizeBytes = StackSize - StackOffset;
   }
 
+  // Combine fixed allocations into SpillAreaSizeBytes if we are emitting the
+  // fixed allocations in the prolog.
+  if (PrologEmitsFixedAllocas)
+    SpillAreaSizeBytes += FixedAllocaSizeBytes;
   // Generate "sub esp, SpillAreaSizeBytes"
-  if (SpillAreaSizeBytes)
+  if (SpillAreaSizeBytes) {
     _sub(getPhysicalRegister(Traits::RegisterSet::Reg_esp),
          Ctx->getConstantInt32(SpillAreaSizeBytes));
+    // If the fixed allocas are aligned more than the stack frame, align the
+    // stack pointer accordingly.
+    if (PrologEmitsFixedAllocas &&
+        FixedAllocaAlignBytes > Traits::X86_STACK_ALIGNMENT_BYTES) {
+      assert(IsEbpBasedFrame);
+      _and(getPhysicalRegister(Traits::RegisterSet::Reg_esp),
+           Ctx->getConstantInt32(-FixedAllocaAlignBytes));
+    }
+  }
 
   // Account for alloca instructions with known frame offsets.
-  SpillAreaSizeBytes += FixedAllocaSizeBytes;
+  if (!PrologEmitsFixedAllocas)
+    SpillAreaSizeBytes += FixedAllocaSizeBytes;
 
   Ctx->statsUpdateFrameBytes(SpillAreaSizeBytes);
 
-  // Initialize the stack adjustment so that after all the known-frame-offset
-  // alloca instructions are emitted, the stack adjustment will reach zero.
-  resetStackAdjustment();
-  updateStackAdjustment(-FixedAllocaSizeBytes);
-
   // Fill in stack offsets for stack args, and copy args into registers for
   // those that were register-allocated. Args are pushed right to left, so
   // Arg[0] is closest to the stack/frame pointer.
   Variable *FramePtr = getPhysicalRegister(getFrameOrStackReg());
   size_t BasicFrameOffset =
       PreservedRegsSizeBytes + Traits::X86_RET_IP_SIZE_BYTES;
   if (!IsEbpBasedFrame)
     BasicFrameOffset += SpillAreaSizeBytes;
 
   const VarList &Args = Func->getArgs();
@@ skipping 11 matching lines @@
       assert(isScalarIntegerType(Arg->getType()));
       if (NumGPRArgs < Traits::X86_MAX_GPR_ARGS) {
         ++NumGPRArgs;
         continue;
       }
     }
     // For esp-based frames, the esp value may not stabilize to its home value
     // until after all the fixed-size alloca instructions have executed.  In
     // this case, a stack adjustment is needed when accessing in-args in order
     // to copy them into registers.
-    size_t StackAdjBytes = IsEbpBasedFrame ? 0 : -FixedAllocaSizeBytes;
+    size_t StackAdjBytes = 0;
+    if (!IsEbpBasedFrame && !PrologEmitsFixedAllocas)
+      StackAdjBytes -= FixedAllocaSizeBytes;
     finishArgumentLowering(Arg, FramePtr, BasicFrameOffset, StackAdjBytes,
                            InArgsSizeBytes);
   }
 
   // Fill in stack offsets for locals.
   assignVarStackSlots(SortedSpilledVariables, SpillAreaPaddingBytes,
                       SpillAreaSizeBytes, GlobalsAndSubsequentPaddingSize,
                       IsEbpBasedFrame);
   // Assign stack offsets to variables that have been linked to spilled
   // variables.
   for (Variable *Var : VariablesLinkedToSpillSlots) {
     Variable *Linked =
         (llvm::cast<typename Traits::SpillVariable>(Var))->getLinkedTo();
     Var->setStackOffset(Linked->getStackOffset());
   }
   this->HasComputedFrame = true;
 
   if (BuildDefs::dump() && Func->isVerbose(IceV_Frame)) {
     OstreamLocker L(Func->getContext());
     Ostream &Str = Func->getContext()->getStrDump();
 
     Str << "Stack layout:\n";
     uint32_t EspAdjustmentPaddingSize =
         SpillAreaSizeBytes - LocalsSpillAreaSize -
-        GlobalsAndSubsequentPaddingSize - SpillAreaPaddingBytes;
+        GlobalsAndSubsequentPaddingSize - SpillAreaPaddingBytes -
+        maxOutArgsSizeBytes();
     Str << " in-args = " << InArgsSizeBytes << " bytes\n"
         << " return address = " << Traits::X86_RET_IP_SIZE_BYTES << " bytes\n"
         << " preserved registers = " << PreservedRegsSizeBytes << " bytes\n"
         << " spill area padding = " << SpillAreaPaddingBytes << " bytes\n"
         << " globals spill area = " << GlobalsSize << " bytes\n"
         << " globals-locals spill areas intermediate padding = "
         << GlobalsAndSubsequentPaddingSize - GlobalsSize << " bytes\n"
         << " locals spill area = " << LocalsSpillAreaSize << " bytes\n"
         << " esp alignment padding = " << EspAdjustmentPaddingSize
         << " bytes\n";
 
     Str << "Stack details:\n"
         << " esp adjustment = " << SpillAreaSizeBytes << " bytes\n"
         << " spill area alignment = " << SpillAreaAlignmentBytes << " bytes\n"
+        << " outgoing args size = " << maxOutArgsSizeBytes() << " bytes\n"
         << " locals spill area alignment = " << LocalsSlotsAlignmentBytes
         << " bytes\n"
         << " is ebp based = " << IsEbpBasedFrame << "\n";
   }
 }
 
 void TargetX8664::addEpilog(CfgNode *Node) {
   InstList &Insts = Node->getInsts();
   InstList::reverse_iterator RI, E;
   for (RI = Insts.rbegin(), E = Insts.rend(); RI != E; ++RI) {
@@ skipping 350 matching lines @@
 // case the high-level table has extra entries.
 #define X(tag, sizeLog2, align, elts, elty, str)                               \
   static_assert(_table1_##tag == _table2_##tag,                                \
                 "Inconsistency between ICETYPEX8664_TABLE and ICETYPE_TABLE");
 ICETYPE_TABLE
 #undef X
 } // end of namespace dummy3
 } // end of anonymous namespace
 
 } // end of namespace Ice