switch to having the callee pop stack operands for fastcc. This is currently buggy

do not use


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21984 91177308-0d34-0410-b5e6-96231b3b80d8

3 files changed, 43 insertions, 16 deletions

diff --git a/lib/Target/X86/X86ISelPattern.cpp b/lib/Target/X86/X86ISelPattern.cpp
index eb6babf628..3345afcc48 100644
--- a/lib/Target/X86/X86ISelPattern.cpp
+++ b/lib/Target/X86/X86ISelPattern.cpp
@@ -109,6 +109,11 @@ namespace {
       addLegalFPImmediate(-1.0); // FLD1/FCHS
     }
 
+    // Return the number of bytes that a function should pop when it returns (in
+    // addition to the space used by the return address).
+    //
+    unsigned getBytesToPopOnReturn() const { return BytesToPopOnReturn; }
+
     /// LowerArguments - This hook must be implemented to indicate how we should
     /// lower the arguments for the specified function, into the specified DAG.
     virtual std::vector<SDOperand>
@@ -228,7 +233,8 @@ X86TargetLowering::LowerCCCArguments(Function &F, SelectionDAG &DAG) {
   // the start of the first vararg value... for expansion of llvm.va_start.
   if (F.isVarArg())
     VarArgsFrameIndex = MFI->CreateFixedObject(1, ArgOffset);
-  ReturnAddrIndex = 0;  // No return address slot generated yet.
+  ReturnAddrIndex = 0;     // No return address slot generated yet.
+  BytesToPopOnReturn = 0;  // Callee pops nothing.
 
   // Finally, inform the code generator which regs we return values in.
   switch (getValueType(F.getReturnType())) {
@@ -527,11 +533,9 @@ X86TargetLowering::LowerFastCCArguments(Function &F, SelectionDAG &DAG) {
       ArgOffset += ArgIncrement;   // Move on to the next argument.
   }
 
-  // If the function takes variable number of arguments, make a frame index for
-  // the start of the first vararg value... for expansion of llvm.va_start.
-  if (F.isVarArg())
-    VarArgsFrameIndex = MFI->CreateFixedObject(1, ArgOffset);
-  ReturnAddrIndex = 0;  // No return address slot generated yet.
+  VarArgsFrameIndex = 0xAAAAAAA;   // fastcc functions can't have varargs.
+  ReturnAddrIndex = 0;             // No return address slot generated yet.
+  BytesToPopOnReturn = ArgOffset;  // Callee pops all stack arguments.
 
   // Finally, inform the code generator which regs we return values in.
   switch (getValueType(F.getReturnType())) {
@@ -3518,7 +3522,10 @@ void ISel::Select(SDOperand N) {
       Select(N.getOperand(0));
       break;
     }
-    BuildMI(BB, X86::RET, 0); // Just emit a 'ret' instruction
+    if (X86Lowering.getBytesToPopOnReturn() == 0)
+      BuildMI(BB, X86::RET, 0); // Just emit a 'ret' instruction
+    else
+      BuildMI(BB, X86::RETI, 1).addImm(X86Lowering.getBytesToPopOnReturn());
     return;
   case ISD::BR: {
     Select(N.getOperand(0));
@@ -3721,13 +3728,19 @@ void ISel::Select(SDOperand N) {
     return;
   }
   case ISD::CALLSEQ_START:
+    Select(N.getOperand(0));
+    // Stack amount
+    Tmp1 = cast<ConstantSDNode>(N.getOperand(1))->getValue();
+    BuildMI(BB, X86::ADJCALLSTACKDOWN, 1).addImm(Tmp1);
+    return;
   case ISD::CALLSEQ_END:
     Select(N.getOperand(0));
+    // Stack amount
     Tmp1 = cast<ConstantSDNode>(N.getOperand(1))->getValue();
 
-    Opc = N.getOpcode() == ISD::CALLSEQ_START ? X86::ADJCALLSTACKDOWN :
-                                                X86::ADJCALLSTACKUP;
-    BuildMI(BB, Opc, 1).addImm(Tmp1);
+    // Amount the callee added to the stack pointer.
+    Tmp2 = cast<ConstantSDNode>(N.getOperand(2))->getValue();
+    BuildMI(BB, X86::ADJCALLSTACKUP, 2).addImm(Tmp1).addImm(Tmp2);
     return;
   case ISD::MEMSET: {
     Select(N.getOperand(0));  // Select the chain.
diff --git a/lib/Target/X86/X86ISelSimple.cpp b/lib/Target/X86/X86ISelSimple.cpp
index f27784b7f7..ea5520fbf8 100644
--- a/lib/Target/X86/X86ISelSimple.cpp
+++ b/lib/Target/X86/X86ISelSimple.cpp
@@ -1695,7 +1695,7 @@ void X86ISel::doCall(const ValueRecord &Ret, MachineInstr *CallMI,
 
   BB->push_back(CallMI);
 
-  BuildMI(BB, X86::ADJCALLSTACKUP, 1).addImm(NumBytes);
+  BuildMI(BB, X86::ADJCALLSTACKUP, 2).addImm(NumBytes).addImm(0);
 
   // If there is a return value, scavenge the result from the location the call
   // leaves it in...
diff --git a/lib/Target/X86/X86RegisterInfo.cpp b/lib/Target/X86/X86RegisterInfo.cpp
index 5ef4862d39..4fb5ccf5b5 100644
--- a/lib/Target/X86/X86RegisterInfo.cpp
+++ b/lib/Target/X86/X86RegisterInfo.cpp
@@ -375,17 +375,31 @@ eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB,
       unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
       Amount = (Amount+Align-1)/Align*Align;
 
-      MachineInstr *New;
+      MachineInstr *New = 0;
       if (Old->getOpcode() == X86::ADJCALLSTACKDOWN) {
 	New=BuildMI(X86::SUB32ri, 1, X86::ESP, MachineOperand::UseAndDef)
               .addZImm(Amount);
       } else {
 	assert(Old->getOpcode() == X86::ADJCALLSTACKUP);
-	New=BuildMI(X86::ADD32ri, 1, X86::ESP, MachineOperand::UseAndDef)
-              .addZImm(Amount);
+        // factor out the amount the callee already popped.
+        unsigned CalleeAmt = Old->getOperand(1).getImmedValue();
+        Amount -= CalleeAmt;
+        if (Amount)
+          New = BuildMI(X86::ADD32ri, 1, X86::ESP,
+                        MachineOperand::UseAndDef).addZImm(Amount);
       }
 
       // Replace the pseudo instruction with a new instruction...
+      if (New) MBB.insert(I, New);
+    }
+  } else if (I->getOpcode() == X86::ADJCALLSTACKUP) {
+    // If we are performing frame pointer elimination and if the callee pops
+    // something off the stack pointer, add it back.  We do this until we have
+    // more advanced stack pointer tracking ability.
+    if (unsigned CalleeAmt = I->getOperand(1).getImmedValue()) {
+      MachineInstr *New = 
+        BuildMI(X86::SUB32ri, 1, X86::ESP,
+                MachineOperand::UseAndDef).addZImm(CalleeAmt);
       MBB.insert(I, New);
     }
   }
@@ -513,8 +527,8 @@ void X86RegisterInfo::emitEpilogue(MachineFunction &MF,
     unsigned NumBytes = MFI->getStackSize();
 
     if (NumBytes) {    // adjust stack pointer back: ESP += numbytes
-      MI =BuildMI(X86::ADD32ri, 1, X86::ESP, MachineOperand::UseAndDef)
-            .addZImm(NumBytes);
+      MI = BuildMI(X86::ADD32ri, 1, X86::ESP, MachineOperand::UseAndDef)
+        .addZImm(NumBytes);
       MBB.insert(MBBI, MI);
     }
   }