Propagate the AlignStack bit in InlineAsm's to the

PrologEpilog code, and use it to determine whether
the asm forces stack alignment or not.  gcc consistently
does not do this for GCC-style asms; Apple gcc inconsistently
sometimes does it for asm blocks.  There is no
convenient place to put a bit in either the SDNode or
the MachineInstr form, so I've added an extra operand
to each; unlovely, but it does allow for expansion for
more bits, should we need it.  PR 5125.  Some
existing testcases are affected.
The operand lists of the SDNode and MachineInstr forms
are indexed with awesome mnemonics, like "2"; I may
fix this someday, but not now.  I'm not making it any
worse.  If anyone is inspired I think you can find all
the right places from this patch.

llvm-svn: 107506

include/llvm/InlineAsm.h

@@ -154,7 +154,8 @@ public:
     Op_InputChain = 0,
     Op_AsmString = 1,
     Op_MDNode = 2,
-    Op_FirstOperand = 3,
+    Op_IsAlignStack = 3,
+    Op_FirstOperand = 4,
     
     Kind_RegUse = 1,
     Kind_RegDef = 2,

lib/CodeGen/AsmPrinter/AsmPrinterInlineAsm.cpp

@@ -279,7 +279,7 @@ void AsmPrinter::EmitInlineAsm(const MachineInstr *MI) const {
       // Okay, we finally have a value number.  Ask the target to print this
       // operand!
       if (CurVariant == -1 || CurVariant == AsmPrinterVariant) {
-        unsigned OpNo = 1;
+        unsigned OpNo = 2;
 
         bool Error = false;
 

lib/CodeGen/MachineInstr.cpp

@@ -881,14 +881,14 @@ int MachineInstr::findFirstPredOperandIdx() const {
 bool MachineInstr::
 isRegTiedToUseOperand(unsigned DefOpIdx, unsigned *UseOpIdx) const {
   if (isInlineAsm()) {
-    assert(DefOpIdx >= 2);
+    assert(DefOpIdx >= 3);
     const MachineOperand &MO = getOperand(DefOpIdx);
     if (!MO.isReg() || !MO.isDef() || MO.getReg() == 0)
       return false;
     // Determine the actual operand index that corresponds to this index.
     unsigned DefNo = 0;
     unsigned DefPart = 0;
-    for (unsigned i = 1, e = getNumOperands(); i < e; ) {
+    for (unsigned i = 2, e = getNumOperands(); i < e; ) {
       const MachineOperand &FMO = getOperand(i);
       // After the normal asm operands there may be additional imp-def regs.
       if (!FMO.isImm())
@@ -903,7 +903,7 @@ isRegTiedToUseOperand(unsigned DefOpIdx, unsigned *UseOpIdx) const {
       }
       ++DefNo;
     }
-    for (unsigned i = 1, e = getNumOperands(); i != e; ++i) {
+    for (unsigned i = 2, e = getNumOperands(); i != e; ++i) {
       const MachineOperand &FMO = getOperand(i);
       if (!FMO.isImm())
         continue;
@@ -946,7 +946,7 @@ isRegTiedToDefOperand(unsigned UseOpIdx, unsigned *DefOpIdx) const {
 
     // Find the flag operand corresponding to UseOpIdx
     unsigned FlagIdx, NumOps=0;
-    for (FlagIdx = 1; FlagIdx < UseOpIdx; FlagIdx += NumOps+1) {
+    for (FlagIdx = 2; FlagIdx < UseOpIdx; FlagIdx += NumOps+1) {
       const MachineOperand &UFMO = getOperand(FlagIdx);
       // After the normal asm operands there may be additional imp-def regs.
       if (!UFMO.isImm())
@@ -964,9 +964,9 @@ isRegTiedToDefOperand(unsigned UseOpIdx, unsigned *DefOpIdx) const {
       if (!DefOpIdx)
         return true;
 
-      unsigned DefIdx = 1;
+      unsigned DefIdx = 2;
-      // Remember to adjust the index. First operand is asm string, then there
+      // Remember to adjust the index. First operand is asm string, second is
-      // is a flag for each.
+      // the AlignStack bit, then there is a flag for each.
       while (DefNo) {
         const MachineOperand &FMO = getOperand(DefIdx);
         assert(FMO.isImm());

lib/CodeGen/PrologEpilogInserter.cpp

@@ -158,9 +158,9 @@ void PEI::calculateCallsInformation(MachineFunction &Fn) {
         AdjustsStack = true;
         FrameSDOps.push_back(I);
       } else if (I->isInlineAsm()) {
-        // An InlineAsm might be a call; assume it is to get the stack frame
+        // Some inline asm's need a stack frame, as indicated by operand 1.
-        // aligned correctly for calls.
+        if (I->getOperand(1).getImm())
-        AdjustsStack = true;
+          AdjustsStack = true;
       }
 
   MFI->setAdjustsStack(AdjustsStack);

lib/CodeGen/SelectionDAG/InstrEmitter.cpp

@@ -834,6 +834,12 @@ EmitSpecialNode(SDNode *Node, bool IsClone, bool IsCloned,
     const char *AsmStr = cast<ExternalSymbolSDNode>(AsmStrV)->getSymbol();
     MI->addOperand(MachineOperand::CreateES(AsmStr));
       
+    // Add the isAlignStack bit.
+    int64_t isAlignStack =
+      cast<ConstantSDNode>(Node->getOperand(InlineAsm::Op_IsAlignStack))->
+                          getZExtValue();
+    MI->addOperand(MachineOperand::CreateImm(isAlignStack));
+
     // Add all of the operand registers to the instruction.
     for (unsigned i = InlineAsm::Op_FirstOperand; i != NumOps;) {
       unsigned Flags =

lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

@@ -5454,6 +5454,10 @@ void SelectionDAGBuilder::visitInlineAsm(ImmutableCallSite CS) {
   const MDNode *SrcLoc = CS.getInstruction()->getMetadata("srcloc");
   AsmNodeOperands.push_back(DAG.getMDNode(SrcLoc));
 
+  // Remember the AlignStack bit as operand 3.
+  AsmNodeOperands.push_back(DAG.getTargetConstant(IA->isAlignStack() ? 1 : 0,
+                                            MVT::i1));
+
   // Loop over all of the inputs, copying the operand values into the
   // appropriate registers and processing the output regs.
   RegsForValue RetValRegs;
@@ -5642,7 +5646,7 @@ void SelectionDAGBuilder::visitInlineAsm(ImmutableCallSite CS) {
   }
 
   // Finish up input operands.  Set the input chain and add the flag last.
-  AsmNodeOperands[0] = Chain;
+  AsmNodeOperands[InlineAsm::Op_InputChain] = Chain;
   if (Flag.getNode()) AsmNodeOperands.push_back(Flag);
 
   Chain = DAG.getNode(ISD::INLINEASM, getCurDebugLoc(),

lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp

@@ -1089,6 +1089,7 @@ SelectInlineAsmMemoryOperands(std::vector<SDValue> &Ops) {
   Ops.push_back(InOps[InlineAsm::Op_InputChain]); // 0
   Ops.push_back(InOps[InlineAsm::Op_AsmString]);  // 1
   Ops.push_back(InOps[InlineAsm::Op_MDNode]);     // 2, !srcloc
+  Ops.push_back(InOps[InlineAsm::Op_IsAlignStack]);  // 3
 
   unsigned i = InlineAsm::Op_FirstOperand, e = InOps.size();
   if (InOps[e-1].getValueType() == MVT::Flag)

test/CodeGen/Thumb/push.ll

@@ -5,6 +5,6 @@ define void @t() nounwind {
 ; CHECK: t:
 ; CHECK: push {r7}
 entry:
-  call void asm sideeffect ".long 0xe7ffdefe", ""() nounwind
+  call void asm sideeffect alignstack ".long 0xe7ffdefe", ""() nounwind
   ret void
 }

test/CodeGen/X86/2009-04-20-LinearScanOpt.ll

@@ -1,4 +1,4 @@
-; RUN: llc < %s -mtriple=x86_64-apple-darwin10.0 -relocation-model=pic -disable-fp-elim -stats |& grep asm-printer | grep 83
+; RUN: llc < %s -mtriple=x86_64-apple-darwin10.0 -relocation-model=pic -disable-fp-elim -stats |& grep asm-printer | grep 82
 ; rdar://6802189
 
 ; Test if linearscan is unfavoring registers for allocation to allow more reuse

test/CodeGen/X86/2010-07-02-asm-alignstack.ll

@@ -0,0 +1,31 @@
+; RUN: llc < %s -mtriple=x86_64-apple-darwin10 | FileCheck %s
+
+define void @foo() nounwind ssp {
+entry:
+; CHECK: foo
+; CHECK: subq $8, %rsp
+; CHECK: int $3
+  call void asm sideeffect alignstack "# top of block", "~{dirflag},~{fpsr},~{flags},~{edi},~{esi},~{edx},~{ecx},~{eax}"() nounwind
+  call void asm sideeffect alignstack ".file \22small.c\22", "~{dirflag},~{fpsr},~{flags}"() nounwind
+  call void asm sideeffect alignstack ".line 3", "~{dirflag},~{fpsr},~{flags}"() nounwind
+  call void asm sideeffect alignstack "int $$3", "~{dirflag},~{fpsr},~{flags},~{memory}"() nounwind
+  br label %return
+
+return:                                           ; preds = %entry
+  ret void
+}
+
+define void @bar() nounwind ssp {
+entry:
+; CHECK: bar
+; CHECK-NOT: subq $8, %rsp
+; CHECK: int $3
+  call void asm sideeffect "# top of block", "~{dirflag},~{fpsr},~{flags},~{edi},~{esi},~{edx},~{ecx},~{eax}"() nounwind
+  call void asm sideeffect ".file \22small.c\22", "~{dirflag},~{fpsr},~{flags}"() nounwind
+  call void asm sideeffect ".line 3", "~{dirflag},~{fpsr},~{flags}"() nounwind
+  call void asm sideeffect "int $$3", "~{dirflag},~{fpsr},~{flags},~{memory}"() nounwind
+  br label %return
+
+return:                                           ; preds = %entry
+  ret void
+}

test/CodeGen/X86/inline-asm-tied.ll

@@ -1,4 +1,4 @@
-; RUN: llc < %s -mtriple=i386-apple-darwin9 -O0 -regalloc=linearscan | grep {movl	%edx, 12(%esp)} | count 2
+; RUN: llc < %s -mtriple=i386-apple-darwin9 -O0 -regalloc=linearscan | grep {movl	%edx, 4(%esp)} | count 2
 ; rdar://6992609
 
 target triple = "i386-apple-darwin9.0"

test/CodeGen/X86/iv-users-in-other-loops.ll

@@ -1,7 +1,7 @@
 ; RUN: llc < %s -march=x86-64 -o %t
 ; RUN: not grep inc %t
 ; RUN: grep dec %t | count 2
-; RUN: grep addq %t | count 13
+; RUN: grep addq %t | count 12
 ; RUN: not grep addb %t
 ; RUN: not grep leaq %t
 ; RUN: not grep leal %t