[Sparc] Support user-specified stack object overalignment.

Note: I do not implement a base pointer, so it's still impossible to
have dynamic realignment AND dynamic alloca in the same function.

This also moves the code for determining the frame index reference
into getFrameIndexReference, where it belongs, instead of inline in
eliminateFrameIndex.

[Begin long-winded screed]

Now, stack realignment for Sparc is actually a silly thing to support,
because the Sparc ABI has no need for it -- unlike the situation on
x86, the stack is ALWAYS aligned to the required alignment for the CPU
instructions: 8 bytes on sparcv8, and 16 bytes on sparcv9.

However, LLVM unfortunately implements user-specified overalignment
using stack realignment support, so for now, I'm going to go along
with that tradition. GCC instead treats objects which have alignment
specification greater than the maximum CPU-required alignment for the
target as a separate block of stack memory, with their own virtual
base pointer (which gets aligned). Doing it that way avoids needing to
implement per-target support for stack realignment, except for the
targets which *actually* have an ABI-specified stack alignment which
is too small for the CPU's requirements.

Further unfortunately in LLVM, the default canRealignStack for all
targets effectively returns true, despite that implementing that is
something a target needs to do specifically. So, the previous behavior
on Sparc was to silently ignore the user's specified stack
alignment. Ugh.

Yet MORE unfortunate, if a target actually does return false from
canRealignStack, that also causes the user-specified alignment to be
*silently ignored*, rather than emitting an error.

(I started looking into fixing that last, but it broke a bunch of
tests, because LLVM actually *depends* on having it silently ignored:
some architectures (e.g. non-linux i386) have smaller stack alignment
than spilled-register alignment. But, the fact that a register needs
spilling is not known until within the register allocator. And by that
point, the decision to not reserve the frame pointer has been frozen
in place. And without a frame pointer, stack realignment is not
possible. So, canRealignStack() returns false, and
needsStackRealignment() then returns false, assuming everyone can just
go on their merry way assuming the alignment requirements were
probably just suggestions after-all. Sigh...)

Differential Revision: http://reviews.llvm.org/D12208

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@245668 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/Sparc/SparcFrameLowering.cpp

@@ -90,8 +90,21 @@ void SparcFrameLowering::emitPrologue(MachineFunction &MF,
   MachineFrameInfo *MFI = MF.getFrameInfo();
   const SparcInstrInfo &TII =
       *static_cast<const SparcInstrInfo *>(MF.getSubtarget().getInstrInfo());
+  const SparcRegisterInfo &RegInfo =
+      *static_cast<const SparcRegisterInfo *>(MF.getSubtarget().getRegisterInfo());
   MachineBasicBlock::iterator MBBI = MBB.begin();
   DebugLoc dl = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc();
+  bool NeedsStackRealignment = RegInfo.needsStackRealignment(MF);
+
+  // FIXME: unfortunately, returning false from canRealignStack
+  // actually just causes needsStackRealignment to return false,
+  // rather than reporting an error, as would be sensible. This is
+  // poor, but fixing that bogosity is going to be a large project.
+  // For now, just see if it's lied, and report an error here.
+  if (!NeedsStackRealignment && MFI->getMaxAlignment() > getStackAlignment())
+    report_fatal_error("Function \"" + Twine(MF.getName()) + "\" required "
+                       "stack re-alignment, but LLVM couldn't handle it "
+                       "(probably because it has a dynamic alloca).");
 
   // Get the number of bytes to allocate from the FrameInfo
   int NumBytes = (int) MFI->getStackSize();
@@ -104,12 +117,14 @@ void SparcFrameLowering::emitPrologue(MachineFunction &MF,
     SAVEri = SP::ADDri;
     SAVErr = SP::ADDrr;
   }
-  NumBytes = -MF.getSubtarget<SparcSubtarget>().getAdjustedFrameSize(NumBytes);
+
-  emitSPAdjustment(MF, MBB, MBBI, NumBytes, SAVErr, SAVEri);
+  NumBytes = MF.getSubtarget<SparcSubtarget>().getAdjustedFrameSize(NumBytes);
+  MFI->setStackSize(NumBytes); // Update stack size with corrected value.
+
+  emitSPAdjustment(MF, MBB, MBBI, -NumBytes, SAVErr, SAVEri);
 
   MachineModuleInfo &MMI = MF.getMMI();
-  const MCRegisterInfo *MRI = MMI.getContext().getRegisterInfo();
+  unsigned regFP = RegInfo.getDwarfRegNum(SP::I6, true);
-  unsigned regFP = MRI->getDwarfRegNum(SP::I6, true);
 
   // Emit ".cfi_def_cfa_register 30".
   unsigned CFIIndex =
@@ -122,13 +137,19 @@ void SparcFrameLowering::emitPrologue(MachineFunction &MF,
   BuildMI(MBB, MBBI, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
       .addCFIIndex(CFIIndex);
 
-  unsigned regInRA = MRI->getDwarfRegNum(SP::I7, true);
+  unsigned regInRA = RegInfo.getDwarfRegNum(SP::I7, true);
-  unsigned regOutRA = MRI->getDwarfRegNum(SP::O7, true);
+  unsigned regOutRA = RegInfo.getDwarfRegNum(SP::O7, true);
   // Emit ".cfi_register 15, 31".
   CFIIndex = MMI.addFrameInst(
       MCCFIInstruction::createRegister(nullptr, regOutRA, regInRA));
   BuildMI(MBB, MBBI, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
       .addCFIIndex(CFIIndex);
+
+  if (NeedsStackRealignment) {
+    // andn %o6, MaxAlign-1, %o6
+    int MaxAlign = MFI->getMaxAlignment();
+    BuildMI(MBB, MBBI, dl, TII.get(SP::ANDNri), SP::O6).addReg(SP::O6).addImm(MaxAlign - 1);
+  }
 }
 
 void SparcFrameLowering::
@@ -167,7 +188,6 @@ void SparcFrameLowering::emitEpilogue(MachineFunction &MF,
   if (NumBytes == 0)
     return;
 
-  NumBytes = MF.getSubtarget<SparcSubtarget>().getAdjustedFrameSize(NumBytes);
   emitSPAdjustment(MF, MBB, MBBI, NumBytes, SP::ADDrr, SP::ADDri);
 }
 
@@ -180,12 +200,60 @@ bool SparcFrameLowering::hasReservedCallFrame(const MachineFunction &MF) const {
 // pointer register.  This is true if the function has variable sized allocas or
 // if frame pointer elimination is disabled.
 bool SparcFrameLowering::hasFP(const MachineFunction &MF) const {
+  const TargetRegisterInfo *RegInfo = MF.getSubtarget().getRegisterInfo();
+
   const MachineFrameInfo *MFI = MF.getFrameInfo();
   return MF.getTarget().Options.DisableFramePointerElim(MF) ||
-    MFI->hasVarSizedObjects() || MFI->isFrameAddressTaken();
+      RegInfo->needsStackRealignment(MF) ||
+      MFI->hasVarSizedObjects() ||
+      MFI->isFrameAddressTaken();
 }
 
 
+int SparcFrameLowering::getFrameIndexReference(const MachineFunction &MF, int FI,
+                                               unsigned &FrameReg) const {
+  const SparcSubtarget &Subtarget = MF.getSubtarget<SparcSubtarget>();
+  const MachineFrameInfo *MFI = MF.getFrameInfo();
+  const SparcRegisterInfo *RegInfo = Subtarget.getRegisterInfo();
+  const SparcMachineFunctionInfo *FuncInfo = MF.getInfo<SparcMachineFunctionInfo>();
+  bool isFixed = MFI->isFixedObjectIndex(FI);
+
+  // Addressable stack objects are accessed using neg. offsets from
+  // %fp, or positive offsets from %sp.
+  bool UseFP;
+
+  // Sparc uses FP-based references in general, even when "hasFP" is
+  // false. That function is rather a misnomer, because %fp is
+  // actually always available, unless isLeafProc.
+  if (FuncInfo->isLeafProc()) {
+    // If there's a leaf proc, all offsets need to be %sp-based,
+    // because we haven't caused %fp to actually point to our frame.
+    UseFP = false;
+  } else if (isFixed) {
+    // Otherwise, argument access should always use %fp.
+    UseFP = true;
+  } else if (RegInfo->needsStackRealignment(MF)) {
+    // If there is dynamic stack realignment, all local object
+    // references need to be via %sp, to take account of the
+    // re-alignment.
+    UseFP = false;
+  } else {
+    // Finally, default to using %fp.
+    UseFP = true;
+  }
+
+  int64_t FrameOffset = MF.getFrameInfo()->getObjectOffset(FI) +
+      Subtarget.getStackPointerBias();
+
+  if (UseFP) {
+    FrameReg = RegInfo->getFrameRegister(MF);
+    return FrameOffset;
+  } else {
+    FrameReg = SP::O6; // %sp
+    return FrameOffset + MF.getFrameInfo()->getStackSize();
+  }
+}
+
 static bool LLVM_ATTRIBUTE_UNUSED verifyLeafProcRegUse(MachineRegisterInfo *MRI)
 {
 

lib/Target/Sparc/SparcFrameLowering.h

@@ -39,6 +39,8 @@ public:
   void determineCalleeSaves(MachineFunction &MF, BitVector &SavedRegs,
                             RegScavenger *RS = nullptr) const override;
 
+  int getFrameIndexReference(const MachineFunction &MF, int FI,
+                             unsigned &FrameReg) const override;
 private:
   // Remap input registers to output registers for leaf procedure.
   void remapRegsForLeafProc(MachineFunction &MF) const;

lib/Target/Sparc/SparcRegisterInfo.cpp

@@ -170,21 +170,15 @@ SparcRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
   MachineInstr &MI = *II;
   DebugLoc dl = MI.getDebugLoc();
   int FrameIndex = MI.getOperand(FIOperandNum).getIndex();
-
-  // Addressable stack objects are accessed using neg. offsets from %fp
   MachineFunction &MF = *MI.getParent()->getParent();
   const SparcSubtarget &Subtarget = MF.getSubtarget<SparcSubtarget>();
-  int64_t Offset = MF.getFrameInfo()->getObjectOffset(FrameIndex) +
+  const SparcFrameLowering *TFI = getFrameLowering(MF);
-                   MI.getOperand(FIOperandNum + 1).getImm() +
+
-                   Subtarget.getStackPointerBias();
+  unsigned FrameReg;
-  SparcMachineFunctionInfo *FuncInfo = MF.getInfo<SparcMachineFunctionInfo>();
+  int Offset;
-  unsigned FramePtr = SP::I6;
+  Offset = TFI->getFrameIndexReference(MF, FrameIndex, FrameReg);
-  if (FuncInfo->isLeafProc()) {
+
-    // Use %sp and adjust offset if needed.
+  Offset += MI.getOperand(FIOperandNum + 1).getImm();
-    FramePtr = SP::O6;
-    int stackSize = MF.getFrameInfo()->getStackSize();
-    Offset += (stackSize) ? Subtarget.getAdjustedFrameSize(stackSize) : 0 ;
-  }
 
   if (!Subtarget.isV9() || !Subtarget.hasHardQuad()) {
     if (MI.getOpcode() == SP::STQFri) {
@@ -194,8 +188,8 @@ SparcRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
       unsigned SrcOddReg  = getSubReg(SrcReg, SP::sub_odd64);
       MachineInstr *StMI =
         BuildMI(*MI.getParent(), II, dl, TII.get(SP::STDFri))
-        .addReg(FramePtr).addImm(0).addReg(SrcEvenReg);
+        .addReg(FrameReg).addImm(0).addReg(SrcEvenReg);
-      replaceFI(MF, II, *StMI, dl, 0, Offset, FramePtr);
+      replaceFI(MF, II, *StMI, dl, 0, Offset, FrameReg);
       MI.setDesc(TII.get(SP::STDFri));
       MI.getOperand(2).setReg(SrcOddReg);
       Offset += 8;
@@ -206,8 +200,8 @@ SparcRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
       unsigned DestOddReg  = getSubReg(DestReg, SP::sub_odd64);
       MachineInstr *StMI =
         BuildMI(*MI.getParent(), II, dl, TII.get(SP::LDDFri), DestEvenReg)
-        .addReg(FramePtr).addImm(0);
+        .addReg(FrameReg).addImm(0);
-      replaceFI(MF, II, *StMI, dl, 1, Offset, FramePtr);
+      replaceFI(MF, II, *StMI, dl, 1, Offset, FrameReg);
 
       MI.setDesc(TII.get(SP::LDDFri));
       MI.getOperand(0).setReg(DestOddReg);
@@ -215,10 +209,33 @@ SparcRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
     }
   }
 
-  replaceFI(MF, II, MI, dl, FIOperandNum, Offset, FramePtr);
+  replaceFI(MF, II, MI, dl, FIOperandNum, Offset, FrameReg);
 
 }
 
 unsigned SparcRegisterInfo::getFrameRegister(const MachineFunction &MF) const {
   return SP::I6;
 }
+
+// Sparc has no architectural need for stack realignment support,
+// except that LLVM unfortunately currently implements overaligned
+// stack objects by depending upon stack realignment support.
+// If that ever changes, this can probably be deleted.
+bool SparcRegisterInfo::canRealignStack(const MachineFunction &MF) const {
+  if (!TargetRegisterInfo::canRealignStack(MF))
+    return false;
+
+  // Sparc always has a fixed frame pointer register, so don't need to
+  // worry about needing to reserve it. [even if we don't have a frame
+  // pointer for our frame, it still cannot be used for other things,
+  // or register window traps will be SADNESS.]
+
+  // If there's a reserved call frame, we can use SP to access locals.
+  if (getFrameLowering(MF)->hasReservedCallFrame(MF))
+    return true;
+
+  // Otherwise, we'd need a base pointer, but those aren't implemented
+  // for SPARC at the moment.
+
+  return false;
+}

lib/Target/Sparc/SparcRegisterInfo.h

@@ -42,8 +42,10 @@ struct SparcRegisterInfo : public SparcGenRegisterInfo {
   void processFunctionBeforeFrameFinalized(MachineFunction &MF,
                                        RegScavenger *RS = nullptr) const;
 
-  // Debug information queries.
   unsigned getFrameRegister(const MachineFunction &MF) const override;
+
+  bool canRealignStack(const MachineFunction &MF) const override;
+
 };
 
 } // end namespace llvm

test/CodeGen/SPARC/fp128.ll

@@ -45,14 +45,14 @@ entry:
 ; HARD:       std %f{{.+}}, [%[[S1:.+]]]
 ; HARD-DAG:   ldd [%[[S0]]], %f{{.+}}
 ; HARD-DAG:   ldd [%[[S1]]], %f{{.+}}
-; HARD:       jmp %o7+12
+; HARD:       jmp {{%[oi]7}}+12
 
 ; SOFT-LABEL: f128_spill
 ; SOFT:       std %f{{.+}}, [%[[S0:.+]]]
 ; SOFT:       std %f{{.+}}, [%[[S1:.+]]]
 ; SOFT-DAG:   ldd [%[[S0]]], %f{{.+}}
 ; SOFT-DAG:   ldd [%[[S1]]], %f{{.+}}
-; SOFT:       jmp %o7+12
+; SOFT:       jmp {{%[oi]7}}+12
 
 define void @f128_spill(fp128* noalias sret %scalar.result, fp128* byval %a) {
 entry:

test/CodeGen/SPARC/stack-align.ll

@@ -0,0 +1,22 @@
+; RUN: llc -march=sparc < %s | FileCheck %s
+declare void @stack_realign_helper(i32 %a, i32* %b)
+
+@foo = global i32 1
+
+;; This is a function where we have a local variable of 64-byte
+;; alignment.  We want to see that the stack is aligned (the initial
+;; andn), that the local var is accessed via stack pointer (to %o0), and that
+;; the argument is accessed via frame pointer not stack pointer (to %o1).
+
+;; CHECK-LABEL: stack_realign:
+;; CHECK:      andn %sp, 63, %sp
+;; CHECK-NEXT: ld [%fp+92], %o0
+;; CHECK-NEXT: call stack_realign_helper
+;; CHECK-NEXT: add %sp, 96, %o1
+
+define void @stack_realign(i32 %a, i32 %b, i32 %c, i32 %d, i32 %e, i32 %f, i32 %g) {
+entry:
+  %aligned = alloca i32, align 64
+  call void @stack_realign_helper(i32 %g, i32* %aligned)
+  ret void
+}