[SPARC] Allow overaligned allocas (#107223)

SPARC ABI doesn't use stack realignment, so let LLVM know about it in
`SparcFrameLowering`. This has the side effect of making all overaligned
allocations go through `LowerDYNAMIC_STACKALLOC`, so implement the
missing logic there too for overaligned allocations.
This makes the SPARC backend not crash on overaligned `alloca`s and fix
#89569.

compiler-rt/test/asan/TestCases/alloca_vla_interact.cpp

@@ -3,6 +3,9 @@
 //
 // REQUIRES: stable-runtime
 
+// See https://github.com/llvm/llvm-project/issues/110956
+// XFAIL: target=sparc{{.*}}
+
 // This testcase checks correct interaction between VLAs and allocas.
 
 #include <assert.h>

llvm/lib/Target/Sparc/SparcFrameLowering.cpp

@@ -35,7 +35,8 @@ DisableLeafProc("disable-sparc-leaf-proc",
 SparcFrameLowering::SparcFrameLowering(const SparcSubtarget &ST)
     : TargetFrameLowering(TargetFrameLowering::StackGrowsDown,
                           ST.is64Bit() ? Align(16) : Align(8), 0,
-                          ST.is64Bit() ? Align(16) : Align(8)) {}
+                          ST.is64Bit() ? Align(16) : Align(8),
+                          /*StackRealignable=*/false) {}
 
 void SparcFrameLowering::emitSPAdjustment(MachineFunction &MF,
                                           MachineBasicBlock &MBB,
@@ -97,12 +98,6 @@ void SparcFrameLowering::emitPrologue(MachineFunction &MF,
   // Debug location must be unknown since the first debug location is used
   // to determine the end of the prologue.
   DebugLoc dl;
-  bool NeedsStackRealignment = RegInfo.shouldRealignStack(MF);
-
-  if (NeedsStackRealignment && !RegInfo.canRealignStack(MF))
-    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();
@@ -168,31 +163,6 @@ void SparcFrameLowering::emitPrologue(MachineFunction &MF,
       MCCFIInstruction::createRegister(nullptr, regOutRA, regInRA));
   BuildMI(MBB, MBBI, dl, TII.get(TargetOpcode::CFI_INSTRUCTION))
       .addCFIIndex(CFIIndex);
-
-  if (NeedsStackRealignment) {
-    int64_t Bias = Subtarget.getStackPointerBias();
-    unsigned regUnbiased;
-    if (Bias) {
-      // This clobbers G1 which we always know is available here.
-      regUnbiased = SP::G1;
-      // add %o6, BIAS, %g1
-      BuildMI(MBB, MBBI, dl, TII.get(SP::ADDri), regUnbiased)
-        .addReg(SP::O6).addImm(Bias);
-    } else
-      regUnbiased = SP::O6;
-
-    // andn %regUnbiased, MaxAlign-1, %regUnbiased
-    Align MaxAlign = MFI.getMaxAlign();
-    BuildMI(MBB, MBBI, dl, TII.get(SP::ANDNri), regUnbiased)
-        .addReg(regUnbiased)
-        .addImm(MaxAlign.value() - 1U);
-
-    if (Bias) {
-      // add %g1, -BIAS, %o6
-      BuildMI(MBB, MBBI, dl, TII.get(SP::ADDri), SP::O6)
-        .addReg(regUnbiased).addImm(-Bias);
-    }
-  }
 }
 
 MachineBasicBlock::iterator SparcFrameLowering::
@@ -257,8 +227,7 @@ bool SparcFrameLowering::hasFPImpl(const MachineFunction &MF) const {
 
   const MachineFrameInfo &MFI = MF.getFrameInfo();
   return MF.getTarget().Options.DisableFramePointerElim(MF) ||
-         RegInfo->hasStackRealignment(MF) || MFI.hasVarSizedObjects() ||
-         MFI.isFrameAddressTaken();
+         MFI.hasVarSizedObjects() || MFI.isFrameAddressTaken();
 }
 
 StackOffset
@@ -284,11 +253,6 @@ SparcFrameLowering::getFrameIndexReference(const MachineFunction &MF, int FI,
   } else if (isFixed) {
     // Otherwise, argument access should always use %fp.
     UseFP = true;
-  } else if (RegInfo->hasStackRealignment(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;

llvm/lib/Target/Sparc/SparcISelLowering.cpp

@@ -2762,22 +2762,16 @@ static SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) {
 
 static SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG,
                                        const SparcSubtarget *Subtarget) {
-  SDValue Chain = Op.getOperand(0);  // Legalize the chain.
-  SDValue Size  = Op.getOperand(1);  // Legalize the size.
-  MaybeAlign Alignment =
-      cast<ConstantSDNode>(Op.getOperand(2))->getMaybeAlignValue();
-  Align StackAlign = Subtarget->getFrameLowering()->getStackAlign();
+  SDValue Chain = Op.getOperand(0);
+  SDValue Size = Op.getOperand(1);
+  SDValue Alignment = Op.getOperand(2);
+  MaybeAlign MaybeAlignment =
+      cast<ConstantSDNode>(Alignment)->getMaybeAlignValue();
   EVT VT = Size->getValueType(0);
   SDLoc dl(Op);
 
-  // TODO: implement over-aligned alloca. (Note: also implies
-  // supporting support for overaligned function frames + dynamic
-  // allocations, at all, which currently isn't supported)
-  if (Alignment && *Alignment > StackAlign) {
-    const MachineFunction &MF = DAG.getMachineFunction();
-    report_fatal_error("Function \"" + Twine(MF.getName()) + "\": "
-                       "over-aligned dynamic alloca not supported.");
-  }
+  unsigned SPReg = SP::O6;
+  SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
 
   // The resultant pointer needs to be above the register spill area
   // at the bottom of the stack.
@@ -2811,16 +2805,29 @@ static SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG,
     regSpillArea = 96;
   }
 
-  unsigned SPReg = SP::O6;
-  SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
-  SDValue NewSP = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
-  Chain = DAG.getCopyToReg(SP.getValue(1), dl, SPReg, NewSP);    // Output chain
-
-  regSpillArea += Subtarget->getStackPointerBias();
-
-  SDValue NewVal = DAG.getNode(ISD::ADD, dl, VT, NewSP,
-                               DAG.getConstant(regSpillArea, dl, VT));
-  SDValue Ops[2] = { NewVal, Chain };
+  int64_t Bias = Subtarget->getStackPointerBias();
+
+  // Debias and increment SP past the reserved spill area.
+  // We need the SP to point to the first usable region before calculating
+  // anything to prevent any of the pointers from becoming out of alignment when
+  // we rebias the SP later on.
+  SDValue StartOfUsableStack = DAG.getNode(
+      ISD::ADD, dl, VT, SP, DAG.getConstant(regSpillArea + Bias, dl, VT));
+  SDValue AllocatedPtr =
+      DAG.getNode(ISD::SUB, dl, VT, StartOfUsableStack, Size);
+
+  bool IsOveraligned = MaybeAlignment.has_value();
+  SDValue AlignedPtr =
+      IsOveraligned
+          ? DAG.getNode(ISD::AND, dl, VT, AllocatedPtr,
+                        DAG.getConstant(-MaybeAlignment->value(), dl, VT))
+          : AllocatedPtr;
+
+  // Now that we are done, restore the bias and reserved spill area.
+  SDValue NewSP = DAG.getNode(ISD::SUB, dl, VT, AlignedPtr,
+                              DAG.getConstant(regSpillArea + Bias, dl, VT));
+  Chain = DAG.getCopyToReg(SP.getValue(1), dl, SPReg, NewSP);
+  SDValue Ops[2] = {AlignedPtr, Chain};
   return DAG.getMergeValues(Ops, dl);
 }
 

llvm/lib/Target/Sparc/SparcRegisterInfo.cpp

@@ -226,26 +226,3 @@ SparcRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
 Register 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;
-}

llvm/lib/Target/Sparc/SparcRegisterInfo.h

@@ -40,9 +40,6 @@ struct SparcRegisterInfo : public SparcGenRegisterInfo {
                            RegScavenger *RS = nullptr) const override;
 
   Register getFrameRegister(const MachineFunction &MF) const override;
-
-  bool canRealignStack(const MachineFunction &MF) const override;
-
 };
 
 } // end namespace llvm

llvm/test/CodeGen/Generic/ForceStackAlign.ll

@@ -5,9 +5,6 @@
 ; CHECK-LABEL: @f
 ; CHECK-LABEL: @g
 
-; Stack realignment not supported.
-; XFAIL: target=sparc{{.*}}
-
 ; NVPTX can only select dynamic_stackalloc on sm_52+ and with ptx73+
 ; XFAIL: target=nvptx{{.*}}
 

llvm/test/CodeGen/SPARC/2013-05-17-CallFrame.ll

@@ -5,14 +5,10 @@
 ; (this should ideally be doing "add 4+7; and -8", instead of
 ;  "add 7; and -8; add 8"; see comments in LowerDYNAMIC_STACKALLOC)
 
-define void @variable_alloca_with_adj_call_stack(i32 %num) {
+define void @variable_alloca_with_adj_call_stack(i32 %num) nounwind {
 ; V8-LABEL: variable_alloca_with_adj_call_stack:
-; V8:         .cfi_startproc
-; V8-NEXT:  ! %bb.0: ! %entry
+; V8:       ! %bb.0: ! %entry
 ; V8-NEXT:    save %sp, -96, %sp
-; V8-NEXT:    .cfi_def_cfa_register %fp
-; V8-NEXT:    .cfi_window_save
-; V8-NEXT:    .cfi_register %o7, %i7
 ; V8-NEXT:    add %i0, 7, %i0
 ; V8-NEXT:    and %i0, -8, %i0
 ; V8-NEXT:    sub %sp, %i0, %i0
@@ -34,12 +30,8 @@ define void @variable_alloca_with_adj_call_stack(i32 %num) {
 ; V8-NEXT:    restore
 ;
 ; SPARC64-LABEL: variable_alloca_with_adj_call_stack:
-; SPARC64:         .cfi_startproc
-; SPARC64-NEXT:  ! %bb.0: ! %entry
+; SPARC64:       ! %bb.0: ! %entry
 ; SPARC64-NEXT:    save %sp, -128, %sp
-; SPARC64-NEXT:    .cfi_def_cfa_register %fp
-; SPARC64-NEXT:    .cfi_window_save
-; SPARC64-NEXT:    .cfi_register %o7, %i7
 ; SPARC64-NEXT:    srl %i0, 0, %i0
 ; SPARC64-NEXT:    add %i0, 15, %i0
 ; SPARC64-NEXT:    sethi 4194303, %i1

llvm/test/CodeGen/SPARC/alloca-align.ll

@@ -0,0 +1,93 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -march=sparc < %s | FileCheck %s --check-prefixes=CHECK32
+; RUN: llc -march=sparcv9 < %s | FileCheck %s --check-prefixes=CHECK64
+
+define void @variable_alloca_with_overalignment(i32 %num) nounwind {
+; CHECK32-LABEL: variable_alloca_with_overalignment:
+; CHECK32:       ! %bb.0:
+; CHECK32-NEXT:    save %sp, -96, %sp
+; CHECK32-NEXT:    add %sp, 80, %i1
+; CHECK32-NEXT:    and %i1, -64, %o0
+; CHECK32-NEXT:    add %o0, -96, %sp
+; CHECK32-NEXT:    add %i0, 7, %i0
+; CHECK32-NEXT:    and %i0, -8, %i0
+; CHECK32-NEXT:    sub %sp, %i0, %i0
+; CHECK32-NEXT:    add %i0, -8, %sp
+; CHECK32-NEXT:    call foo
+; CHECK32-NEXT:    add %i0, 88, %o1
+; CHECK32-NEXT:    ret
+; CHECK32-NEXT:    restore
+;
+; CHECK64-LABEL: variable_alloca_with_overalignment:
+; CHECK64:       ! %bb.0:
+; CHECK64-NEXT:    save %sp, -128, %sp
+; CHECK64-NEXT:    add %sp, 2159, %i1
+; CHECK64-NEXT:    and %i1, -64, %o0
+; CHECK64-NEXT:    add %o0, -2175, %sp
+; CHECK64-NEXT:    srl %i0, 0, %i0
+; CHECK64-NEXT:    add %i0, 15, %i0
+; CHECK64-NEXT:    sethi 4194303, %i1
+; CHECK64-NEXT:    or %i1, 1008, %i1
+; CHECK64-NEXT:    sethi 0, %i2
+; CHECK64-NEXT:    or %i2, 1, %i2
+; CHECK64-NEXT:    sllx %i2, 32, %i2
+; CHECK64-NEXT:    or %i2, %i1, %i1
+; CHECK64-NEXT:    and %i0, %i1, %i0
+; CHECK64-NEXT:    sub %sp, %i0, %i0
+; CHECK64-NEXT:    add %i0, 2175, %o1
+; CHECK64-NEXT:    mov %i0, %sp
+; CHECK64-NEXT:    call foo
+; CHECK64-NEXT:    add %sp, -48, %sp
+; CHECK64-NEXT:    add %sp, 48, %sp
+; CHECK64-NEXT:    ret
+; CHECK64-NEXT:    restore
+  %aligned = alloca i32, align 64
+  %var_size = alloca i8, i32 %num, align 4
+  call void @foo(ptr %aligned, ptr %var_size)
+  ret void
+}
+
+;; Same but with the alloca itself overaligned
+define void @variable_alloca_with_overalignment_2(i32 %num) nounwind {
+; CHECK32-LABEL: variable_alloca_with_overalignment_2:
+; CHECK32:       ! %bb.0:
+; CHECK32-NEXT:    save %sp, -96, %sp
+; CHECK32-NEXT:    add %i0, 7, %i0
+; CHECK32-NEXT:    and %i0, -8, %i0
+; CHECK32-NEXT:    sub %sp, %i0, %i0
+; CHECK32-NEXT:    add %i0, 88, %i0
+; CHECK32-NEXT:    and %i0, -64, %o1
+; CHECK32-NEXT:    add %o1, -96, %sp
+; CHECK32-NEXT:    call foo
+; CHECK32-NEXT:    mov %g0, %o0
+; CHECK32-NEXT:    ret
+; CHECK32-NEXT:    restore
+;
+; CHECK64-LABEL: variable_alloca_with_overalignment_2:
+; CHECK64:       ! %bb.0:
+; CHECK64-NEXT:    save %sp, -128, %sp
+; CHECK64-NEXT:    srl %i0, 0, %i0
+; CHECK64-NEXT:    add %i0, 15, %i0
+; CHECK64-NEXT:    sethi 4194303, %i1
+; CHECK64-NEXT:    or %i1, 1008, %i1
+; CHECK64-NEXT:    sethi 0, %i2
+; CHECK64-NEXT:    or %i2, 1, %i2
+; CHECK64-NEXT:    sllx %i2, 32, %i2
+; CHECK64-NEXT:    or %i2, %i1, %i1
+; CHECK64-NEXT:    and %i0, %i1, %i0
+; CHECK64-NEXT:    sub %sp, %i0, %i0
+; CHECK64-NEXT:    add %i0, 2175, %i0
+; CHECK64-NEXT:    and %i0, -64, %o1
+; CHECK64-NEXT:    add %o1, -2175, %sp
+; CHECK64-NEXT:    add %sp, -48, %sp
+; CHECK64-NEXT:    call foo
+; CHECK64-NEXT:    mov %g0, %o0
+; CHECK64-NEXT:    add %sp, 48, %sp
+; CHECK64-NEXT:    ret
+; CHECK64-NEXT:    restore
+  %var_size = alloca i8, i32 %num, align 64
+  call void @foo(ptr null, ptr %var_size)
+  ret void
+}
+
+declare void @foo(ptr, ptr);

llvm/test/CodeGen/SPARC/fp128.ll

@@ -54,18 +54,10 @@ entry:
 
 ; CHECK-LABEL: f128_spill_large:
 ; CHECK:       sethi 4, %g1
-; CHECK:       sethi 4, %g1
-; CHECK-NEXT:  add %g1, %sp, %g1
-; CHECK-NEXT:  std %f{{.+}}, [%g1]
-; CHECK:       sethi 4, %g1
-; CHECK-NEXT:  add %g1, %sp, %g1
-; CHECK-NEXT:  std %f{{.+}}, [%g1+8]
-; CHECK:       sethi 4, %g1
-; CHECK-NEXT:  add %g1, %sp, %g1
-; CHECK-NEXT:  ldd [%g1], %f{{.+}}
-; CHECK:       sethi 4, %g1
-; CHECK-NEXT:  add %g1, %sp, %g1
-; CHECK-NEXT:  ldd [%g1+8], %f{{.+}}
+; CHECK:       std %f{{.+}}, [%fp+-16]
+; CHECK-NEXT:  std %f{{.+}}, [%fp+-8]
+; CHECK:       ldd [%fp+-16], %f{{.+}}
+; CHECK-NEXT:  ldd [%fp+-8], %f{{.+}}
 
 define void @f128_spill_large(ptr noalias sret(<251 x fp128>) %scalar.result, ptr byval(<251 x fp128>) %a) {
 entry: