[ARM] Correctly handle combining segmented stacks with execute-only

[john.brawn]

Summary

Using segmented stacks with execute-only mostly works, but we need to use the correct movi32 opcode in 6-M, and there's one place where for thumb1 (i.e. 6-M and 8-M.base) a constant pool was unconditionally used which needed to be fixed.

[olista01]

LGTM

llvm/lib/Target/ARM/ARMFrameLowering.cpp

if (!MFI.needsSplitStackProlog())
     return;
 
   uint64_t StackSize = MFI.getStackSize();
 
   // Use R4 and R5 as scratch registers.
   // We save R4 and R5 before use and restore them before leaving the function.
   unsigned ScratchReg0 = ARM::R4;
   unsigned ScratchReg1 = ARM::R5;
+  unsigned MovOp = ST->useMovt() ? ARM::t2MOVi32imm : ARM::tMOVi32imm;
   uint64_t AlignedStackSize;
 
   MachineBasicBlock *PrevStackMBB = MF.CreateMachineBasicBlock();
   MachineBasicBlock *PostStackMBB = MF.CreateMachineBasicBlock();
   MachineBasicBlock *AllocMBB = MF.CreateMachineBasicBlock();
   MachineBasicBlock *GetMBB = MF.CreateMachineBasicBlock();
   MachineBasicBlock *McrMBB = MF.CreateMachineBasicBlock();
 
void ARMFrameLowering::adjustForSegmentedStacks(
   if (!CompareStackPointer && Thumb) {
     if (AlignedStackSize < 256) {
       BuildMI(McrMBB, DL, TII.get(ARM::tSUBi8), ScratchReg1)
           .add(condCodeOp())
           .addReg(ScratchReg1)
           .addImm(AlignedStackSize)
           .add(predOps(ARMCC::AL));
     } else {
-      if (Thumb2) {
-        BuildMI(McrMBB, DL, TII.get(ARM::t2MOVi32imm), ScratchReg0)
+      if (Thumb2 || ST->genExecuteOnly()) {
+        BuildMI(McrMBB, DL, TII.get(MovOp), ScratchReg0)
             .addImm(AlignedStackSize);
       } else {
         auto MBBI = McrMBB->end();
         auto RegInfo = STI.getRegisterInfo();
         RegInfo->emitLoadConstPool(*McrMBB, MBBI, DL, ScratchReg0, 0,
                                    AlignedStackSize);
       }
       BuildMI(McrMBB, DL, TII.get(ARM::tSUBrr), ScratchReg1)
if (AlignedStackSize < 256) {
           .addReg(ARM::SP)
           .addReg(ScratchReg0)
           .add(predOps(ARMCC::AL))
           .add(condCodeOp());
     }
   }
 
   if (Thumb && ST->isThumb1Only()) {
+    if (ST->genExecuteOnly()) {
+      BuildMI(GetMBB, DL, TII.get(MovOp), ScratchReg0)
+          .addExternalSymbol("__STACK_LIMIT");
+    } else {
       unsigned PCLabelId = ARMFI->createPICLabelUId();
       ARMConstantPoolValue *NewCPV = ARMConstantPoolSymbol::Create(
         MF.getFunction().getContext(), "__STACK_LIMIT", PCLabelId, 0);
       MachineConstantPool *MCP = MF.getConstantPool();
       unsigned CPI = MCP->getConstantPoolIndex(NewCPV, Align(4));
 
       // ldr SR0, [pc, offset(STACK_LIMIT)]
       BuildMI(GetMBB, DL, TII.get(ARM::tLDRpci), ScratchReg0)
           .addConstantPoolIndex(CPI)
           .add(predOps(ARMCC::AL));
+    }
 
     // ldr SR0, [SR0]
     BuildMI(GetMBB, DL, TII.get(ARM::tLDRi), ScratchReg0)
         .addReg(ScratchReg0)
         .addImm(0)
         .add(predOps(ARMCC::AL));
   } else {
     // Get TLS base address from the coprocessor
void ARMFrameLowering::adjustForSegmentedStacks(
   //   The amount size of stack required
   if (Thumb) {
     if (AlignedStackSize < 256) {
       BuildMI(AllocMBB, DL, TII.get(ARM::tMOVi8), ScratchReg0)
           .add(condCodeOp())
           .addImm(AlignedStackSize)
           .add(predOps(ARMCC::AL));
     } else {
-      if (Thumb2) {
-        BuildMI(AllocMBB, DL, TII.get(ARM::t2MOVi32imm), ScratchReg0)
+      if (Thumb2 || ST->genExecuteOnly()) {
+        BuildMI(AllocMBB, DL, TII.get(MovOp), ScratchReg0)
             .addImm(AlignedStackSize);
       } else {
         auto MBBI = AllocMBB->end();
         auto RegInfo = STI.getRegisterInfo();
         RegInfo->emitLoadConstPool(*AllocMBB, MBBI, DL, ScratchReg0, 0,
                                    AlignedStackSize);
       }
     }
void ARMFrameLowering::adjustForSegmentedStacks(
   //   The amount size of stack consumed to save function arguments.
   if (Thumb) {
     if (ARMFI->getArgumentStackSize() < 256) {
       BuildMI(AllocMBB, DL, TII.get(ARM::tMOVi8), ScratchReg1)
           .add(condCodeOp())
           .addImm(alignToARMConstant(ARMFI->getArgumentStackSize()))
           .add(predOps(ARMCC::AL));
     } else {
-      if (Thumb2) {
-        BuildMI(AllocMBB, DL, TII.get(ARM::t2MOVi32imm), ScratchReg1)
+      if (Thumb2 || ST->genExecuteOnly()) {
+        BuildMI(AllocMBB, DL, TII.get(MovOp), ScratchReg1)
             .addImm(alignToARMConstant(ARMFI->getArgumentStackSize()));
       } else {
         auto MBBI = AllocMBB->end();
         auto RegInfo = STI.getRegisterInfo();
         RegInfo->emitLoadConstPool(
             *AllocMBB, MBBI, DL, ScratchReg1, 0,
             alignToARMConstant(ARMFI->getArgumentStackSize()));
       }

llvm/test/CodeGen/ARM/execute-only-segmented-stack.ll

+; RUN: llc -mtriple=thumbv6m-linux-eabi      -mattr=+execute-only %s -o - | FileCheck %s --check-prefixes=CHECK,CHECK-NOMOVW-SIZE,CHECK-NOMOVW-LIMIT
+; RUN: llc -mtriple=thumbv8m.base-linux-eabi -mattr=+execute-only %s -o - | FileCheck %s --check-prefixes=CHECK,CHECK-MOVW-SIZE,CHECK-MOVW-LIMIT
+; RUN: llc -mtriple=thumbv7m-linux-eabi      -mattr=+execute-only %s -o - | FileCheck %s --check-prefixes=CHECK,CHECK-MOVW-SIZE,CHECK-MRC-LIMIT
+; RUN: llc -mtriple=thumbv8m.main-linux-eabi -mattr=+execute-only %s -o - | FileCheck %s --check-prefixes=CHECK,CHECK-MOVW-SIZE,CHECK-MRC-LIMIT
+
+%struct.large_struct = type { [500 x i8] }
+declare void @fn(ptr)
+
+; CHECK-LABEL:             test:
+; CHECK:                   mov  [[SP:r[0-9]+]], sp
+; CHECK-MOVW-SIZE-NEXT:    movw [[SIZE:r[0-9]+]], #4032
+; CHECK-NOMOVW-SIZE-NEXT:  movs [[SIZE:r[0-9]+]], #15
+; CHECK-NOMOVW-SIZE-NEXT:  lsls [[SIZE]], [[SIZE]], #8
+; CHECK-NOMOVW-SIZE-NEXT:  adds [[SIZE]], #192
+; CHECK-NEXT:              sub{{s?}} [[SP]], [[SP]], [[SIZE]]
+; CHECK-MOVW-LIMIT-NEXT:   movw [[LIMIT:r[0-9]+]], :lower16:__STACK_LIMIT
+; CHECK-MOVW-LIMIT-NEXT:   movt [[LIMIT]], :upper16:__STACK_LIMIT
+; CHECK-NOMOVW-LIMIT-NEXT: movs [[LIMIT:r[0-9]+]], :upper8_15:__STACK_LIMIT
+; CHECK-NOMOVW-LIMIT-NEXT: lsls [[LIMIT]], [[LIMIT]], #8
+; CHECK-NOMOVW-LIMIT-NEXT: adds [[LIMIT]], :upper0_7:__STACK_LIMIT
+; CHECK-NOMOVW-LIMIT-NEXT: lsls [[LIMIT]], [[LIMIT]], #8
+; CHECK-NOMOVW-LIMIT-NEXT: adds [[LIMIT]], :lower8_15:__STACK_LIMIT
+; CHECK-NOMOVW-LIMIT-NEXT: lsls [[LIMIT]], [[LIMIT]], #8
+; CHECK-NOMOVW-LIMIT-NEXT: adds [[LIMIT]], :lower0_7:__STACK_LIMIT
+; CHECK-MRC-LIMIT-NEXT:    p15, #0, [[LIMIT:r[0-9]+]], c13, c0, #3
+; CHECK-NEXT:              ldr [[LIMIT]], [[[LIMIT]]{{.*}}]
+; CHECK-NEXT:              cmp [[LIMIT]], [[SP]]
+; CHECK-NEXT:              bls
+; CHECK-NEXT:              @{{.*}}:
+; CHECK-MOVW-SIZE-NEXT:    movw r4, #4032
+; CHECK-NOMOVW-SIZE-NEXT:  movs r4, #15
+; CHECK-NOMOVW-SIZE-NEXT:  lsls r4, r4, #8
+; CHECK-NOMOVW-SIZE-NEXT:  adds r4, #192
+; CHECK-MOVW-SIZE-NEXT:    movw r5, #484
+; CHECK-NOMOVW-SIZE-NEXT:  movs r5, #1
+; CHECK-NOMOVW-SIZE-NEXT:  lsls r5, r5, #8
+; CHECK-NOMOVW-SIZE-NEXT:  adds r5, #228
+; CHECK-NEXT:              push {lr}
+; CHECK-NEXT:              bl __morestack
+
+define void @test(ptr byval(%struct.large_struct) align 4 %arg) #0 {
+  %ptr = alloca i32, i32 1000
+  call void @fn(ptr %ptr)
+  ret void
+}
+
+attributes #0 = { "split-stack" }