ARM: use AAPCS-style prologues for embedded MachO.

Darwin prologues save their GPRs in two stages: a narrow push of r0-r7 & lr,
followed by a wide push of the remaining registers if there are any. AAPCS uses
a single push.w instruction.

It turns out that, on average, enough registers get pushed that code is smaller
in the AAPCS prologue, which is a nice property for M-class programmers. They
also have other options available for back-traces, so can hopefully deal with
the fact that FP & LR aren't adjacent in memory.

rdar://problem/15909583

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@209895 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/ARM/ARMBaseRegisterInfo.cpp

@@ -45,9 +45,12 @@ using namespace llvm;
 
 ARMBaseRegisterInfo::ARMBaseRegisterInfo(const ARMSubtarget &sti)
     : ARMGenRegisterInfo(ARM::LR, 0, 0, ARM::PC), STI(sti), BasePtr(ARM::R6) {
-  if (STI.isTargetMachO())
-    FramePtr = ARM::R7;
-  else if (STI.isTargetWindows())
+  if (STI.isTargetMachO()) {
+    if (STI.isTargetDarwin() || STI.isThumb1Only())
+      FramePtr = ARM::R7;
+    else
+      FramePtr = ARM::R11;
+  } else if (STI.isTargetWindows())
     FramePtr = ARM::R11;
   else // ARM EABI
     FramePtr = STI.isThumb() ? ARM::R7 : ARM::R11;

lib/Target/ARM/ARMFrameLowering.cpp

@@ -220,7 +220,7 @@ void ARMFrameLowering::emitPrologue(MachineFunction &MF) const {
     case ARM::R10:
     case ARM::R11:
     case ARM::R12:
-      if (STI.isTargetMachO()) {
+      if (STI.isTargetDarwin()) {
         GPRCS2Size += 4;
         break;
       }
@@ -380,7 +380,7 @@ void ARMFrameLowering::emitPrologue(MachineFunction &MF) const {
       case ARM::R10:
       case ARM::R11:
       case ARM::R12:
-        if (STI.isTargetMachO())
+        if (STI.isTargetDarwin())
           break;
         // fallthrough
       case ARM::R0:
@@ -445,7 +445,7 @@ void ARMFrameLowering::emitPrologue(MachineFunction &MF) const {
       case ARM::R10:
       case ARM::R11:
       case ARM::R12:
-        if (STI.isTargetMachO()) {
+        if (STI.isTargetDarwin()) {
           unsigned DwarfReg =  MRI->getDwarfRegNum(Reg, true);
           unsigned Offset = MFI->getObjectOffset(FI);
           unsigned CFIIndex = MMI.addFrameInst(
@@ -810,7 +810,7 @@ void ARMFrameLowering::emitPushInst(MachineBasicBlock &MBB,
     unsigned LastReg = 0;
     for (; i != 0; --i) {
       unsigned Reg = CSI[i-1].getReg();
-      if (!(Func)(Reg, STI.isTargetMachO())) continue;
+      if (!(Func)(Reg, STI.isTargetDarwin())) continue;
 
       // D-registers in the aligned area DPRCS2 are NOT spilled here.
       if (Reg >= ARM::D8 && Reg < ARM::D8 + NumAlignedDPRCS2Regs)
@@ -888,7 +888,7 @@ void ARMFrameLowering::emitPopInst(MachineBasicBlock &MBB,
     bool DeleteRet = false;
     for (; i != 0; --i) {
       unsigned Reg = CSI[i-1].getReg();
-      if (!(Func)(Reg, STI.isTargetMachO())) continue;
+      if (!(Func)(Reg, STI.isTargetDarwin())) continue;
 
       // The aligned reloads from area DPRCS2 are not inserted here.
       if (Reg >= ARM::D8 && Reg < ARM::D8 + NumAlignedDPRCS2Regs)
@@ -1438,7 +1438,7 @@ ARMFrameLowering::processFunctionBeforeCalleeSavedScan(MachineFunction &MF,
     if (Spilled) {
       NumGPRSpills++;
 
-      if (!STI.isTargetMachO()) {
+      if (!STI.isTargetDarwin()) {
         if (Reg == ARM::LR)
           LRSpilled = true;
         CS1Spilled = true;
@@ -1460,7 +1460,7 @@ ARMFrameLowering::processFunctionBeforeCalleeSavedScan(MachineFunction &MF,
         break;
       }
     } else {
-      if (!STI.isTargetMachO()) {
+      if (!STI.isTargetDarwin()) {
         UnspilledCS1GPRs.push_back(Reg);
         continue;
       }

test/CodeGen/ARM/fold-stack-adjust.ll

@@ -12,11 +12,11 @@ declare void @bar(i8*)
 
 define void @check_simple() minsize {
 ; CHECK-LABEL: check_simple:
-; CHECK: push {r3, r4, r5, r6, r7, lr}
+; CHECK: push.w {r7, r8, r9, r10, r11, lr}
 ; CHECK-NOT: sub sp, sp,
 ; ...
 ; CHECK-NOT: add sp, sp,
-; CHECK: pop {r0, r1, r2, r3, r7, pc}
+; CHECK: pop.w {r0, r1, r2, r3, r11, pc}
 
 ; CHECK-T1-LABEL: check_simple:
 ; CHECK-T1: push {r3, r4, r5, r6, r7, lr}
@@ -44,11 +44,11 @@ define void @check_simple() minsize {
 
 define void @check_simple_too_big() minsize {
 ; CHECK-LABEL: check_simple_too_big:
-; CHECK: push {r7, lr}
+; CHECK: push.w {r11, lr}
 ; CHECK: sub sp,
 ; ...
 ; CHECK: add sp,
-; CHECK: pop {r7, pc}
+; CHECK: pop.w {r11, pc}
   %var = alloca i8, i32 64
   call void @bar(i8* %var)
   ret void
@@ -93,11 +93,11 @@ define void @check_vfp_fold() minsize {
 ; folded in except that doing so would clobber the value being returned.
 define i64 @check_no_return_clobber() minsize {
 ; CHECK-LABEL: check_no_return_clobber:
-; CHECK: push {r1, r2, r3, r4, r5, r6, r7, lr}
+; CHECK: push.w {r5, r6, r7, r8, r9, r10, r11, lr}
 ; CHECK-NOT: sub sp,
 ; ...
 ; CHECK: add sp, #24
-; CHECK: pop {r7, pc}
+; CHECK: pop.w {r11, pc}
 
   ; Just to keep iOS FileCheck within previous function:
 ; CHECK-IOS-LABEL: check_no_return_clobber:
@@ -176,9 +176,9 @@ define void @test_varsize(...) minsize {
 
 ; CHECK-LABEL: test_varsize:
 ; CHECK: sub	sp, #16
-; CHECK: push	{r5, r6, r7, lr}
+; CHECK: push.w {r9, r10, r11, lr}
 ; ...
-; CHECK: pop.w	{r2, r3, r7, lr}
+; CHECK: pop.w	{r2, r3, r11, lr}
 ; CHECK: add	sp, #16
 ; CHECK: bx	lr
 

test/CodeGen/ARM/interrupt-attr.ll

@@ -35,15 +35,15 @@ define arm_aapcscc void @irq_fn() alignstack(8) "interrupt"="IRQ" {
   ; Normal AAPCS function (r0-r3 pushed onto stack by hardware, lr set to
   ; appropriate sentinel so no special return needed).
 ; CHECK-M-LABEL: irq_fn:
-; CHECK-M: push {r4, r6, r7, lr}
-; CHECK-M: add r7, sp, #8
+; CHECK-M: push.w {r4, r10, r11, lr}
+; CHECK-M: add.w r11, sp, #8
 ; CHECK-M: mov r4, sp
 ; CHECK-M: bic r4, r4, #7
 ; CHECK-M: mov sp, r4
 ; CHECK-M: blx _bar
-; CHECK-M: sub.w r4, r7, #8
+; CHECK-M: sub.w r4, r11, #8
 ; CHECK-M: mov sp, r4
-; CHECK-M: pop {r4, r6, r7, pc}
+; CHECK-M: pop.w {r4, r10, r11, pc}
 
   call arm_aapcscc void @bar()
   ret void

test/CodeGen/ARM/none-macho.ll

@@ -48,8 +48,8 @@ define i32 @test_frame_ptr() {
 ; CHECK-LABEL: test_frame_ptr:
   call void @test_trap()
 
-  ; Frame pointer is r7 as for Darwin
-; CHECK: mov r7, sp
+  ; Frame pointer is r11.
+; CHECK: mov r11, sp
   ret i32 42
 }
 
@@ -63,11 +63,9 @@ define void @test_two_areas(%big_arr* %addr) {
   ; This goes with the choice of r7 as FP (largely). FP and LR have to be stored
   ; consecutively on the stack for the frame record to be valid, which means we
   ; need the 2 register-save areas employed by iOS.
-; CHECK-NON-FAST: push {r4, r5, r6, r7, lr}
-; CHECK-NON-FAST: push.w {r8, r9, r10, r11}
+; CHECK-NON-FAST: push.w {r4, r5, r6, r7, r8, r9, r10, r11, lr}
 ; ...
-; CHECK-NON-FAST: pop.w {r8, r9, r10, r11}
-; CHECK-NON-FAST: pop {r4, r5, r6, r7, pc}
+; CHECK-NON-FAST: pop.w {r4, r5, r6, r7, r8, r9, r10, r11, pc}
   ret void
 }