[cpu-profiler] Fix stack iterability during deopt (arm port)

We didn't think this affected arm but it turns out it does.

This is a port of
https://chromium-review.googlesource.com/c/v8/v8/+/1688049

Bug: v8:9057
Change-Id: I671737107e589673fc32c5fa1760d887d1bcc240
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/1826665Reviewed-by: Benedikt Meurer <bmeurer@chromium.org>
Commit-Queue: Peter Marshall <petermarshall@chromium.org>
Cr-Commit-Position: refs/heads/master@{#64024}

src/deoptimizer/arm/deoptimizer-arm.cc

@@ -123,6 +123,17 @@ void Deoptimizer::GenerateDeoptimizationEntries(MacroAssembler* masm,
     __ vstr(d0, r1, dst_offset);
   }
 
+  // Mark the stack as not iterable for the CPU profiler which won't be able to
+  // walk the stack without the return address.
+  {
+    UseScratchRegisterScope temps(masm);
+    Register is_iterable = temps.Acquire();
+    Register zero = r4;
+    __ Move(is_iterable, ExternalReference::stack_is_iterable_address(isolate));
+    __ mov(zero, Operand(0));
+    __ strb(zero, MemOperand(is_iterable));
+  }
+
   // Remove the saved registers from the stack.
   __ add(sp, sp, Operand(kSavedRegistersAreaSize));
 
@@ -209,6 +220,15 @@ void Deoptimizer::GenerateDeoptimizationEntries(MacroAssembler* masm,
   // Restore the registers from the stack.
   __ ldm(ia_w, sp, restored_regs);  // all but pc registers.
 
+  {
+    UseScratchRegisterScope temps(masm);
+    Register is_iterable = temps.Acquire();
+    Register one = r4;
+    __ Move(is_iterable, ExternalReference::stack_is_iterable_address(isolate));
+    __ mov(one, Operand(1));
+    __ strb(one, MemOperand(is_iterable));
+  }
+
   // Remove sp, lr and pc.
   __ Drop(3);
   {
@@ -218,6 +238,7 @@ void Deoptimizer::GenerateDeoptimizationEntries(MacroAssembler* masm,
     __ pop(lr);
     __ Jump(scratch);
   }
+
   __ stop();
 }
 

src/deoptimizer/arm64/deoptimizer-arm64.cc

@@ -189,6 +189,15 @@ void Deoptimizer::GenerateDeoptimizationEntries(MacroAssembler* masm,
   CopyRegListToFrame(masm, x1, FrameDescription::double_registers_offset(),
                      saved_double_registers, x2, x3, kDoubleRegistersOffset);
 
+  // Mark the stack as not iterable for the CPU profiler which won't be able to
+  // walk the stack without the return address.
+  {
+    UseScratchRegisterScope temps(masm);
+    Register is_iterable = temps.AcquireX();
+    __ Mov(is_iterable, ExternalReference::stack_is_iterable_address(isolate));
+    __ strb(xzr, MemOperand(is_iterable));
+  }
+
   // Remove the saved registers from the stack.
   DCHECK_EQ(kSavedRegistersAreaSize % kXRegSize, 0);
   __ Drop(kSavedRegistersAreaSize / kXRegSize);
@@ -251,6 +260,15 @@ void Deoptimizer::GenerateDeoptimizationEntries(MacroAssembler* masm,
   RestoreRegList(masm, saved_double_registers, x1,
                  FrameDescription::double_registers_offset());
 
+  {
+    UseScratchRegisterScope temps(masm);
+    Register is_iterable = temps.AcquireX();
+    Register one = x4;
+    __ Mov(is_iterable, ExternalReference::stack_is_iterable_address(isolate));
+    __ Mov(one, Operand(1));
+    __ strb(one, MemOperand(is_iterable));
+  }
+
   // TODO(all): ARM copies a lot (if not all) of the last output frame onto the
   // stack, then pops it all into registers. Here, we try to load it directly
   // into the relevant registers. Is this correct? If so, we should improve the