Reland: [llvm][clang] Allocate a new stack instead of spawning a new thread to get more stack space (#136046)

Reland https://github.com/llvm/llvm-project/pull/133173

Clang spawns a new thread to avoid running out of stack space. This can
make debugging and performance analysis more difficult as how the
threads are connected is difficult to recover.

This patch introduces `runOnNewStack` and applies it in Clang. On
platforms that have good support for it this allocates a new stack and
moves to it using assembly. Doing split stacks like this actually runs
on most platforms, but many debuggers and unwinders reject the large or
backwards stack offsets that occur. Apple platforms and tools are known
to support this, so this only enables it there for now.

1 files changed, 127 insertions, 0 deletions

diff --git a/llvm/lib/Support/ProgramStack.cpp b/llvm/lib/Support/ProgramStack.cpp
new file mode 100644
index 0000000..66f74ff
--- /dev/null
+++ b/llvm/lib/Support/ProgramStack.cpp
@@ -0,0 +1,127 @@
+//===--- RunOnNewStack.cpp - Crash Recovery -------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Support/ProgramStack.h"
+#include "llvm/Config/config.h"
+#include "llvm/Support/Compiler.h"
+
+#ifdef LLVM_ON_UNIX
+# include <sys/resource.h> // for getrlimit
+#endif
+
+#ifdef _MSC_VER
+# include <intrin.h>  // for _AddressOfReturnAddress
+#endif
+
+#ifndef LLVM_HAS_SPLIT_STACKS
+# include "llvm/Support/thread.h"
+#endif
+
+using namespace llvm;
+
+uintptr_t llvm::getStackPointer() {
+#if __GNUC__ || __has_builtin(__builtin_frame_address)
+  return (uintptr_t)__builtin_frame_address(0);
+#elif defined(_MSC_VER)
+  return (uintptr_t)_AddressOfReturnAddress();
+#else
+  volatile char CharOnStack = 0;
+  // The volatile store here is intended to escape the local variable, to
+  // prevent the compiler from optimizing CharOnStack into anything other
+  // than a char on the stack.
+  //
+  // Tested on: MSVC 2015 - 2019, GCC 4.9 - 9, Clang 3.2 - 9, ICC 13 - 19.
+  char *volatile Ptr = &CharOnStack;
+  return (uintptr_t)Ptr;
+#endif
+}
+
+unsigned llvm::getDefaultStackSize() {
+#ifdef LLVM_ON_UNIX
+  rlimit RL;
+  getrlimit(RLIMIT_STACK, &RL);
+  return RL.rlim_cur;
+#else
+  // Clang recursively parses, instantiates templates, and evaluates constant
+  // expressions. We've found 8MiB to be a reasonable stack size given the way
+  // Clang works and the way C++ is commonly written.
+  return 8 << 20;
+#endif
+}
+
+// Not an anonymous namespace to avoid warning about undefined local function.
+namespace llvm {
+#ifdef LLVM_HAS_SPLIT_STACKS_AARCH64
+void runOnNewStackImpl(void *Stack, void (*Fn)(void *), void *Ctx) __asm__(
+    "_ZN4llvm17runOnNewStackImplEPvPFvS0_ES0_");
+
+// This can't use naked functions because there is no way to know if cfi
+// directives are being emitted or not.
+//
+// When adding new platforms it may be better to move to a .S file with macros
+// for dealing with platform differences.
+__asm__ (
+    ".globl  _ZN4llvm17runOnNewStackImplEPvPFvS0_ES0_\n\t"
+    ".p2align  2\n\t"
+    "_ZN4llvm17runOnNewStackImplEPvPFvS0_ES0_:\n\t"
+    ".cfi_startproc\n\t"
+    "mov       x16, sp\n\t"
+    "sub       x0, x0, #0x20\n\t"            // subtract space from stack
+    "stp       xzr, x16, [x0, #0x00]\n\t"    // save old sp
+    "stp       x29, x30, [x0, #0x10]\n\t"    // save fp, lr
+    "mov       sp, x0\n\t"                   // switch to new stack
+    "add       x29, x0, #0x10\n\t"           // switch to new frame
+    ".cfi_def_cfa w29, 16\n\t"
+    ".cfi_offset w30, -8\n\t"                // lr
+    ".cfi_offset w29, -16\n\t"               // fp
+
+    "mov       x0, x2\n\t"                   // Ctx is the only argument
+    "blr       x1\n\t"                       // call Fn
+
+    "ldp       x29, x30, [sp, #0x10]\n\t"    // restore fp, lr
+    "ldp       xzr, x16, [sp, #0x00]\n\t"    // load old sp
+    "mov       sp, x16\n\t"
+    "ret\n\t"
+    ".cfi_endproc"
+);
+#endif
+} // namespace llvm
+
+namespace {
+#ifdef LLVM_HAS_SPLIT_STACKS
+void callback(void *Ctx) {
+  (*reinterpret_cast<function_ref<void()> *>(Ctx))();
+}
+#endif
+} // namespace
+
+#ifdef LLVM_HAS_SPLIT_STACKS
+void llvm::runOnNewStack(unsigned StackSize, function_ref<void()> Fn) {
+  if (StackSize == 0)
+    StackSize = getDefaultStackSize();
+
+  // We use malloc here instead of mmap because:
+  //   - it's simpler,
+  //   - many malloc implementations will reuse the allocation in cases where
+  //     we're bouncing accross the edge of a stack boundry, and
+  //   - many malloc implemenations will already provide guard pages for
+  //     allocations this large.
+  void *Stack = malloc(StackSize);
+  void *BottomOfStack = (char *)Stack + StackSize;
+
+  runOnNewStackImpl(BottomOfStack, callback, &Fn);
+
+  free(Stack);
+}
+#else
+void llvm::runOnNewStack(unsigned StackSize, function_ref<void()> Fn) {
+  llvm::thread Thread(
+      StackSize == 0 ? std::nullopt : std::optional<unsigned>(StackSize), Fn);
+  Thread.join();
+}
+#endif