[compiler-rt] r355738 - [HWASan] Save + print registers when tag mismatch occurs in AArch64.

[Adhemerval Zanella via llvm-commits]

On 08/03/2019 18:22, Mitch Phillips via llvm-commits wrote:
> Author: hctim
> Date: Fri Mar  8 13:22:35 2019
> New Revision: 355738
> 
> URL: http://llvm.org/viewvc/llvm-project?rev=355738&view=rev
> Log:
> [HWASan] Save + print registers when tag mismatch occurs in AArch64.
> 

[...]

> Added: compiler-rt/trunk/lib/hwasan/hwasan_tag_mismatch_aarch64.S
> URL: http://llvm.org/viewvc/llvm-project/compiler-rt/trunk/lib/hwasan/hwasan_tag_mismatch_aarch64.S?rev=355738&view=auto
> ==============================================================================
> --- compiler-rt/trunk/lib/hwasan/hwasan_tag_mismatch_aarch64.S (added)
> +++ compiler-rt/trunk/lib/hwasan/hwasan_tag_mismatch_aarch64.S Fri Mar  8 13:22:35 2019
> @@ -0,0 +1,108 @@
> +#include "sanitizer_common/sanitizer_asm.h"
> +
> +// The content of this file is AArch64-only:
> +#if defined(__aarch64__)
> +
> +// The responsibility of the HWASan entry point in compiler-rt is to primarily
> +// readjust the stack from the callee and save the current register values to
> +// the stack.
> +// This entry point function should be called from a __hwasan_check_* symbol.
> +// These are generated during a lowering pass in the backend, and are found in
> +// AArch64AsmPrinter::EmitHwasanMemaccessSymbols(). Please look there for
> +// further information.
> +// The __hwasan_check_* caller of this function should have expanded the stack
> +// and saved the previous values of x0, x1, x29, and x30. This function will
> +// "consume" these saved values and treats it as part of its own stack frame.
> +// In this sense, the __hwasan_check_* callee and this function "share" a stack
> +// frame. This allows us to omit having unwinding information (.cfi_*) present
> +// in every __hwasan_check_* function, therefore reducing binary size. This is
> +// particularly important as hwasan_check_* instances are duplicated in every
> +// translation unit where HWASan is enabled.
> +// This function calls HwasanTagMismatch to step back into the C++ code that
> +// completes the stack unwinding and error printing. This function is is not
> +// permitted to return.
> +
> +
> +// Frame from __hwasan_check_:
> +// |              ...                |
> +// |              ...                |
> +// | Previous stack frames...        |
> +// +=================================+
> +// | Unused 8-bytes for maintaining  |
> +// | 16-byte SP alignment.           |
> +// +---------------------------------+
> +// | Return address (x30) for caller |
> +// | of __hwasan_check_*.            |
> +// +---------------------------------+
> +// | Frame address (x29) for caller  |
> +// | of __hwasan_check_*             |
> +// +---------------------------------+ <-- [SP + 232]
> +// |              ...                |
> +// |                                 |
> +// | Stack frame space for x2 - x28. |
> +// |                                 |
> +// |              ...                |
> +// +---------------------------------+ <-- [SP + 16]
> +// |                                 |
> +// | Saved x1, as __hwasan_check_*   |
> +// | clobbers it.                    |
> +// +---------------------------------+
> +// | Saved x0, likewise above.       |
> +// +---------------------------------+ <-- [x30 / SP]
> +
> +// This function takes two arguments:
> +//   * x0: The address of read/write instruction that caused HWASan check fail.
> +//   * x1: The tag size.
> +
> +.section .text
> +.file "hwasan_tag_mismatch_aarch64.S"
> +.global __hwasan_tag_mismatch
> +.type __hwasan_tag_mismatch, %function
> +__hwasan_tag_mismatch:
> +  CFI_STARTPROC
> +
> +  // Set the CFA to be the return address for caller of __hwasan_check_*. Note
> +  // that we do not emit CFI predicates to describe the contents of this stack
> +  // frame, as this proxy entry point should never be debugged. The contents
> +  // are static and are handled by the unwinder after calling
> +  // __hwasan_tag_mismatch. The frame pointer is already correctly setup
> +  // by __hwasan_check_*.
> +  add x29, sp, #232
> +  CFI_DEF_CFA(w29, 16)
> +  CFI_OFFSET(w30, -8)
> +  CFI_OFFSET(w29, -16)
> +
> +  // Save the rest of the registers into the preallocated space left by
> +  // __hwasan_check.
> +  str     x28,      [sp, #224]
> +  stp     x26, x27, [sp, #208]
> +  stp     x24, x25, [sp, #192]
> +  stp     x22, x23, [sp, #176]
> +  stp     x20, x21, [sp, #160]
> +  stp     x18, x19, [sp, #144]
> +  stp     x16, x17, [sp, #128]
> +  stp     x14, x15, [sp, #112]
> +  stp     x12, x13, [sp, #96]
> +  stp     x10, x11, [sp, #80]
> +  stp     x8,  x9,  [sp, #64]
> +  stp     x6,  x7,  [sp, #48]
> +  stp     x4,  x5,  [sp, #32]
> +  stp     x2,  x3,  [sp, #16]
> +
> +  // Pass the address of the frame to __hwasan_tag_mismatch_stub, so that it can
> +  // extract the saved registers from this frame without having to worry about
> +  // finding this frame.
> +  mov x2, sp
> +
> +  bl __hwasan_tag_mismatch_stub
> +  CFI_ENDPROC
> +
> +.Lfunc_end0:
> +  .size __hwasan_tag_mismatch, .Lfunc_end0-__hwasan_tag_mismatch
> +
> +.addrsig

This is causing a regression on llvm aarch bots [1] on the case of using system
compiler (in this case gcc/binutils) to build compiler-rt.

My understanding it should be ok to just remove the the assembly directive.


[1] http://lab.llvm.org:8011/builders/clang-cmake-aarch64-lld/builds/6162