arm: Fix ttype encoding assertion failure.

[logan]

Summary

GCC 4.7 or newer emits 0x90 (indirect | pcrel) as the ttype encoding.
This would hit an assertion in cxa_personality.cpp. This commit fixes
the problem by relaxing the assertion.

http://llvm.org/PR28884

[mclow.lists]

src/cxa_personality.cpp

355

It's not clear to me how this accomplishes what you want.
You're looking for 00/10/90, right? Why not just check for that?

Why are you anding with 0x0f ?
Before, this would pass only a single value - DW_EH_PE_absptr (aka 0)
With this change, it passes 32 values: 00, 03, 10, 13, 20, 23, and so on.

Was that your intent?

[logan]

src/cxa_personality.cpp

355

ttypeEncoding is encoded with following rules:

1. Lower 4 bits stand for the representation of the data, such as absptr, uleb128, udata1, udata2, udata4, udata8, etc. These bits control the way we extract the bytes from the exception table.

2. Upper 4 bits stand for the post processing action, such as pcrel, indirect, etc. For example, if pcrel is specified, then we should add the value, which was read in step 1, with the address of the value.

My intention is to weaken the assertion (only assert the essential assumption) so that we don't have to alter the assertion if there are new configurations that I am not aware of or new compiler which is generating different ttypeEncoding.

Since the upcoming lines (L365) only uses sizeof(uintptr_t) to decode the TType pointer, it is not necessary to impose restriction on the upper 4 bits. That's the reason why I wrote ttypeEncoding & 0xf. For the same reason, both absptr and udata have the same meaning (4 bytes in the exception table) in this context, thus I am adding extra (ttypeEncoding & 0x0f) == DW_EH_PE_udata4.

[mclow.lists]

src/cxa_personality.cpp

355

Ok; thanks for the explanation. However, I'm still concerned.
The assert is there to catch bad assumptions. (i.e, this can never happen).

the old code would assert on 255 of 256 possible values.
It turns out that this is too strict - there are at least three values that we need to pass.

But your assertion passes 32 possible values (out of 256). So if something goes wrong, and a random value for ttypeEncoding gets passed in here, there's a 1 in 8 chance that the assertion will not fire. *Thats* my concern. It should never be an issue, because we don't have bugs, and never pass random values around, right? ;-)

As for "dealing with new configurations" or "new compilers", I would say those are very infrequent events; and I wouldn't worry about them until they happen. (Demonstrations that I am misinformed here are welcome)

[logan]

src/cxa_personality.cpp

355

Thanks. I'll update the patch soon.

[logan]

Refine assertions to address the comments from mclow.lists.

[mclow.lists]

Thanks.

[logan]

Thanks for reviewing. Committed as rL286760.

src/cxa_personality.cpp

get_shim_type_info(uint64_t ttypeIndex, const uint8_t* classInfo,
                    _Unwind_Exception* unwind_exception)
 {
     if (classInfo == 0)
     {
         // this should not happen.  Indicates corrupted eh_table.
         call_terminate(native_exception, unwind_exception);
     }
 
-    assert(ttypeEncoding == DW_EH_PE_absptr && "Unexpected TTypeEncoding");
+    assert(((ttypeEncoding == DW_EH_PE_absptr) ||  // LLVM or GCC 4.6
+            (ttypeEncoding == DW_EH_PE_pcrel) ||  // GCC 4.7 baremetal
+            (ttypeEncoding == (DW_EH_PE_pcrel | DW_EH_PE_indirect))) &&  // GCC 4.7 linux
+           "Unexpected TTypeEncoding");
     (void)ttypeEncoding;
 
     const uint8_t* ttypePtr = classInfo - ttypeIndex * sizeof(uintptr_t);
     return reinterpret_cast<const __shim_type_info *>(
         read_target2_value(ttypePtr));
 }
 #else // !LIBCXXABI_ARM_EHABI
 static
 const __shim_type_info*
exception_spec_can_catch(int64_t specIndex, const uint8_t* classInfo,
                          void* adjustedPtr, _Unwind_Exception* unwind_exception)
 {
     if (classInfo == 0)
     {
         // this should not happen.   Indicates corrupted eh_table.
         call_terminate(false, unwind_exception);
     }
 
-    assert(ttypeEncoding == DW_EH_PE_absptr && "Unexpected TTypeEncoding");
+    assert(((ttypeEncoding == DW_EH_PE_absptr) ||  // LLVM or GCC 4.6
+            (ttypeEncoding == DW_EH_PE_pcrel) ||  // GCC 4.7 baremetal
+            (ttypeEncoding == (DW_EH_PE_pcrel | DW_EH_PE_indirect))) &&  // GCC 4.7 linux
+           "Unexpected TTypeEncoding");
     (void)ttypeEncoding;
 
     // specIndex is negative of 1-based byte offset into classInfo;
     specIndex = -specIndex;
     --specIndex;
     const void** temp = reinterpret_cast<const void**>(
         reinterpret_cast<uintptr_t>(classInfo) +
         static_cast<uintptr_t>(specIndex) * sizeof(uintptr_t));

test/lit.cfg

 if 'PYLINT_IMPORT' in os.environ:
     config = object()
     lit_config = object()
 
 # name: The name of this test suite.
 config.name = 'libc++abi'
 
 # suffixes: A list of file extensions to treat as test files.
-config.suffixes = ['.cpp']
+config.suffixes = ['.cpp', '.s']
 
 # test_source_root: The root path where tests are located.
 config.test_source_root = os.path.dirname(__file__)
 
 # Infer the libcxx_test_source_root for configuration import.
 # If libcxx_source_root isn't specified in the config, assume that the libcxx
 # and libcxxabi source directories are sibling directories.
 libcxx_src_root = getattr(config, 'libcxx_src_root', None)

test/lit.site.cfg.in

 config.enable_threads           = "@LIBCXXABI_ENABLE_THREADS@"
 config.use_sanitizer            = "@LLVM_USE_SANITIZER@"
 config.enable_32bit             = "@LIBCXXABI_BUILD_32_BITS@"
 config.target_info              = "@LIBCXXABI_TARGET_INFO@"
 config.executor                 = "@LIBCXXABI_EXECUTOR@"
 config.libcxxabi_shared         = "@LIBCXXABI_ENABLE_SHARED@"
 config.enable_shared            = "@LIBCXX_ENABLE_SHARED@"
 config.enable_exceptions        = "@LIBCXXABI_ENABLE_EXCEPTIONS@"
+config.host_triple              = "@LLVM_HOST_TRIPLE@"
+config.target_triple            = "@TARGET_TRIPLE@"
 
 # Let the main config do the real work.
 lit_config.load_config(config, "@LIBCXXABI_SOURCE_DIR@/test/lit.cfg")

test/native/arm-linux-eabi/lit.local.cfg

+def is_arm_linux_eabi(triple):
+    return ('arm' in triple) and ('linux' in triple) and ('eabi' in triple)
+
+is_native = config.root.host_triple == config.root.target_triple
+if not is_native or not is_arm_linux_eabi(config.root.host_triple):
+    config.unsupported = True

test/native/arm-linux-eabi/ttype-encoding-00.pass.sh.s

+@ RUN: %cxx %flags %link_flags %s -o %t.exe
+@ RUN: %exec %t.exe
+
+@ PURPOSE: Check that 0x00 is a valid value for ttype encoding.  LLVM and
+@ GCC 4.6 are generating 0x00 as ttype encoding.  libc++abi should provide
+@ legacy support.
+
+@ NOTE:
+@
+@ This file is generated from the following C++ source code:
+@
+@ ```
+@ int main() {
+@   try {
+@     throw 5;
+@   } catch (int i) {
+@     if (i != 5)
+@       abort();
+@     return 0;
+@   }
+@ }
+@ ```
+
+  .syntax unified
+
+  .text
+  .globl  main
+  .p2align  2
+  .type main,%function
+main:                                   @ @main
+.Lfunc_begin0:
+  .fnstart
+@ BB#0:                                 @ %entry
+  .save {r11, lr}
+  push  {r11, lr}
+  .setfp  r11, sp
+  mov r11, sp
+  mov r0, #4
+  bl  __cxa_allocate_exception
+  mov r1, #5
+  str r1, [r0]
+.Ltmp0:
+  ldr r1, .LCPI0_0
+  mov r2, #0
+  bl  __cxa_throw
+.Ltmp1:
+
+@ BB#2:                                 @ %lpad
+.Ltmp2:
+  bl  __cxa_begin_catch
+  ldr r0, [r0]
+  cmp r0, #5
+  bne .LBB0_4
+@ BB#3:                                 @ %if.end
+  bl  __cxa_end_catch
+  mov r0, #0
+  pop {r11, lr}
+  bx  lr
+.LBB0_4:                                @ %if.then
+  bl  abort
+  .p2align  2
+@ BB#5:
+.LCPI0_0:
+  .long _ZTIi
+.Lfunc_end0:
+
+  .size main, .Lfunc_end0-main
+  .globl  __gxx_personality_v0
+  .personality __gxx_personality_v0
+  .handlerdata
+  .p2align  2
+GCC_except_table0:
+.Lexception0:
+  .byte 255                     @ @LPStart Encoding = omit
+  .byte 0                       @ @TType Encoding = absptr
+  .asciz  "\257\200"              @ @TType base offset
+  .byte 3                       @ Call site Encoding = udata4
+  .byte 39                      @ Call site table length
+  .long  .Lfunc_begin0-.Lfunc_begin0 @ >> Call Site 1 <<
+  .long  .Ltmp0-.Lfunc_begin0    @   Call between .Lfunc_begin0 and .Ltmp0
+  .long 0                       @     has no landing pad
+  .byte 0                       @   On action: cleanup
+  .long  .Ltmp0-.Lfunc_begin0    @ >> Call Site 2 <<
+  .long  .Ltmp1-.Ltmp0           @   Call between .Ltmp0 and .Ltmp1
+  .long .Ltmp2-.Lfunc_begin0    @     jumps to .Ltmp2
+  .byte 1                       @   On action: 1
+  .long  .Ltmp1-.Lfunc_begin0    @ >> Call Site 3 <<
+  .long  .Lfunc_end0-.Ltmp1      @   Call between .Ltmp1 and .Lfunc_end0
+  .long 0                       @     has no landing pad
+  .byte 0                       @   On action: cleanup
+  .byte  1                       @ >> Action Record 1 <<
+                                        @   Catch TypeInfo 1
+  .byte 0                       @   No further actions
+                                        @ >> Catch TypeInfos <<
+  .long _ZTIi(target2)          @ TypeInfo 1
+  .p2align  2
+  .fnend

test/native/arm-linux-eabi/ttype-encoding-90.pass.sh.s

+@ RUN: %cxx %flags %link_flags %s -o %t.exe
+@ RUN: %exec %t.exe
+
+@ PURPOSE: Check that 0x90 is a valid value for ttype encoding.
+
+@ NOTE:
+@
+@ This file is generated from the following C++ source code and then change the
+@ `TType Encoding` to 0x90.
+@
+@ ```
+@ int main() {
+@   try {
+@     throw 5;
+@   } catch (int i) {
+@     if (i != 5)
+@       abort();
+@     return 0;
+@   }
+@ }
+@ ```
+
+  .syntax unified
+
+  .text
+  .globl  main
+  .p2align  2
+  .type main,%function
+main:                                   @ @main
+.Lfunc_begin0:
+  .fnstart
+@ BB#0:                                 @ %entry
+  .save {r11, lr}
+  push  {r11, lr}
+  .setfp  r11, sp
+  mov r11, sp
+  mov r0, #4
+  bl  __cxa_allocate_exception
+  mov r1, #5
+  str r1, [r0]
+.Ltmp0:
+  ldr r1, .LCPI0_0
+  mov r2, #0
+  bl  __cxa_throw
+.Ltmp1:
+
+@ BB#2:                                 @ %lpad
+.Ltmp2:
+  bl  __cxa_begin_catch
+  ldr r0, [r0]
+  cmp r0, #5
+  bne .LBB0_4
+@ BB#3:                                 @ %if.end
+  bl  __cxa_end_catch
+  mov r0, #0
+  pop {r11, lr}
+  bx  lr
+.LBB0_4:                                @ %if.then
+  bl  abort
+  .p2align  2
+@ BB#5:
+.LCPI0_0:
+  .long _ZTIi
+.Lfunc_end0:
+
+  .size main, .Lfunc_end0-main
+  .globl  __gxx_personality_v0
+  .personality __gxx_personality_v0
+  .handlerdata
+  .p2align  2
+GCC_except_table0:
+.Lexception0:
+  .byte 255                     @ @LPStart Encoding = omit
+  .byte  0x90                    @ @TType Encoding = indirect | pcrel
+  .asciz  "\257\200"              @ @TType base offset
+  .byte 3                       @ Call site Encoding = udata4
+  .byte 39                      @ Call site table length
+  .long  .Lfunc_begin0-.Lfunc_begin0 @ >> Call Site 1 <<
+  .long  .Ltmp0-.Lfunc_begin0    @   Call between .Lfunc_begin0 and .Ltmp0
+  .long 0                       @     has no landing pad
+  .byte 0                       @   On action: cleanup
+  .long  .Ltmp0-.Lfunc_begin0    @ >> Call Site 2 <<
+  .long  .Ltmp1-.Ltmp0           @   Call between .Ltmp0 and .Ltmp1
+  .long .Ltmp2-.Lfunc_begin0    @     jumps to .Ltmp2
+  .byte 1                       @   On action: 1
+  .long  .Ltmp1-.Lfunc_begin0    @ >> Call Site 3 <<
+  .long  .Lfunc_end0-.Ltmp1      @   Call between .Ltmp1 and .Lfunc_end0
+  .long 0                       @     has no landing pad
+  .byte 0                       @   On action: cleanup
+  .byte  1                       @ >> Action Record 1 <<
+                                        @   Catch TypeInfo 1
+  .byte 0                       @   No further actions
+                                        @ >> Catch TypeInfos <<
+  .long _ZTIi(target2)          @ TypeInfo 1
+  .p2align  2
+  .fnend