Linux/ARM: Fix +3000 bus errors of unit-test in case of O2/O3 levels

**PROBLEM**
This patch is to resolve +3000 bus errors that are generated whenever we use
the aggressive optimization levels of clang (issue #5844 ).

When we enable the -O3 optimization level of the clang version(from clang 3.5
to clang 3.9(snapshot)), we have always got the lots of bus errors from the
coreCLR's unit tests. Actually, we can easily monitor SIGBUS signals (e.g.,
"misaligned memory access") with /proc/cpu/alignment facility in kernel space.
Using "echo 2 > /proc/cpu/alignment" makes Linux kernel fixes the problems
but the performance of the application will be degraded.
.source: http://lxr.free-electrons.com/source/Documentation/arm/mem_alignment

**VERSION 2**
.Add UNALIGNED_ARM macro for handling ARM core specific optimisation levels.
.Add RISC-based ARM core handling into the existing infra-structure of the
platform adaptation layer (PAL) for aggressive optimization cases on Linux/ARM.

**VERSION 1**
Basically, RISC-based ARM architecture requires aligned access with 4byte reads.
In order to support aggressive optimization levels such as O2/O3, let's use
attribute keyword of aligned(1) instead of using memcpy(2) in into
a properly aligned buffer or the packing attribute.

**BACKGROUND**
According to ARM information center(infocenter.arm.com), By default,
the ARM compiler expects normal C and C++ pointers to point
to an aligned word in memory. A type qualifier __packed is provided to
enable unaligned pointer access. If you want to define a pointer to a word
that can be at any address (that is, that can be at a non-natural alignment),
you must specify this using the __packed qualifier when defining the pointer:
__packed int *pi; // pointer to unaligned int

However, clang/llvm does not support the __packed qualifier such as
__attribute__((packed)) or __attribute__((packed, aligned(4)))

In -O0 (debug build) the innermost block is emitted as the following assembly,
which works properly:
 ldr r1, [r0, dotnet#24]
 ldr r2, [r0, dotnet#20]

In -O3 (release build) however the compiler realizes these fields are adjacent
and generates this assembly:
 ldrdeq  r2, r3, [r0, dotnet#20]
Unfortunately, vldr/ldrdb instructions always generate an alignment fault
(in practice). It seems that clang uses ldrb instruction although GCC uses
ldr instruction because armv7 supports unaligned ldr instruction.

Note: If some arm architectures (e.g., Linux/ARM Emulator) does not support
unaligned ldr, this issue is not generated with aggressive optimization
levels (e.g., -O2 and -O3).

* source: How does the ARM Compiler support unaligned accesses?
  http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.faqs/ka15414.html

Signed-off-by: Geunsik Lim <geunsik.lim@samsung.com>

Author: leemgs

src/jit/compiler.hpp

@@ -757,13 +757,16 @@ inline
   signed __int16    getI2LittleEndian(const BYTE * ptr)
 { return * (UNALIGNED signed __int16 *)ptr; }
 
+typedef signed __int32 UNALIGNED_ARM __unaligned_int32;
+typedef signed __int64 UNALIGNED_ARM __unaligned_int64;
+
 inline
   signed __int32    getI4LittleEndian(const BYTE * ptr)
-{ return *(UNALIGNED signed __int32*)ptr; }
+{ return *(__unaligned_int32 *)ptr; }
 
 inline
   signed __int64    getI8LittleEndian(const BYTE * ptr)
-{ return *(UNALIGNED signed __int64*)ptr; }
+{ return *(__unaligned_int64 *)ptr; }
 
 inline
 float               getR4LittleEndian(const BYTE * ptr)

src/pal/inc/pal.h

@@ -201,7 +201,13 @@ extern "C" {
 
 #else // _MSC_VER
 
+#if defined (_M_ARM) && !defined (DEBUG)
 #define UNALIGNED
+#define UNALIGNED_ARM     __attribute__((aligned(1))) //for -O2 and -O3
+#else
+#define UNALIGNED
+#define UNALIGNED_ARM  //for -O0
+#endif
 
 #endif // _MSC_VER