[Lldb-commits] [PATCH] Add assertion that byte-swapping is only performed on register-sized copies

[Ed Maste]

On 17 September 2013 14:51, Ed Maste <emaste at freebsd.org> wrote:
> I believe that we should only ever reverse a block that corresponds to (some) register size.

Adding the assert shows that ObjectFile::CopyData is the source of the
reversed output for "memory read" on a big-endian target.  There are
other instances that copy a block of target memory to a buffer
intended to remain in target byte order.  If there are any CopyData
callers that directly fetch a target word (and thus need endian
swapping), they can be changed to use a DataExtractor instead (via
ObjectFile::GetData).

This change fixes the "memory read" issue:

--- a/source/Symbol/ObjectFile.cpp
+++ b/source/Symbol/ObjectFile.cpp
@@ -459,7 +459,8 @@ size_t
 ObjectFile::CopyData (off_t offset, size_t length, void *dst) const
 {
     // The entire file has already been mmap'ed into m_data, so just
copy from there
-    return m_data.CopyByteOrderedData (offset, length, dst, length,
lldb::endian::InlHostByteOrder());
+    // Note that the data remains in target byte order.
+    return m_data.CopyByteOrderedData (offset, length, dst, length,
m_data.GetByteOrder());
}

This change and the assert in review D1701 should be NOPs for the
current x86 and arm little-endian targets.  I'd like to commit these
sorts of changes as I encounter them (even though there are many
remaining endianness issues), both to help Félix Cloutier's work on
the PPC target, and to help solidify expectations on endianness as
other development continues.  Does that seem reasonable?