在 2018-08-30 01:36, Nathan Sidwell 写道:
But, it would be bad to make that particular ABI fix in a point release,
so this patch just reverts the regression I caused. Sadly, because it
requires understanding TEMPLATE_DECL, we can't simply update
place_field. Instead I temporarily stitch out undesired DECLs around
the call to place_field. This seems the least intrusive, and happens
only when ms_bitfield_layout_p is in effect.
It is strictly an ABI break but I doubt whether code in real world that
got broken by this bug ever exists. Usually when people expect
consecutive bitfields to be packed into a single word they wouldn't put
irrelevant declarations between them.
An important reason why such code could be rare is that the following
code compiles differently as C and C++:
```
E:\Desktop>cat test.cc
#include <stdio.h>
struct foo
{
unsigned a : 21;
union x1 { char x; };
unsigned b : 11;
union x1 u;
};
struct bar
{
union x2 { char x; };
unsigned a : 21;
unsigned b : 11;
union x2 u;
};
int main(void)
{
printf("%d\n", (int)sizeof(struct foo));
printf("%d\n", (int)sizeof(struct bar));
}
```
When compiled as C the output is:
```
E:\Desktop>cl /nologo /Tctest.cc /Fe:a.exe && a.exe
test.cc
16
12
```
while as C++ the output is:
```
E:\Desktop>cl /nologo /Tptest.cc /Fe:a.exe && a.exe
test.cc
8
8
```
Basing on the fact that such code is rare I think it should be safe to
trade tolerance (or 'compatibility for this bug') for the correct behavior.
--
Best regards,
LH_Mouse
