[cfe-dev] size of pointer and size of long

Eli Friedman eli.friedman at gmail.com
Tue Oct 21 13:36:20 PDT 2008


On Tue, Oct 21, 2008 at 10:38 AM,  <Alireza.Moshtaghi at microchip.com> wrote:
> In the recent changes to the preprocessor, new assumptions are added
> that assume the size of pointer to be same as size of long (or size of
> long long for 64 bit targets)
> Although this is to mimic gcc behavior, not all ports necessarily want
> to follow suite.

Definitely... that isn't intended to be a permanent restriction.  In
addition to 16-bit targets, Win64 breaks that assumption.

> Our target does not support long long and assumes 32 bit long and 16 bit
> pointer.

Okay, good to know; so for PIC16, the correct specification is that
everything has alignment of 8 bits, int and pointer have size 16 bits,
long has size 32 bits, and long long is illegal?  While I'm asking, is
there any floating-point support?

> There is an assertion in the Preprocessor.cpp line 600 that assumes
> current targets have sizeof(long) == sizeof(void *). This assumption is
> incorrect.

Yeah; the only reason Chris did that was to avoid adding a few extra
cases to the code.  Feel free to submit a patch, or else I'm sure
Chris will get to it soon.

> Are these assumptions only in the preprocessor? Or there are other
> places in clang that these assumptions are also made?

There are a few places scattered through the code making bad
assumptions about types because target data didn't have sufficient
information, although not many; we've been pretty good about making
Sema match the actual C type model and not making assumptions about
type sizes, so most places are fine.

The places I can think of easily that have issues are
AST::getSizeType() and friends, and HandleModeAttr in
SemaDeclAttr.cpp, although it's possible there are others.

Targets really need to have a way to explicitly define the types of
intmax_t, size_t, ptrdiff_t, and wchar_t.  In addition to it being
messy to derive these types from the widths of the integer/pointer
types, it's actually impossible: for example, size_t should be
"unsigned int" on Linux and "unsigned long" on OSX.  Here's a quick
proposal: we add an enum IntType {UnsignedShort, SignedShort,
UnsignedInt, SignedInt, UnsignedLong, SignedLong, UnsignedLongLong,
SignedLongLong} to TargetInfo.  We add 4 new methods to TargetInfo,
SizeType, IntMaxType, PtrDiffType, and WCharType, returning values
from the enum.  The preprocessor and ASTContext then key off of these
methods for the relevant types rather than trying to guess.

For HandleModeAttr, I'm not really sure what gcc does; could someone
check the output of the following on OSX 32-bit?

typedef long a __attribute__((__mode__(__SI__)));
int x[__builtin_types_compatible_p(a, long) ? 1 : -1];

-Eli



More information about the cfe-dev mailing list