[llvm-commits] [llvm-gcc-4.2] r108554 - /llvm-gcc-4.2/trunk/gcc/llvm-convert.cpp

[Bob Wilson]

On Jul 19, 2010, at 10:40 AM, Dale Johannesen wrote:

> 
> On Jul 19, 2010, at 9:20 AMPDT, Bob Wilson wrote:
> 
>> Dale, I don't think this is correct, at least for ARM.  The 'p' constraint is for a memory address, not the actual memory reference.  If I remember correctly, the use of this reported in pr4521 was kind of questionable, and I haven't yet looked in any detail at pr5314 and pr5533.  Maybe we can talk about this in person sometime today.
> 
> It wouldn't surprise me if "p" isn't completely correct yet; inline asm is implemented incrementally as bugs are found in it.   I believe this one is an improvement, but could be wrong.  Do you know of anything it breaks?

There may be some breakage, but it definitely allows things through that will not work for ARM.  I've fixed it for now by adding a target-specific macro (svn 108963).  There's such a wide variation in what different targets allow as a valid memory address that I don't see any other way to handle this.