fix Issue 17237 - Wrong alignment for 256-bit vectors. #6582
Conversation
|
|
Auto-merge toggled on |
| if ((global.params.is64bit || global.params.isOSX) && (size == 16 || size == 32)) | ||
| return size; | ||
|
|
||
| return (8 < size) ? 8 : size; |
There was a problem hiding this comment.
@ibuclaw: What's the reason for this upper bound of 8/32 bytes (why not allow 64 for cache-line aligned fields etc.)? And I find the name size for actual alignment rather confusing.
There was a problem hiding this comment.
Meh, I don't use this code. It doesn't make sense to me either. You shouldn't be using this code either.
There was a problem hiding this comment.
As for max field alignment, it should be 4 bytes for x86, 16 bytes for 64bit, excluding vectors and x87 types.
There was a problem hiding this comment.
Your backend should handle this for you, however, but I won't feel sorry for you if it doesn't. :D
There was a problem hiding this comment.
Lol, thanks for the hint (LLVM obviously does), I've just seen that I messed up a front-end merge, so that we now have 2 We don't use it either. ;)Target::alignsize() implementations in LDC (D and C++), don't know which one gets used. ;)
Relaxed the condition to allow 256-bit vector alignment, but I won't change the logic of anything else incase it affects something, you can do this in a much better way, but it's not my place to say.
Moved alignment adjusting code to
Target::fieldalign, as the frontend should adhere to the result returned from this function, unless some explicit alignment was requested, and not overwrite it in frontend-only code.