urnathan wrote:
@rjmccall thanks for your comments. Here's a reimplementation using a single
loop -- the downside is that we may end up repeating the initial grouping scan,
if we search more than 1 Access Unit ahead and fail to find a 'good' access
unit. This update changes the algorithm slightly, as I realized that `struct A
{ short a: 16; char b:8; int c;};` and `struct B { char b: 8; short a:16; int
c;};` should have the same (32-bit) access unit. Originally only `A` got that.
1) I noticed that several `CGRecordLowering` member fns could either be `const`
or `static` -- would you like that as a separate NFC PR?
2) I have corrected the error about merging across zero-sized members.
3) It may not be obvious from GH, but this PR breaks down to 3 (or 4) separate
commits:
a) A set of test cases, marked up for the current scheme
b) A new Target hook indicating whether unaligned accesses are cheap
c) [optional] the CGRecordLowering change just mentioned
d) the algorithm change, which updates those tests and makes it easier to
see how the behaviour changes.
Do you want that commit structure maintained?
https://github.com/llvm/llvm-project/pull/65742
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
