On Tue, Feb 18, 2020 at 03:27:39AM -0600, Xionghu Luo wrote:
> Store-merging pass should run twice, the reason is pass fre/pre will do
> some kind of optimizations to instructions by:
> 1. Converting the load from address to load from function arguments
> (store_merging_30.c:foo1).
> 2. Converting the byte access to BIT_FIELD_REF(store_merging_30.c:foo2).
> 3. Other bitfield combinations or potential interference optimizations etc.
> These optimizations will break the store chain, store-merging pass fails
> to catch such kind of pattern so stores are not merged in middle end,
> then consecutive stb/sth instructions(should be merged to stw) are emitted
> finally.
This seems to be way too early, won't it e.g. stand in a way for SRA because
it will make harder to break those appart again? Or anything that uses
alias info and store-merging needs to use a conservative alias set?
For constant stores into bitfields followed by loads I'm working on something
ATM (PR93582).
Jakub
