On Thu, Feb 4, 2021 at 11:18 PM Alexandre Oliva <ol...@adacore.com> wrote:
>
> On Feb 4, 2021, Richard Biener <richard.guent...@gmail.com> wrote:
>
> >> > b) if expansion would use BY_PIECES then expand to an unrolled loop
> >>
> >> Why would that be better than keeping the constant-length memset call,
> >> that would be turned into an unrolled loop during expand?
>
> > Well, because of the possibly lost ctz and alignment info.
>
> Funny you should mention that. I got started with the expand-time
> expansion yesterday, and found out that we're not using the alignment
> information that is available. Though the pointer is known to point to
> an aligned object, we are going for 8-bit alignment for some reason.
>
> The strategy I used there was to first check whether by_pieces would
> expand inline a constant length near the max known length, then loop
> over the bits in the variable length, expand in each iteration a
> constant-length store-by-pieces for the fixed length corresponding to
> that bit, and a test comparing the variable length with the fixed length
> guarding the expansion of the store-by-pieces. We may get larger code
> this way (no loops), but only O(log(len)) compares.
>
> I've also fixed some bugs in the ldist expander, so now it bootstraps,
> but with a few regressions in the testsuite, that I'm yet to look into.
>
> >> Uhh, thanks, but... you realize nearly all of the gimple-building code
> >> is one and the same for the loop and for trailing count misalignment?
>
> > Sorry, the code lacked comments and so I didn't actually try decipering
> > the code you generate ;)
>
> Oh, come on, it was planly obscure ;-D
>
> Sorry for posting an early-draft before polishing it up.
>
> > The original motivation was really that esp. for small trip count loops
> > the target knows best how to implement them. Now, that completely
> > fails of course in case the target doesn't implement any of this or
> > the generic code fails because we lost ctz and alignment info.
>
> In our case, generic code fails because it won't handle variable-sized
> clear-by-pieces. But then, I found out, when it's fixed-size, it also
> makes the code worse, because it seems to expand to byte stores even
> when the store-to object is known to have wider alignment:
>
> union u {
> long long i;
> char c[8];
> } x[8];
> int s(union u *p, int k) {
> for (int i = k ? 0 : 3; i < 8; i++) {
> for (int j = 0; j < 8; j++) {
> p[i].c[j] = 0;
> } // becomes a memset to an 8-byte-aligned 8-byte object, then 8 byte
> stores
> }
> }
On x86_64 I see two DImode stores generated, but that appears to be
done via setmem.
> >> > I think the builtins with alignment and calloc-style element count
> >> > will be useful on its own.
> >>
> >> Oh, I see, you're suggesting actual separate builtin functions. Uhh...
> >> I'm not sure I want to go there. I'd much rather recover the ctz of the
> >> length, and use it in existing code.
>
> > Yeah, but when we generate memcpy there might not be a way to
> > store the ctz info until RTL expansion where the magic should really happen
> > ...
>
> True. It can be recovered without much difficulty in the cases I've
> looked at, but it could be lost in others.
>
> > So I'd say go for improving RTL expansion.
>
> 'k, thanks
>
> --
> Alexandre Oliva, happy hacker https://FSFLA.org/blogs/lxo/
> Free Software Activist GNU Toolchain Engineer
> Vim, Vi, Voltei pro Emacs -- GNUlius Caesar
