ahatanak added inline comments.
================
Comment at: clang/lib/CodeGen/Address.h:67
return;
- // Currently the max supported alignment is much less than 1 << 63 and is
+ // Currently the max supported alignment is much less than 1 << 32 and is
// guaranteed to be a power of 2, so we can store the log of the alignment
----------------
aeubanks wrote:
> ahatanak wrote:
> > ahatanak wrote:
> > > efriedma wrote:
> > > > ahatanak wrote:
> > > > > efriedma wrote:
> > > > > > ahatanak wrote:
> > > > > > > efriedma wrote:
> > > > > > > > This comment isn't right. The max alignment is, as far as I
> > > > > > > > can tell, 1<<32 exactly. (But there's something weird going on
> > > > > > > > with very large values... somehow `int a[1LL<<32]
> > > > > > > > __attribute((aligned(1ULL<<32))) = {};` ignores the alignment.)
> > > > > > > The following function generated by tablegen (and a few others
> > > > > > > directly or indirectly calling the function) returns a 32-bit
> > > > > > > int, but it should be returning a 64-bit int.
> > > > > > >
> > > > > > > https://github.com/llvm/llvm-project/blob/main/clang/utils/TableGen/ClangAttrEmitter.cpp#L532
> > > > > > Filed https://github.com/llvm/llvm-project/issues/60752 so we don't
> > > > > > lose track of this.
> > > > > I just realized we can't reduce the number of bits used for alignment
> > > > > here as we need 6 bits for alignment of `1 << 32`.
> > > > >
> > > > > Should we allocate additional memory when `AlignLog` is either 31 or
> > > > > 32? If the 5-bit alignment is equal to `0b11111`, it would mean that
> > > > > there is an out-of-line storage large enough to hold the alignment
> > > > > and any other extra information that is needed. I think
> > > > > https://reviews.llvm.org/D117262#3267899 proposes a similar idea.
> > > > How much does `sizeof(Address)` actually matter, anyway? If it's going
> > > > to get that nasty to implement the packing, I'm not sure it's worth the
> > > > effort to optimize.
> > > I'm not sure, but apparently https://reviews.llvm.org/D117262 was needed
> > > to reduce the memory usage.
> > >
> > > I don't see a significant increase in stack usage (a little over 1%) in
> > > the files in `clang/lib/CodeGen` when I build clang with `-fstack-usage`.
> > @aeubanks do we have to use the specialization `AddressImpl<T, true>`?
> >
> > https://github.com/llvm/llvm-project/blob/main/clang/lib/CodeGen/Address.h#L31
> looking back at the internal bug motivating the patch, it was because some we
> had to lower some thresholds in a stress test, so that in it of itself isn't
> super important. but `Address` is used a lot and it's nice to keep resource
> usage down (1% isn't trivial). does this noticeably affect max RSS? could run
> a potential patch through llvm-compile-time-tracker (I can help with that) to
> see the impact
Note that 1% is the increase in the files in `clang/lib/CodeGen`. The files in
other directories (e.g., `clang/lib/Sema`) aren't affected by the change as
they don't use `Address`.
I see an increase of 0.33% in max RSS when I run the tests in
`clang/test/CodeGen*`, but I'm not sure how reliable that is.
How do you run the potential patch through `llvm-compile-time-tracker`? The
patch just changes the type to `AddressImpl<void, false> A;` on this line:
https://github.com/llvm/llvm-project/blob/main/clang/lib/CodeGen/Address.h#L92
Repository:
rG LLVM Github Monorepo
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D142584/new/
https://reviews.llvm.org/D142584
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
