[PATCH] D78756: [SveEmitter] Add builtins for svreinterpret
Eli Friedman via Phabricator via cfe-commits
cfe-commits at lists.llvm.org
Wed Apr 29 14:33:50 PDT 2020
efriedma added inline comments.
================
Comment at: clang/lib/CodeGen/CGBuiltin.cpp:7880
+ return Builder.CreateBitCast(Val, Ty);
+ }
+
----------------
sdesmalen wrote:
> efriedma wrote:
> > sdesmalen wrote:
> > > efriedma wrote:
> > > > I'm vaguely suspicious this might be wrong for big-endian targets. I mean, this isn't unreasonable, but users might be surprised if svreinterpret isn't a no-op.
> > > For SVE the loads and stores (svld1 and svst1) are all endian safe, so no special consideration needs to be taken for big endian targets.
> > >
> > > The ACLE specifies that:
> > > > The svreinterpret functions simply reinterpret a vector of one type as a vector of another type, without changing any of the bits.
> > "bitcast" is specified to mean "reinterpret the bits like a store+load". On big-endian NEON (and, I assume, SVE), that isn't a no-op. See http://llvm.org/docs/BigEndianNEON.html .
> >
> > I mean, if the answer here is "yes, svreinterpret is supposed to lower to a REV", then that's fine. But I'd like to see some explciit acknowledgement that that's intentional.
> Thanks for pointing out that page, but for SVE I don't think the svreinterpret should lower to a REV.
>
> This is probably where things are different from Neon. The ACLE SVE vectors such as `svint32_t` are opaque vector types and the only way to load/store them from/to memory is through the use of the svld1 and svst1 intrinsics which are endian safe (in that they use the ld1/st1 instructions that do endianess conversion on big endian targets). The ACLE does not expose any full-vector load/store (ldr/str) operations.
Like that page describes, we use ld1/st1 for big-endian NEON, to match the LLVM IR rules for laying out a vector. If you use ld1/st1 to load/store vectors on big-endian NEON, a bitcast is not a no-op. As far as I know, SVE ld1/st1 is equivalent to NEON ld1/st1 in the case where vscale=1. Therefore, on big-endian SVE, a bitcast is not a no-op.
That leaves the following options:
1. svreinterpret is not a no-op.
2. svreinterpret is not equivalent to an LLVM IR bitcast, so this patch needs to be changed.
Repository:
rG LLVM Github Monorepo
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D78756/new/
https://reviews.llvm.org/D78756
More information about the cfe-commits
mailing list