[PATCH] D84556: [WebAssembly] Remove intrinsics for SIMD widening ops

Tue Jul 28 19:59:57 PDT 2020

tlively added inline comments.

================
Comment at: llvm/test/CodeGen/WebAssembly/simd-widening.ll:113
+
+;; Also test that similar patterns are still expanded correctly
+
----------------
aheejin wrote:
> tlively wrote:
> > aheejin wrote:
> > > It'd be clearer to say starting indices of these don't start with 0 or [lanecount - 1] so they can't be widened using `widen_low` or `widen_high` instructions.
> > > 
> > > Question: Can we also widen these using shifts?
> > Sure, since I didn't end up testing more patterns, I can make the comment more specific.
> > 
> > Regarding shifts, I don't think it's possible to do widening with shifts because widening has to fundamentally change the number of lanes, which shifts can't do.
> What I meant was, in case of i16x8->i32x4, the current code can widen i16x8 input vector with elements in the indices 0 to 4. If those elements are instead in 1 to 5, can we first shift that to 0~4 and widen it?
Oh gotcha. No, unfortunately I don't think that would work. The SIMD shift instructions shift bytes within lanes but they can't shift data into a different lane. Even if we used shifts on larger lanes to try to overcome that limitation, a 64x2 shift would still not be able to shift data from the high half of the vector to the low half or vice versa, which we would need to do to implement your suggestion.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D84556/new/

https://reviews.llvm.org/D84556