[PATCH] D151782: Improve WebAssembly vector bitmask, mask reduction, and extending

[Markus Everling via Phabricator via llvm-commits]

Sp00ph added inline comments.


================
Comment at: llvm/test/CodeGen/WebAssembly/simd-vecreduce-bool.ll:27-30
+; CHECK-NEXT:    i32.const $push0=, 15
+; CHECK-NEXT:    i16x8.shl $push1=, $0, $pop0
+; CHECK-NEXT:    i32.const $push4=, 15
+; CHECK-NEXT:    i16x8.shr_s $push2=, $pop1, $pop4
----------------
calebzulawski wrote:
> tlively wrote:
> > Could we skip this sign extension before the `v128.any_true`? As long as the high bits of each input lane are zeroed (which I believe they should be), the sign extension doesn't affect the outcome of `v128.any_true`.
> I originally thought this as well, but I think the shifts are inserted because the other lanes are undef instead of zero, if that's possible.  I added the case at the end (test_cmp_v16i8) where all bits are known, and the shifts disappear.
The high bits of `<N x i1>` are unspecified, which also previously led to miscompilations on AArch64, see this issue: https://github.com/llvm/llvm-project/issues/62211


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D151782/new/

https://reviews.llvm.org/D151782