[PATCH] D60618: [ARM] Add bitcast/extract_subvec. of fp16 vectors

[Oliver Stannard (Linaro) via Phabricator via llvm-commits]

ostannard added a comment.

Thanks for fixing the order of the patterns, it's much easier to check for completeness now.



================
Comment at: lib/Target/ARM/ARMInstrNEON.td:7223
+  def : Pat<(v8f16 (bitconvert (v2f64 QPR:$src))), (v8f16 QPR:$src)>;
+  def : Pat<(v8f16 (bitconvert (v2f64 QPR:$src))), (v8f16 QPR:$src)>;
+  def : Pat<(v8f16 (bitconvert (v2i64 QPR:$src))), (v8f16 QPR:$src)>;
----------------
Duplicate of the line above.


================
Comment at: test/CodeGen/ARM/fp16-vector-basic-instructions.ll:1
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -o - -mtriple=arm-eabi -mattr=+v8.2a,+fullfp16,+neon  -float-abi=hard -target-abi=apcs-gnu %s | FileCheck %s --check-prefix=HARD
----------------
The expected output for this file is much more complicated than it needs to be, and depends on lots of things other than the changes you've made here. The shuffles in particular can generate a lot of extra code which we aren't trying to test here. Something like big-endian-neon-bitconv.ll would be better, using a similar pattern of code to test each type.


Repository:
  rL LLVM

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D60618/new/

https://reviews.llvm.org/D60618