https://gcc.gnu.org/g:440f657945f8114ebdc78dbc93c611d651d7bef0
commit 440f657945f8114ebdc78dbc93c611d651d7bef0 Author: Michael Meissner <[email protected]> Date: Tue Oct 20 21:27:21 2026 -0400 Update BF to V4SF comment. 2026-10-20 Michael Meissner <[email protected]> * config/rs6000/float16.cc: Update comment about the advantage of using V4SFmode vector operations to do arithmetic on BFmode scalar variables. Diff: --- gcc/config/rs6000/float16.cc | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/gcc/config/rs6000/float16.cc b/gcc/config/rs6000/float16.cc index c20925bdad98..5064a4611133 100644 --- a/gcc/config/rs6000/float16.cc +++ b/gcc/config/rs6000/float16.cc @@ -263,7 +263,11 @@ fp16_vectorization (enum rtx_code icode, xvcvspbf16 0,0 stxsihx 0,0,3 - */ + For non-vectorized power10 code, I measured that using the V4SFmode vector + instructions were 20-45 times faster than using the scalar conversions and + scalar SFmode instructions. The HFmode code which has the appropriate + scalar conversions operates at roughly the same speed as the BFmode code + using the V4SFmode vector instructions. */ /* Stats for whole expression. */
