[PATCH] D66050: Improve division estimation of floating points.

Wed Aug 14 01:19:18 PDT 2019

qiucf added a comment.

In D66050#1624254 <https://reviews.llvm.org/D66050#1624254>, @lebedev.ri wrote:

> Missing test coverage; performance/precision overview - how much of performance do we loose for how much extra precision, comparison with native division?

The patch won't affect the performance by either comparing running time or analyzing it in theory, since actually it just does a re-order for them, producing no more DAG nodes/instructions.

Compared with native division (tested with our own math library on PPC), the maximum error from estimation by GCC is 0.5ulp, while that of Clang is >4ulp.

In D66050#1624257 <https://reviews.llvm.org/D66050#1624257>, @xbolva00 wrote:

> ICC behaviour?

Do you mean the behavior is the same as ICC? I've not tested it, but the patch is to make result as precise as GCC.

Repository:
  rL LLVM

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D66050/new/

https://reviews.llvm.org/D66050