[PATCH] D26855: New unsafe-fp-math implementation for X86 target

[Gerolf Hoflehner via Phabricator via llvm-commits]

Gerolf added a comment.

I'm leaning towards a LGTM since you addressed basically all my issues, but more people mushroomed and are curious about your performance data. So I think can't dodge that question anymore and need to share some good data for the benchmark(s) you are tuning for before you get the nod.

Cheers
Gerolf



================
Comment at: lib/Target/X86/X86InstrInfo.cpp:9523
+///
+///   vmovss  xmm1, dword ptr [rip + .LCPI0_0] # xmm1 = mem[0],zero,zero,zero
+///   vrcpss  xmm2, xmm0, xmm0
----------------
avt77 wrote:
> RKSimon wrote:
> > avt77 wrote:
> > > Gerolf wrote:
> > > > input to this function are 7  parameters, the comment only lists 6.
> > > I did not understand this comment: what should I do here?
> > I think it means that while ArrayRef<int> Instrs have 7 instructions listed, the codegen in the comment only shows 6 instructions
> But in fact all 7 instructions are shown but from index 0 to index 6 (maybe in "strange" order: 0,2,1,3,4,5,6). If you'd like I could change the order and/or start numbering from 1. Gerolf, should I do it or we fixed everything?
In the comments I only count 6 instructions from vmovss to vaddss, but the code handles 7. What am I missing?


https://reviews.llvm.org/D26855