CSE optimization confuses -0.0 with 0.0

[vicuna]

Original bug ID: 7042
Reporter: @xavierleroy
Assigned to: @xavierleroy
Status: closed (set by @xavierleroy on 2017-02-16T14:16:46Z)
Resolution: fixed
Priority: normal
Severity: major
Platform: x86-64
Version: 4.03.0+dev / +beta1
Target version: 4.03.0+dev / +beta1
Fixed in version: 4.03.0+dev / +beta1
Category: back end (clambda to assembly)
Related to: #6442

Bug description

Owing to the use of Pervasives.compare, the new CSE pass of 4.03 considers as equal the instructions "Iconst_float 0.0" and "Iconst_float (-0.0)", replacing the latter by a move of the result of the former. This is not correct, as demonstrated by the attached repro case.

Possible fixes:

• use a specific comparison function, as done for #6442
• at some point in the back-end (C--? Mach?), use int64 instead of float to represent FP constants.

Steps to reproduce

Compile mzero.ml using ocamlc and using ocamlopt, then compare the outputs.

File attachments

• mzero.ml

[vicuna]

Comment author: @mshinwell

I think Flambda may suffer from a similar problem; in fact, we already have a code review comment about it. Maybe we should consider a different representation for floats even before Cmm.

https://github.com/chambart/ocaml-1/blob/flambda_trunk/middle_end/simplify_common.mli

[vicuna]

Comment author: @xavierleroy

Fixed by commit [trunk 6cd8656].