With have important issues with ICC-17, and no time to debug those
issues. Let's postpone.
Anyway, there is probably a bug with this compiler version. I have
filled at least one issue, that has escalated to Intel developers:
https://businessportal.intel.com/irj/portal/MyIssue2?issueNum=6000161285
The test-case is:
https://gist.github.com/lrineau/59da15e8a02958474aa74ef1cc1b1fb2
The bug I have filled is:
```
The attached file is not compiled correctly, and the run of it fails. Here is the command lines I have used, and the results:
icpc -V -O1 -o /tmp/test_bug test_bug.cpp && /tmp/test_bug
Intel(R) C++ Intel(R) 64 Compiler for applications running on Intel(R) 64, Version 17.0.0.064 Beta Build 20160517
Copyright (C) 1985-2016 Intel Corporation. All rights reserved.
icpc: NOTE: The Beta evaluation period for this product ends on 7-oct-2016 UTC.
Edison Design Group C/C++ Front End, version 4.10.1 (May 18 2016 01:11:31)
Copyright 1988-2015 Edison Design Group, Inc.
GNU ld version 2.25-17.fc23
(Construction from 0x1a3bc90)
(Construction from 0x1a3bcb8)
(Copy-construction from 0x1a3bc90)
operator++(int)
before: 0x1a3bc90 : 1
(Copy-construction from 0x1a3bc90)
operator++()
before: 0x1a3bc90 : 1
after: 0x1a3bc98 : 2
after: 0x1a3bc98 : 2
tmp: 0x1a3bc90 : 1
test_bug: test_bug.cpp:79: void test_N_step_adaptor(): Assertion `4 == (*begin)' failed.
If I compile with the flag -O3 instead of -O1, or if I remove the lines with std::cerr, then the code compiles and run correctly. Previous version Intel C++ Compiler 16.0 compiled that code correctly, and all g++ and clang versions as well. That code is part of CGAL, and is tested daily on a lot of platforms.
```
We found out with Jane @jtournois that a testsuite with VLD was a lot
slower than the same testsuite without it. That can explain why picasso
can no longer process 7 or 8 testsuites per day, but only barely 3 or 4.
Picasso will no longer use VLD, and Jane's machine, Sosno, will use it
only once per week, on Saturdays.