Optimization versus broken versions of Intel compiler

[werdna]

This is a follow up to a discussion in the email list last year, which started at ​http://lists.osgeo.org/pipermail/gdal-dev/2015-November/043025.html (some emails, including mine, in the discussion have not made it onto the list).

I have access to a machine which has many versions and subversions of the intel compiler from v9 to v15. Compiler versions 9.1.043, 9.1.046, 10.0.017, 10.0.026 and 10.1.015 fail before attempting to compile thinplatespline.cpp (on gdalmediancut.cpp). Versions 10.1.022 and all the v11 and v13 versions I have (11.0.074, 11.0.081, 11.0.083, 1 1.1.046, 11.1.056, 11.1.059, 11.1.064, 11.1.069, 11.1.072 and 11.1.073 13.0.0.079, 13.0.1. 117, 13.1.0.146 and 13.1.2.183) fail to compile thinplatespline.cpp. Compilation succeeds with all version 12 and 15 versions I have (12.0.0.084, 12.0.1.107, 1 2.0.2.137, 12.0.4.191, 12.0.5.220, 12.0.7.256, 12.1.10.319, 12.1.11.339, 12.1.8.273, 12.1. 9.293 and 15.0.0.090, 15.0.1.133, 15.0.5.223). Version 14.0.3.174 succeeds, but all other v14 versions (all older 14.0.0.080, 14.0.1.106 and 14.0.2.144) I have access to fail.

There are two compiler predefined macros related to the intel compiler version: INTEL_CO MPILER and INTEL_COMPILER_BUILD_DATE . The Intel compilers also also define GNUC and x86_64 .

Version	INTEL_COMPILER	INTEL_COMPILER_BUILD_DATE	OK?
9.1.043	910	20060818	gdalmediancut.lo fails
9.1.046	910	20070109	gdalmediancut.lo fails
10.0.017	1000	20070307	gdalmediancut.lo fails
10.0.026	1000	20070809	gdalmediancut.lo fails
10.1.015	1010	20080312	gdalmediancut.lo fails
10.1.022	1010	20090203	thinplatespline.lo fails
11.0.074	1100	20081105	thinplatespline.lo fails
11.0.081	1100	20090131	thinplatespline.lo fails
11.0.083	1100	20090318	thinplatespline.lo fails
11.1.046	1110	20090630	thinplatespline.lo fails
11.1.056	1110	20090827	thinplatespline.lo fails
11.1.059	1110	20091012	thinplatespline.lo fails
11.1.064	1110	20091130	thinplatespline.lo fails
11.1.069	1110	20100203	thinplatespline.lo fails
11.1.072	1110	20100414	thinplatespline.lo fails
11.1.073	1110	20100806	thinplatespline.lo fails
--	--	--	--
12.0.0.084	1200	20101006	compiles
12.0.1.107	1200	20101116	compiles
12.0.2.137	1200	20110112	compiles
12.0.4.191	1200	20110427	compiles
12.0.5.220	1200	20110719	compiles
12.0.7.256	1210	20111011	compiles
12.1.10.319	1210	20120410	compiles
12.1.11.339	1210	20120612	compiles
12.1.8.273	1210	20111128	compiles
12.1.9.293	1210	20120212	compiles
--	--	--	--
13.0.0.079	1300	20120731	thinplatespline.lo fails
13.0.1.117	1300	20121010	thinplatespline.lo fails
13.1.0.146	1310	20130121	thinplatespline.lo fails
13.1.2.183	1310	20130514	thinplatespline.lo fails
14.0.0.080	1400	20130728	thinplatespline.lo fails
14.0.1.106	1400	20131008	thinplatespline.lo fails
14.0.2.144	1400	20140120	thinplatespline.lo fails
--	--	--	--
14.0.3.174	1400	20140422	compiles
15.0.0.090	1500	20140723	compiles
15.0.1.133	1500	20141023	compiles
15.0.5.223	1500	20150805	compiles

The attached patch enables all my versions of the compiler to compile thinplatespline.c, and uses the optimization where posible - except for v14.0.3.174 where it uses the unoptimized version (this could be fixed with further work if necessary).

[werdna]

Proposed patch

[Even Rouault]

I've just changed the name of the #ifdef to avoid reserved names (leading double underscore)

trunk r33364, branches/2.0 r33365 "thinplatespline.cpp: avoid using optimized version of VizGeorefSpline2DBase_func4() with ICC versions that fail on it (#6350)"