編譯指示SWP，ivdep，在不同的編譯器

Question

在良好的Altix手動支持預取（altix_single_cpu_optimization http://wwwuser.gwdg.de/~parallel/parallelrechner/altix_documentation/altix_single_cpu_optimization.pdf）我發現的編譯指示，暗示此列表優化的編譯器

#pragma ivdep    // no aliasing 
#pragma swp     // try to software-pipeline 
#pragma noswp    // disable software-pipelining 
#pragma loop count (NN)  // hint for SWP 
#pragma distribute point // split this large loop 
#pragma unroll (n)   // unroll n times 
#pragma nounroll   // do not unroll 
#pragma prefetch a   // prefetch array 「a」 
#pragma noprefetch c  // do not prefetch array 「c」 

什麼編譯器不支持這些編譯？什麼是實現鏈接優化（軟件流水線是低ipc和低reg計數處理器相當困難的優化之一）？

Answer 1

的Sun Studio 12 http://docs.oracle.com/cd/E19205-01/819-5265/bjaby/index.html

#pragma nomemorydepend 
#pragma no_side_effect (funcs) 
#pragma opt N (funcs) 
#pragma pipeloop(n)   // like swp, m.b. noswp 
#pragma rarely_called(funcs) 
#pragma unroll (N) 

英特爾http://software.intel.com/en-us/articles/use-pragmas-with-the-intel-c-compiler-for-linux-on-64-bit-architecture/

#pragma swp 
#pragma noswp 
#pragma unroll 
#pragma distribute point 
#pragma prefetch *arrays* 
#pragma loop count 
#pragma ivdep 

IBM http://publib.boulder.ibm.com/infocenter/comphelp/v7v91/topic/com.ibm.vacpp7a.doc/compiler/ref/rupragen.htm

#pragma block_loop 
#pragma disjoint 
#pragma execution_frequency 
#pragma isolated_call 
#pragma stream_unroll 
#pragma unroll 
#pragma unrollandfuse 

Open64 http://wiki.open64.net/index.php/Pragmas_Support_in_Open64（有一堆或鏈接到其他編譯器）

#pragma unroll 
#pragma blockable 
#pragma prefetch 
#pragma ivdep 

PGI 9 http://www.math.hkbu.edu.hk/parallel/pgi/doc/pgiws_ug/pgi32u10.htm

#pragma depchk 
#pragma opt 
#pragma unroll 
#pragma vector 

MSvs2003 http://msdn.microsoft.com/en-us/library/d9x1s805(VS.71).aspx

#pragma optimize 

HP http://www.docs.hp.com/en/B3901-90024/ch03s06.html

#pragma if_convert 
#pragma rarely_called 
#pragma unroll // or unroll factor 

http://docs.cray.com/books/004-2179-001/html-004-2179-001/xgaslrwh.html CRAY

#pragma _CRI ivdep 
#pragma _CRI opt // or noopt 
#pragma _CRI novector // novsearch 
#pragma _CRI shortloop 
#pragma _CRI prefervector 
#pragma _CRI split 
#pragma _CRI unroll [n]