Hi, On Thu, 13 Apr 2006 13:13:11 -0700 "Om Narasimhan" <om.turyx@xxxxxxxxx> wrote: > >From linux-2.6.16/include/asm-i386/semaphore.h ( > http://lxr.linux.no/ident?v=2.6.10;i=down ) > 105 static inline void down(struct semaphore * sem) > 106 { > might_sleep(); > __asm__ __volatile__( > "# atomic down operation\n\t" > LOCK "decl %0\n\t" /* --sem->count */ // locked and > decremented. "js 2f\n" // if sign is set jump to 2: > "1:\n" > LOCK_SECTION_START("") // starts a new section in elf, what > does this mean? > "2:\tlea %0,%%eax\n\t" // loads &sem (==&sem->count) into eax. > "call __down_failed\n\t" // this eventually returns when sem > is acquired 116 "jmp 1b\n" // goes back to 1: ... how would > this function ever return? > 117 LOCK_SECTION_END > :"=m" (sem->count) > : > :"memory","ax"); > } Ok, this is quite subtle. The explanation is in the LOCK_SECTION_START() and LOCK_SECTION_END() macros. These macros puts the code in-between in another section, far away from the current. So, in fact, after the 1:, what you have in memory, is not the "lea", but the next instruction after the down(). So, if you write some C code like : a = 2; down (); c = 1; It will end up like this : - some assembly code to set a to 2 - LOCK decl %0 - js 2f /* Jump only in the contended case */ - some assembly code to set c to 1 and then, far away : - lea %0, %%eax - call __down_failed The idea is to optimize the non-contended case (when there's no contention on the semaphore). The idea is to not trash the i-cache with instructions that are useless most of the time, and probably to optimize the pipeline usage by making sure that prefetched instructions are the one that are most likely to be executed. As this is a frequently asked questions, I think I'll contribute an entry to the FAQ, now that the website is a Wiki. Sincerly, Thomas -- PETAZZONI Thomas - thomas.petazzoni@xxxxxxxx http://{thomas,sos,kos}.enix.org - Jabber: thomas.petazzoni@xxxxxxxxx http://{agenda,livret}dulibre.org - http://www.toulibre.org Fingerprint : 0BE1 4CF3 CEA4 AC9D CC6E 1624 F653 CB30 98D3 F7A7
Attachment:
signature.asc
Description: PGP signature
