Julian Elischer schrieb: > I simplified the output of the preprocessor for a PCPU_SET(xx, newval) > (to look at it). > > I came down to: (after formatting) for i386.. > { > __typeof(((struct pcpu *)0)->pc_xx) __val; > struct __s > { > u_char __b[(((sizeof(__val)) < (4)) ? > (sizeof(__val)) : (4))]; > } __s; > > __val = (newval); /* aligned */ > > if (sizeof(__val) == 1 > || sizeof(__val) == 2 > || sizeof(__val) == 4) { > __s = *(struct __s *)(void *)&__val; > __asm volatile("mov %1,%%fs:%0" : "=m" > (*(struct __s *)(__builtin_offsetof( > struct pcpu, pc_xx))) : "r" (__s)); > } else { > *__extension__ ( > { > __typeof(__val) *__p; > __asm volatile("movl %%fs:%1,%0; > addl %2,%0" : "=r" (__p) : "m" > (*(struct pcpu *)(__builtin_offsetof(struct pcpu, pc_prvspace))), > "i" > (__builtin_offsetof(struct pcpu, pc_xx))); > __p; > }) = __val; > } > } > > having had my brain explode on this several times, > I can't figure out exactly what teh clause after the else is doing. > > anyone better at reading __asm better than me care to explain it in > simple words? First, ({}) is a statement expression - a GCC extension (not to be confused with expression statement, which is an expression followed by a semicolon). It wraps a compound statement, i.e. {}, and turns it into an expression. The value of the last statement is the value of the expression. In this case it's __p;. The per-cpu information can be accessed in two ways: Either by accessing it in the %fs segment (the then clause does it) or reading its address from %fs:pc_prvspace and then accessing it in the normal (i.e. %ds) segment (the else clause does this). Let's have a closer look at the else clause: The asm reads the pointer to per-cpu information into __p and the statement expression returns it. This pointer gets dereferenced (mind the * before __extension__ - __extension__ does nothing, it just marks that the following is a GCC extension) and __val is assigned. This else clause is for assignment for things which are not 1, 2 or 4 bytes in size. For sizes 1, 2 and 4 better code can be generated for by not first getting the pointer, but directly assigning the value into the %fs mappinf of the per-cpu info. Regards ChristophReceived on Sun Aug 02 2009 - 09:00:53 UTC
This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:39:53 UTC