https://sourceware.org/bugzilla/show_bug.cgi?id=30967
--- Comment #8 from Frank Ch. Eigler <fche at redhat dot com> ---
> The issue here is that we create (very) large arrays of struct Dwarf_Line_s
> and we do that eagerly, see bug #27405
> So we would like to keep that struct as small as possible.
Do we have an estimate about the numbers we're talking about here? The current
Dwarf_Line_s object is 40 bytes long (on x86-64). Even if packing becomes less
efficient, it could cost us say 2 bytes more per record. These records are
packed together in allocation via realloc() et al. How many records do we see
in programs of interest? readelf -wL /bin/emacs indicates about 800 thousand;
libxul.so about 8 million. Is this about single digit megabytes more RAM in
grand total?
Note that bug #27405 was a request for optimization, not in order to save a few
percent of memory for used data, but to save ALL the related memory & time for
totally unused data.
struct Dwarf_Line_s {
Dwarf_Files * files; /* 0 8 */
Dwarf_Addr addr; /* 8 8 */
unsigned int file; /* 16 4 */
int line; /* 20 4 */
short unsigned int column; /* 24 2 */
/* Bitfield combined with previous fields */
unsigned int is_stmt:1; /* 24:16 4 */
unsigned int basic_block:1; /* 24:17 4 */
unsigned int end_sequence:1; /* 24:18 4 */
unsigned int prologue_end:1; /* 24:19 4 */
unsigned int epilogue_begin:1; /* 24:20 4 */
unsigned int op_index:8; /* 24:21 4 */
/* XXX 3 bits hole, try to pack */
/* Force alignment to the next boundary: */
unsigned int :0;
unsigned int isa:8; /* 28: 0 4 */
unsigned int discriminator:24; /* 28: 8 4 */
unsigned int context; /* 32 4 */
unsigned int function_name; /* 36 4 */
/* size: 40, cachelines: 1, members: 15 */
/* sum members: 34 */
/* sum bitfield members: 45 bits, bit holes: 1, sum bit holes: 3 bits
*/
/* last cacheline: 40 bytes */
};
--
You are receiving this mail because:
You are on the CC list for the bug.
