dwarf_loader: increase the size of lookup hash map

One of the primary use cases for using pahole is BTF deduplication during
Linux kernel build. That means that DWARF contains more than 5 million types
is loaded. So using a hash map with a small number of buckets is quite
expensive due to hash collisions. This patch bumps the size of the hash map
and reduces overhead of this part of the DWARF loading process.

This shaves off about 1 second out of about 20 seconds total for Linux BTF
dedup.

Committer testing:

Before:

  $ perf stat -r5 pahole -J vmlinux

   Performance counter stats for 'pahole -J vmlinux' (5 runs):

            8,953.80 msec task-clock:u              #    0.998 CPUs utilized            ( +-  0.09% )
                   0      context-switches:u        #    0.000 K/sec
                   0      cpu-migrations:u          #    0.000 K/sec
             353,855      page-faults:u             #    0.040 M/sec                    ( +-  0.00% )
      35,775,730,539      cycles:u                  #    3.996 GHz                      ( +-  0.07% )  (83.33%)
         579,534,836      stalled-cycles-frontend:u #    1.62% frontend cycles idle     ( +-  2.21% )  (83.33%)
       5,719,840,144      stalled-cycles-backend:u  #   15.99% backend cycles idle      ( +-  0.93% )  (83.33%)
      73,035,744,786      instructions:u            #    2.04  insn per cycle
                                                    #    0.08  stalled cycles per insn  ( +-  0.02% )  (83.34%)
      16,798,017,844      branches:u                # 1876.077 M/sec                    ( +-  0.05% )  (83.33%)
         237,777,143      branch-misses:u           #    1.42% of all branches          ( +-  0.15% )  (83.34%)

             8.97077 +- 0.00803 seconds time elapsed  ( +-  0.09% )

  $

After:

  $ perf stat -r5 pahole -J vmlinux

   Performance counter stats for 'pahole -J vmlinux' (5 runs):

            8,735.92 msec task-clock:u              #    0.998 CPUs utilized            ( +-  0.34% )
                   0      context-switches:u        #    0.000 K/sec
                   0      cpu-migrations:u          #    0.000 K/sec
             353,978      page-faults:u             #    0.041 M/sec                    ( +-  0.00% )
      34,722,167,335      cycles:u                  #    3.975 GHz                      ( +-  0.12% )  (83.33%)
         555,981,118      stalled-cycles-frontend:u #    1.60% frontend cycles idle     ( +-  1.53% )  (83.33%)
       5,215,370,531      stalled-cycles-backend:u  #   15.02% backend cycles idle      ( +-  1.31% )  (83.33%)
      72,615,773,119      instructions:u            #    2.09  insn per cycle
                                                    #    0.07  stalled cycles per insn  ( +-  0.02% )  (83.34%)
      16,624,959,121      branches:u                # 1903.057 M/sec                    ( +-  0.01% )  (83.33%)
         229,962,327      branch-misses:u           #    1.38% of all branches          ( +-  0.07% )  (83.33%)

              8.7503 +- 0.0301 seconds time elapsed  ( +-  0.34% )

  $

2.94% less cycles, good :-)

Signed-off-by: Andrii Nakryiko <[email protected]>
Tested-by: Arnaldo Carvalho de Melo <[email protected]>
Cc: Alexei Starovoitov <[email protected]>
Cc: Andrii Nakryiko <[email protected]>
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>

Author: anakryiko

dwarf_loader.c

@@ -93,7 +93,7 @@ static void dwarf_tag__set_spec(struct dwarf_tag *dtag, dwarf_off_ref spec)
 	*(dwarf_off_ref *)(dtag + 1) = spec;
 }
 
-#define HASHTAGS__BITS 8
+#define HASHTAGS__BITS 15
 #define HASHTAGS__SIZE (1UL << HASHTAGS__BITS)
 
 #define obstack_chunk_alloc malloc