From 815041d6dc4d62bfe7267576a8acb65e774fc94f Mon Sep 17 00:00:00 2001
From: Arnaldo Carvalho de Melo <acme@redhat.com>
Date: Fri, 6 Aug 2021 16:46:25 -0300
Subject: [PATCH] pahole: Improve the type sorting routine to consider multiple
 types with same name

Out of these different Linux kernel types with the same name (in different object files):

  $ pahole --sizes | sed -r 's/([^\t]+)\t.*/\1/g' | sort | uniq -c | grep -v ' 1 '
        2 chksum_desc_ctx
        2 controller
        2 debug_buffer
        2 dir_entry
        2 disklabel
        2 dma_chan
        2 dma_heap_attachment
        2 d_partition
        2 elf_thread_core_info
        2 intel_community_context
        2 intel_pad_context
        3 irq_info
        2 irte
        2 map_info
        2 mm_slot
        2 netlbl_domhsh_walk_arg
        2 node
        2 pci_root_info
        2 perf_aux_event
        2 pmc_dev
        2 pmc_reg_map
        2 remap_data
        2 slot
        2 sw842_param
        2 syscall_tp_t
        3 urb_priv
        2 walk_control
        3 workspace
  $

Only this one needs a more involved type comparision:

  $ btfdiff vmlinux
  --- /tmp/btfdiff.dwarf.Pksrlr	2021-08-06 16:42:34.823259365 -0300
  +++ /tmp/btfdiff.btf.KOAuwd	2021-08-06 16:42:35.032264038 -0300
  @@ -31035,7 +31035,7 @@ struct elf_note_info {
   	struct memelfnote          auxv;                 /*    56    24 */
   	/* --- cacheline 1 boundary (64 bytes) was 16 bytes ago --- */
   	struct memelfnote          files;                /*    80    24 */
  -	compat_siginfo_t           csigdata;             /*   104   128 */
  +	siginfo_t                  csigdata;             /*   104   128 */
   	/* --- cacheline 3 boundary (192 bytes) was 40 bytes ago --- */
   	size_t                     size;                 /*   232     8 */
   	int                        thread_notes;         /*   240     4 */
  $

It has the same size, number of members.

And this is not always, it all depends on the order in which the btf
encoder gets it from one of the DWARF loading threads:

  $ pahole -j12 --btf_encode vmlinux
  $ btfdiff vmlinux
  $

No changes, but then:

  $ btfdiff vmlinux
  $ perf stat pahole -j12 --btf_encode vmlinux

   Performance counter stats for 'pahole -j12 --btf_encode vmlinux':

           17,920.75 msec task-clock:u              #    2.995 CPUs utilized
                   0      context-switches:u        #    0.000 /sec
                   0      cpu-migrations:u          #    0.000 /sec
              78,004      page-faults:u             #    4.353 K/sec
      42,677,746,170      cycles:u                  #    2.381 GHz                      (83.37%)
         480,920,924      stalled-cycles-frontend:u #    1.13% frontend cycles idle     (83.33%)
       6,470,001,379      stalled-cycles-backend:u  #   15.16% backend cycles idle      (83.39%)
      96,468,468,147      instructions:u            #    2.26  insn per cycle
                                                    #    0.07  stalled cycles per insn  (83.33%)
      19,757,801,968      branches:u                #    1.103 G/sec                    (83.27%)
         143,118,731      branch-misses:u           #    0.72% of all branches          (83.32%)

         5.984348164 seconds time elapsed

        17.234929000 seconds user
         0.398715000 seconds sys

  $ btfdiff vmlinux
  --- /tmp/btfdiff.dwarf.b9FEZI	2021-08-06 16:46:08.810043718 -0300
  +++ /tmp/btfdiff.btf.IawvDY	2021-08-06 16:46:09.026048548 -0300
  @@ -31035,7 +31035,7 @@ struct elf_note_info {
   	struct memelfnote          auxv;                 /*    56    24 */
   	/* --- cacheline 1 boundary (64 bytes) was 16 bytes ago --- */
   	struct memelfnote          files;                /*    80    24 */
  -	compat_siginfo_t           csigdata;             /*   104   128 */
  +	siginfo_t                  csigdata;             /*   104   128 */
   	/* --- cacheline 3 boundary (192 bytes) was 40 bytes ago --- */
   	size_t                     size;                 /*   232     8 */
   	int                        thread_notes;         /*   240     4 */
  $

Next cset will take that into account by traversing both types looking
for differences in the type for a field.

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
---
 pahole.c | 21 +++++++++++++++++++--
 1 file changed, 19 insertions(+), 2 deletions(-)

diff --git a/pahole.c b/pahole.c
index 32eb287..e3e6d66 100644
--- a/pahole.c
+++ b/pahole.c
@@ -123,19 +123,36 @@ static struct rb_root structures__tree = RB_ROOT;
 static LIST_HEAD(structures__list);
 static pthread_mutex_t structures_lock = PTHREAD_MUTEX_INITIALIZER;
 
+static int type__compare(struct type *a, struct type *b)
+{
+	int ret = strcmp(type__name(a), type__name(b));
+
+	if (ret)
+		goto found;
+
+	ret = (int)a->size - (int)b->size;
+	if (ret)
+		goto found;
+
+	ret = (int)a->nr_members - (int)b->nr_members;
+	if (ret)
+		goto found;
+found:
+	return ret;
+}
+
 static struct structure *__structures__add(struct class *class, struct cu *cu, uint32_t id, bool *existing_entry)
 {
         struct rb_node **p = &structures__tree.rb_node;
         struct rb_node *parent = NULL;
 	struct structure *str;
-	const char *new_class_name = class__name(class);
 
         while (*p != NULL) {
 		int rc;
 
                 parent = *p;
                 str = rb_entry(parent, struct structure, rb_node);
-		rc = strcmp(class__name(str->class), new_class_name);
+		rc = type__compare(&str->class->type, &class->type);
 
 		if (rc > 0)
                         p = &(*p)->rb_left;