g_extern.audio_data.data NULL on ios 8 with retroarch 1.0.0.2 / 1.0.3

[jeapostrophe]

I'm building my own version for a non-Jailbroken device. I am using the latest git version as of yesterday.

I got lots of crashes that had stacks like this:

Exception Type: EXC_BAD_ACCESS (SIGSEGV)
Exception Subtype: KERN_INVALID_ADDRESS at 0x00000000

0 RetroArch 0x00013c88 audio_convert_s16_float_asm (utils_neon.S:63)
1 RetroArch 0x0006b476 audio_convert_s16_to_float_neon (utils.c:159)
2 RetroArch 0x000435d8 rarch_audio_flush (retroarch.c:483)
3 RetroArch 0x00042b46 audio_sample_batch (libretro_version_1.c:125)

To simplify debugging, I disabled the NEON assembly by changing utils.c:283 to

audio_convert_s16_to_float_arm = cpu & RETRO_SIMD_NEON & 0 ?

(notice the & 0)

This got me crashes like this:

Exception Type: EXC_BAD_ACCESS (SIGSEGV)
Exception Subtype: KERN_INVALID_ADDRESS at 0x00000000

0 RetroArch 0x0006b4fe audio_convert_s16_to_float_C (utils.c:35)
1 RetroArch 0x00043628 rarch_audio_flush (retroarch.c:483)
2 RetroArch 0x00042b96 audio_sample_batch (libretro_version_1.c:125)

Since that code dereferences out and in, I went back to rarch_audio_flush to look at its arguments. In retroarch.c:484-ish, it is called with:

out = g_extern.audio_data.data and in = data

I added this around line 481

if (!g_extern.audio_data.data)
return false;

to protect against it being NULL as an experiment. At this point, the crashes went away and I got the expected result of the game playing but not having any audio.

The initialization code from driver.c:1107 is wrapped in an rarch_assert, which is surprising because it is definitely NULL here.

Any ideas about how I might fix this problem or provide more useful debugging information?

[jeapostrophe]

I have found that it is being initialized correctly by adding some printfs:

Sep 30 11:28:26 RetroArch[3036] : BEFORE Initializing g_extern.audio_data.data: 0000000000000000
Sep 30 11:28:26 RetroArch[3036] : AFTER Initializing g_extern.audio_data.data: 000000000577c000

But that by the time we go to flush, it has been zero'd

Sep 30 11:28:26 RetroArch[3036] : Getting ready to flush to g_extern.audio_data.data: 0000000000000000

I also added some printfs in the uninit audio code and it is NOT getting uninitialized early. Any where else that audio_data could be effected?

[jeapostrophe]

https://github.com/libretro/RetroArch/blob/master/retroarch.c#L2536

That's the line that kills it.

Sep 30 11:55:00 Bleu RetroArch[3074] : BEFORE Initializing g_extern.audio_data.data: 0000000000000000
Sep 30 11:55:00 Bleu RetroArch[3074] : AFTER Initializing g_extern.audio_data.data: 0000000006eb3000
Sep 30 11:55:00 Bleu RetroArch[3074] : Exiting init_audio: g_extern.audio_data.data: 0000000006eb3000
Sep 30 11:55:00 Bleu RetroArch[3074] : GL VSync => on
Sep 30 11:55:00 Bleu RetroArch[3074] : BEFORE (clearing g_extern) g_extern.audio_data.data: 0000000006eb3000
Sep 30 11:55:00 Bleu RetroArch[3074] : AFTER (clearing g_extern) g_extern.audio_data.data: 0000000000000000

[jeapostrophe]

diff --git a/retroarch.c b/retroarch.c
index abb060d..fb8b658 100644
--- a/retroarch.c
+++ b/retroarch.c
@@ -475,7 +475,7 @@ bool rarch_audio_flush(const int16_t *data, size_t samples)

    if (g_extern.is_paused || g_extern.audio_data.mute)
       return true;
-   if (!g_extern.audio_active)
+   if (!g_extern.audio_active || !g_extern.audio_data.data)
       return false;

    RARCH_PERFORMANCE_INIT(audio_convert_s16);
@@ -2533,7 +2533,12 @@ void rarch_main_clear_state(void)

    deinit_log_file();

+   // XXX This memset is really dangerous. (a) it can leak memory
+   // because the pointers in g_extern aren't freed and (b) it can
+   // zero pointers that the rest of the code will look at.
+   uninit_drivers();
    memset(&g_extern, 0, sizeof(g_extern));
+   init_drivers();

    init_state();