Solo mode went away in v3.22.0

[YetAnotherRussian]

Hi. Crashing again :'(

Win binaries:

Linux (WSL) default build:

I've tried 4 algos I can test, possibly not algo dependant. Will add some debug info soon.

Stratum is OK.

UPD: Debug info:

How about making "build-gdb.sh" script in repo?

Just copy the default one, add "-ggdb" and remove any stripping? So there's an abiity to build right after cloning. I can make pr, but it's just a copy-paste operation.

[JayDDee]

It looks like I misssed a line copying changes from util.c:json_rpc_call, the last line before comment. It's new and I missed it.
I deliberately left out any changes that would obviously affect GBT, I thought json_rpc_call would be adequately tested using stratum.

Edit: the bug is related to merging pr 204 from Pooler. util.c also merged pr 206 (hex2bin).
https://github.com/pooler/cpuminer/commits/master

json_t *json_rpc_call(CURL *curl, const char *url,
const char *userpass, const char *rpc_req,
int *curl_err, int flags)
{
json_t *val, *err_val, *res_val;
int rc;
long http_rc;
struct data_buffer all_data = {0};
char *json_buf;
json_error_t err;
struct curl_slist *headers = NULL;
char curl_err_str[CURL_ERROR_SIZE];
long timeout = (flags & JSON_RPC_LONGPOLL) ? opt_timeout : 30;
struct header_info hi = {0};

all_data.headers = &hi;
/* it is assumed that 'curl' is freshly [re]initialized at this pt */

If this doesn't fix it try replacing util.c with the old version.

Regarding your suggestion for a debug build script, there's a certain skill set associated with that. Debuggng requires some creativity, it's hard to script it. Someone with the skill set to use GDB would also have the skill set to build for it.
Consider it a test, you can use GDB if you know how to build for it.

I've been testing the MSYS build using the standard build.sh. build_msys2.sh has an extra option that was apparently needed at the time. It seems to compile without it now. I'm not sure if it was ever needed. If you've been adding --param=evrp-mode=legacy, try withou it.

[YetAnotherRussian]

If this doesn't fix it

Just checked, seems to work. No need to replace util.c

[JayDDee]

Thanks. Silly error.

Edit: I'm working on another potential optimization and it's going well. I may hold off another day or 2 for a new release to include it.

[JayDDee]

cpuminer-opt-3.22.1 is released.

[JayDDee]

Closing thoughts on the misalingment issue. I think the compiler should have handled this much better.
For dynamically allocated blocks, for which the application has little control, the optimizer should not make changes that require alignment greater than the default malloc alignment.

There are many ways this problem could have been avoided:

1. use AVX2 unaligned memory accesses
2. malloc automatically aligns pointer based on target/array type
3. optimizer realigns block to match stricter requirements
4. not optimizing beyond default alignment or if required alignment can't be guaranteed

I don't think this the end of it. The fix I implemented is a workaound specific to certain instances, it doesn't solve the root of the problem.