I've created workaround for buggy 13.x catalyst driver, needs some testing though (especially the miner and win32).
For the impatient:
linux:
Code:
$ git clone https://github.com/wyuzhe/vanitygen.git
$ cd vanitygen
$ make oclvanitygen
$ ./oclvanitygen -vv -D x:y 1Blah
$ cd vanitygen
$ make oclvanitygen
$ ./oclvanitygen -vv -D x:y 1Blah
In case you get odd output like:
Code:
Match idx: 0
CPU hash: 0ecfec41290a506e784a4521597213398abf9a98
GPU hash: 08e030855f47c3141a5808d8767a89562ec5c655
Found delta: 497669 Start delta: 1
CPU hash: 0ecfec41290a506e784a4521597213398abf9a98
GPU hash: 08e030855f47c3141a5808d8767a89562ec5c655
Found delta: 497669 Start delta: 1
It means the compiler backend is still outputting broken code (although it compiles). To fix that try:
Code:
./catalystwrap ./oclvanitygen -vv -D 0:0 1ov
Windows users can try applying the dll override script in the directory of official vanitygen binaries [edit: link with resulting archive removed due security concerns].
The gory details:
https://github.com/wyuzhe/vanitygen/commit/6f7fd04adc609b19520cdab4cc12d648e364adbe
In short, ATI LLVM-IR backend miscompiles stuff like:
Code:
uint x, b, c;
x += (a < b)
x += (a < b)
The boolean result of (a < b) is expressed as byte (u8 in crash report), but VLIW architectures know no such thing. Workaround is using artificially complex expressions which cannot be readily expressed as setcc in llvm ir (but will be optimized away during R600 lowering anyway).
