can't find a register in class 'GENERAL_REGS' while reloading 'asm'

[Michael Haubenwallner <michael dot haubenwallner at salomon dot at>]

Hi,

in openssl-0.9.8e there is some inlined assembler code, which does not
compile with -fpic (or -fPIC) and -O0 on x86-linux-gnu.

$ gcc -c openssl-error.c
$ gcc -c openssl-error.c -fPIC
openssl-error.c: In function 'padlock_xcrypt_ecb':
openssl-error.c:28: error: can't find a register in class 'GENERAL_REGS' while reloading 'asm'
$ gcc -c openssl-error.c -fPIC -O1
$

It can be built without any -fpic/-fPIC, or with any optimization >= 1.

This sample is reduced from preprocessor-output of
openssl-0.9.8e/crypto/engine/eng_padlock.c on i686-pc-linux-gnu, and the
error is same for gcc-3.3.6, gcc-3.4.6, gcc-4.1.1 on i686-gentoo-linux,
as well as gcc-3.4.4 on i386 "Red Hat Enterprise Linux AS release 4
(Nahant Update 2)".

Is this a bug in that assembler-code or in the compiler, or even both ?

Thanks!

/haubi/

struct aes_key_st {
    unsigned int rd_key[4 *(14 + 1)];
    int rounds;
};
typedef struct aes_key_st AES_KEY;

struct padlock_cipher_data
{
 unsigned char iv[16];
 union { unsigned int pad[4];
  struct {
   int rounds:4;
   int dgst:1;
   int align:1;
   int ciphr:1;
   unsigned int keygen:1;
   int interm:1;
   unsigned int encdec:1;
   int ksize:2;
  } b;
 } cword;
 AES_KEY ks;
};

static inline void *padlock_xcrypt_ecb(unsigned int cnt, struct padlock_cipher_data *cdata, void *out, const void *inp)
{
       void *iv;
       asm volatile (
               "pushl	%%ebx\n"
               "	leal	16(%0),%%edx\n"
               "	leal	32(%0),%%ebx\n"
               ".byte 0xf3,0x0f,0xa7,0xc8" "\n"
               "	popl	%%ebx"
               : "=a"(iv), "=c"(cnt), "=D"(out), "=S"(inp)
               : "0"(cdata), "1"(cnt), "2"(out), "3"(inp), "m"(*cdata)
               : "edx", "cc", "memory"
       );
       return iv;
}

int
padlock_aes_cipher_omnivorous(unsigned char *out_arg,
  const unsigned char *in_arg, unsigned int nbytes)
{
 struct padlock_cipher_data *cdata;

  padlock_xcrypt_ecb(nbytes/16, cdata, out_arg, in_arg);


 return 1;
}