Jakub Bogusz from pld-linux dot org writes:

Hello,

I managed to get ldso (and thus shared linking to uClibc) to work on
sparc (actually sparc64 kernel with 32-bit userspace), at least on
simple "hello world" program (more complex ones not tested).

Some notes on attached patch (against 0.9.26, would require some work
to apply on current CVS - but I tested 0.9.26, not CVS):

- ELF magic cannot be examined by _dl_strncmp so early, probably because of
  string constant, like on ppc/mips/sh
  (note that early SEND_STDERR still crashes when trying to do _dl_strlen
   - I suppose that string constants require relocation; but adding
   load_addr didn't help, just ELF header was displayed instead of crash)
- mmap() is syscall6 like on ppc/mips/sh, not old i386 mmap()
- for generic sparc (i.e. not sparcv8/sparcv9) gcc produces .udiv/.urem
  calls for unsigned integer / and % operators - so these operations
  must be avoided. I copied do_rem definition from arm header.
  But / and % are used also in _dl_simple_ltoa() and
  _dl_simple_ltoahex(); in ltoahex gcc optimizes it to shifts (but
  I think it's safer to use shifts explicitly, not rely on
  optimization...).
  I changed % in ltoa to do_rem, but as there was no do_div definition,
  I changed all "%d" specifiers to "%x" to avoid crashes (this changes
  wouldn't be needed if _dl_simple_ltoa() were fixed to not use
  division on sparc).
- "#define SOLARIS_COMPATIBLE" in ld_sysdep.h broke ldso on Linux
  because of redefining _dl_linux_resolve only in some places (it was
  still referenced in INIT_GOT before redefinition). So
  _dl_linux_resolve redefinition should be moved before INIT_GOT
  definition or removed.
- sparc64 kernel requires mmap() addresses to be aligned to 8192, not
  4096, otherwise mmap() call failed
- reloc_entry must be shifted by 10, not 12 (I found similar operation
  in glibc sources)

Aside of sparc-specific fixes:
- I moved some _dl_dprintf()s inside if(_dl_debug_*) conditions (to avoid
  debugging messages when LD_DEBUG is not defined)
- it seems that there was possible off-by-one in ltoa and ltoahex?
  they are called with char[22] as 1st argument, and then '\0' is stored
  in local[22] (_before_ p decrementation)... or am I missing something?
  If not, fix is included in patch.

1 files changed, 47 insertions, 4 deletions

diff --git a/ldso/ldso/sparc/dl-sysdep.h b/ldso/ldso/sparc/dl-sysdep.h
index c6ac89e11..ddf74e21e 100644
--- a/ldso/ldso/sparc/dl-sysdep.h
+++ b/ldso/ldso/sparc/dl-sysdep.h
@@ -41,9 +41,51 @@ extern unsigned int _dl_linux_resolver(unsigned int reloc_entry,
 /*
  * Define this if you want a dynamic loader that works on Solaris.
  */
+#ifndef __linux__
 #define SOLARIS_COMPATIBLE
+#endif
+
+#ifndef COMPILE_ASM
+/* Cheap modulo implementation, taken from arm/ld_sysdep.h. */
+static inline unsigned long
+sparc_mod(unsigned long m, unsigned long p)
+{
+	unsigned long i, t, inc;
+
+	i = p;
+	t = 0;
+
+	while (!(i & (1 << 31))) {
+		i <<= 1;
+		t++;
+	}
+
+	t--;
+
+	for (inc = t; inc > 2; inc--) {
+		i = p << inc;
 
-#define do_rem(result, n, base)	    result = (n % base)
+		if (i & (1 << 31))
+			break;
+
+		while (m >= i) {
+			m -= i;
+			i <<= 1;
+			if (i & (1 << 31))
+				break;
+			if (i < p)
+				break;
+		}
+	}
+
+	while (m >= p)
+		m -= p;
+
+	return m;
+}
+
+#define do_rem(result, n, base) result = sparc_mod(n, base);
+#endif
 
 /*
  * dbx wants the binder to have a specific name.  Mustn't disappoint it.
@@ -53,9 +95,10 @@ extern unsigned int _dl_linux_resolver(unsigned int reloc_entry,
 #endif
 
 /* 4096 bytes alignment */
-#define PAGE_ALIGN 0xfffff000
-#define ADDR_ALIGN 0xfff
-#define OFFS_ALIGN 0x7ffff000
+/* ...but 8192 is required for mmap() on sparc64 kernel */
+#define PAGE_ALIGN 0xffffe000
+#define ADDR_ALIGN 0x1fff
+#define OFFS_ALIGN 0x7fffe000
 
 /* ELF_RTYPE_CLASS_PLT iff TYPE describes relocation of a PLT entry, so
    PLT entries should not be allowed to define the value.