diff mbox

[1/4] Nios II: elf subdir changes, unsigned divide in ld.so

Message ID 533597A8.10101@codesourcery.com
State Superseded
Headers show

Commit Message

Chung-Lin Tang March 28, 2014, 3:39 p.m. UTC
This patch contains the changes in the elf/ subdir. The main changes are:

(1) The changes to add Nios II to elf/elf.h, machine number,
relocations, etc. Pretty standard stuff copied from BFD.

(2) A change in elf/dynamic-link.h, to allow a port to define a special
unsigned integer divide during RTLD_BOOTSTRAP, by defining
RTLD_BOOTSTRAP_UDIV.

The rationale is: while Nios II does have optional hardware integer
divide defined as part of the ISA spec, the current support we're
working on here is for the common no-divider case, relying on libgcc
__udivsi3.

Unfortunately however, Nios II does not have PC-relative calls (only a
direct transfer call similar to MIPS), so all subroutine calls under PIC
have to go through the GOT, which... is not available yet during
RTLD_BOOTSTRAP.

So the RTLD_BOOTSTRAP_UDIV macro facility allows nios2 to define its own
inline udiv in dl-machine.h, for use when libgcc is still not accessible.

Thanks,
Chung-Lin

        * elf/dynamic-link.h (RTLD_UDIV): New macro.
        (elf_machine_lazy_rel): Change divide operation to use RTLD_UDIV.
        * elf/elf.h (EM_ALTERA_NIOS2): New machine number for Altera
Nios II.
        (DT_NIOS2_GP): New dynamic entry type for Nios II _gp address.
        (R_NIOS2_*): Define Nios II relocations.

Comments

Joseph Myers March 28, 2014, 4:38 p.m. UTC | #1
On Fri, 28 Mar 2014, Chung-Lin Tang wrote:

> (2) A change in elf/dynamic-link.h, to allow a port to define a special
> unsigned integer divide during RTLD_BOOTSTRAP, by defining
> RTLD_BOOTSTRAP_UDIV.
> 
> The rationale is: while Nios II does have optional hardware integer
> divide defined as part of the ISA spec, the current support we're
> working on here is for the common no-divider case, relying on libgcc
> __udivsi3.
> 
> Unfortunately however, Nios II does not have PC-relative calls (only a
> direct transfer call similar to MIPS), so all subroutine calls under PIC
> have to go through the GOT, which... is not available yet during
> RTLD_BOOTSTRAP.
> 
> So the RTLD_BOOTSTRAP_UDIV macro facility allows nios2 to define its own
> inline udiv in dl-machine.h, for use when libgcc is still not accessible.

With this approach, there's a risk of a future GCC version noticing an 
implementation of division and optimizing it back into a library call.  
Did you consider a -minline-divide or similar GCC option that could be 
used for building the dynamic linker (and if so, what was the rationale 
for rejecting it)?  (See below for one possible reason not to use that 
approach.)

>         (R_NIOS2_*): Define Nios II relocations.

ChangeLog entries need to mention each new symbol individually.

> +/* Allow targets that do not have libgcc __udivsi3 available to define
> +   one locally.  */
> +# if defined RTLD_BOOTSTRAP && defined RTLD_BOOTSTRAP_UDIV
> +#  define RTLD_UDIV(A, B) RTLD_BOOTSTRAP_UDIV (A, B)
> +# else
> +#  define RTLD_UDIV(A, B) ((A) / (B))
> +# endif

#ifdef-like conditionals have a risk of typos being quietly unnoticed, as 
discussed recently in the -Wundef discussions, so for a new macro it would 
be better to have it always defined exactly once without any "#if 
defined".  For example, a header dl-udiv.h that defines 
RTLD_BOOTSTRAP_UDIV trivially in the generic version, non-trivially in the 
Nios II version.

The comment here needs to explain the semantics to be provided by the 
macro RTLD_BOOTSTRAP_UDIV.  In particular, can it be assumed (a) that the 
division is always for word-sized values, (b) that it is always exact, (c) 
that it is always dividing by sizeof (ElfW(reloc))?

Given those assumptions, you could do something like

#define RTLD_BOOTSTRAP_UDIV(A, B) \
  ({ \
    _Static_assert ((B) == 12, "bad arguments to RTLD_BOOTSTRAP_UDIV"); \
    ((A) >> 2) * 0xaaaaaaab; \
  })

(untested), to implement exact division in terms of multiplication, if 
that's available inline and smaller / faster than dividing bit-by-bit.  
(And if multiplication isn't inline either, then you can use a few shifts 
and additions for that particular multiplication.)

I'm not saying these other approaches to the division are necessarily 
smaller / faster on Nios II, but if the interface means the division is 
exact then they should at least be considered.  (And if it's exact, but 
the compiler can't tell that, that's a reason against using an option for 
the compiler to inline division.)

Also, it would be worth reviewing the other approaches discussed when this 
issue was raised for the Xtensa port of glibc 
<https://sourceware.org/ml/libc-alpha/2007-04/msg00002.html>.

> +/* Nios II relocations.  */
> +#define R_NIOS2_NONE		0       /* No reloc.  */
> +#define R_NIOS2_S16		1       /* Direct signed 16 bit.  */
> +#define R_NIOS2_U16		2       /* Direct unsigned 16 bit.  */
> +#define R_NIOS2_PCREL16		3       /* PC relative 16 bit.  */
> +#define R_NIOS2_CALL26		4

Each macro should have a comment rather than giving up on comments after 
the first few.
Roland McGrath March 29, 2014, 12:39 a.m. UTC | #2
Send the elf.h changes alone.  Those can go in first, assuming the
corresponding constants and names are already commited in binutils.
But we'd like to see a short comment on each and every R_* macro.


Thanks,
Roland
diff mbox

Patch

diff --git a/elf/dynamic-link.h b/elf/dynamic-link.h
index 7b3e295..67dabd5 100644
--- a/elf/dynamic-link.h
+++ b/elf/dynamic-link.h
@@ -97,6 +97,15 @@  elf_machine_lazy_rel (struct link_map *map,
 #  define ELF_DURING_STARTUP (0)
 # endif
 
+/* Allow targets that do not have libgcc __udivsi3 available to define
+   one locally.  */
+# if defined RTLD_BOOTSTRAP && defined RTLD_BOOTSTRAP_UDIV
+#  define RTLD_UDIV(A, B) RTLD_BOOTSTRAP_UDIV (A, B)
+# else
+#  define RTLD_UDIV(A, B) ((A) / (B))
+# endif
+
+
 /* Get the definitions of `elf_dynamic_do_rel' and `elf_dynamic_do_rela'.
    These functions are almost identical, so we use cpp magic to avoid
    duplicating their code.  It cannot be done in a more general function
@@ -123,7 +132,7 @@  elf_machine_lazy_rel (struct link_map *map,
 	if (map->l_info[VERSYMIDX (DT_##RELOC##COUNT)] != NULL)		      \
 	  ranges[0].nrelative						      \
 	    = MIN (map->l_info[VERSYMIDX (DT_##RELOC##COUNT)]->d_un.d_val,    \
-		   ranges[0].size / sizeof (ElfW(reloc)));		      \
+		   RTLD_UDIV (ranges[0].size, sizeof (ElfW(reloc))));	      \
       }									      \
     if ((map)->l_info[DT_PLTREL]					      \
 	&& (!test_rel || (map)->l_info[DT_PLTREL]->d_un.d_val == DT_##RELOC)) \
diff --git a/elf/elf.h b/elf/elf.h
index 40e87b2..85401c3 100644
--- a/elf/elf.h
+++ b/elf/elf.h
@@ -249,6 +249,7 @@  typedef struct
 #define EM_OPENRISC	92		/* OpenRISC 32-bit embedded processor */
 #define EM_ARC_A5	93		/* ARC Cores Tangent-A5 */
 #define EM_XTENSA	94		/* Tensilica Xtensa Architecture */
+#define EM_ALTERA_NIOS2 113		/* Altera Nios II */
 #define EM_AARCH64	183		/* ARM AARCH64 */
 #define EM_TILEPRO	188		/* Tilera TILEPro */
 #define EM_MICROBLAZE	189		/* Xilinx MicroBlaze */
@@ -3132,6 +3133,57 @@  typedef Elf32_Addr Elf32_Conflict;
 #define R_MICROBLAZE_TLSGOTTPREL32	28	/* TLS Offset From Thread Pointer. */
 #define R_MICROBLAZE_TLSTPREL32 	29	/* TLS Offset From Thread Pointer. */
 
+/* Legal values for d_tag (dynamic entry type).  */
+#define DT_NIOS2_GP             0x70000002 /* Address of _gp.  */
+
+/* Nios II relocations.  */
+#define R_NIOS2_NONE		0       /* No reloc.  */
+#define R_NIOS2_S16		1       /* Direct signed 16 bit.  */
+#define R_NIOS2_U16		2       /* Direct unsigned 16 bit.  */
+#define R_NIOS2_PCREL16		3       /* PC relative 16 bit.  */
+#define R_NIOS2_CALL26		4
+#define R_NIOS2_IMM5		5
+#define R_NIOS2_CACHE_OPX	6
+#define R_NIOS2_IMM6		7
+#define R_NIOS2_IMM8		8
+#define R_NIOS2_HI16		9
+#define R_NIOS2_LO16		10
+#define R_NIOS2_HIADJ16		11
+#define R_NIOS2_BFD_RELOC_32	12
+#define R_NIOS2_BFD_RELOC_16	13
+#define R_NIOS2_BFD_RELOC_8	14
+#define R_NIOS2_GPREL		15
+#define R_NIOS2_GNU_VTINHERIT	16
+#define R_NIOS2_GNU_VTENTRY	17
+#define R_NIOS2_UJMP		18
+#define R_NIOS2_CJMP		19
+#define R_NIOS2_CALLR		20
+#define R_NIOS2_ALIGN		21
+#define R_NIOS2_GOT16		22
+#define R_NIOS2_CALL16		23
+#define R_NIOS2_GOTOFF_LO	24
+#define R_NIOS2_GOTOFF_HA	25
+#define R_NIOS2_PCREL_LO	26
+#define R_NIOS2_PCREL_HA	27
+#define R_NIOS2_TLS_GD16	28
+#define R_NIOS2_TLS_LDM16	29
+#define R_NIOS2_TLS_LDO16	30
+#define R_NIOS2_TLS_IE16	31
+#define R_NIOS2_TLS_LE16	32
+#define R_NIOS2_TLS_DTPMOD	33
+#define R_NIOS2_TLS_DTPREL	34
+#define R_NIOS2_TLS_TPREL	35
+#define R_NIOS2_COPY		36
+#define R_NIOS2_GLOB_DAT	37
+#define R_NIOS2_JUMP_SLOT	38
+#define R_NIOS2_RELATIVE	39
+#define R_NIOS2_GOTOFF		40
+#define R_NIOS2_CALL26_NOAT	41
+#define R_NIOS2_GOT_LO		42
+#define R_NIOS2_GOT_HA		43
+#define R_NIOS2_CALL_LO		44
+#define R_NIOS2_CALL_HA		45
+
 /* TILEPro relocations.  */
 #define R_TILEPRO_NONE		0	/* No reloc */
 #define R_TILEPRO_32		1	/* Direct 32 bit */