On Wed, 23 Jan 2013, Richard Sandiford wrote:
> > Index: config/mips/gnu-user.h
> > ===================================================================
> > --- config/mips/gnu-user.h (revision 195351)
> > +++ config/mips/gnu-user.h (working copy)
> > @@ -137,3 +137,12 @@ extern const char *host_detect_local_cpu (int argc
> > #define ENDFILE_SPEC \
> > GNU_USER_TARGET_MATHFILE_SPEC " " \
> > GNU_USER_TARGET_ENDFILE_SPEC
> > +
> > +#undef SUBTARGET_OVERRIDE_OPTIONS
> > +#define SUBTARGET_OVERRIDE_OPTIONS \
> > +do { \
> > + /* microMIPS PLT entries are non-microMIPS. */ \
> > + TARGET_INTERLINK_COMPRESSED = 1; \
> > +} while (0)
> > +
> > +
>
> Hmm, that sounds like a reason to set TARGET_INTERLINK_UNCOMPRESSED
> rather than TARGET_INTERLINK_COMPRESSED. I.e. we need to avoid direct
> branches from microMIPS code to standard PLTs.
>
> But that means that microMIPS code can't even jump directly to functions
> that have a micromips attribute when the call might go via a PLT.
> TARGET_INTERLINK_(UN)COMPRESSED doesn't cover that case. I think instead
> we need to handle it directly in mips_function_ok_for_sibcall,
> keyed off TARGET_ABICALLS_PIC0. Specific suggestion below.
[...]
> > /* We can't do a sibcall if the called function is a MIPS16 function
> > because there is no direct "jx" instruction equivalent to "jalx" to
> > switch the ISA mode. We only care about cases where the sibling
> > and normal calls would both be direct. */
> > if (decl
> > - && mips_use_mips16_mode_p (decl)
> > + && ((compression_mode & (MASK_MIPS16 | MASK_MICROMIPS)) != 0)
> > && const_call_insn_operand (XEXP (DECL_RTL (decl), 0), VOIDmode))
> > return false;
> >
> > - /* When -minterlink-mips16 is in effect, assume that non-locally-binding
> > - functions could be MIPS16 ones unless an attribute explicitly tells
> > + /* When -minterlink-compressed is in effect, assume that
> > non-locally-binding
> > + functions could be MIPS16 or microMIPS unless an attribute explicitly
> > tells
> > us otherwise. */
> > - if (TARGET_INTERLINK_MIPS16
> > + if (TARGET_INTERLINK_COMPRESSED
> > && decl
> > && (DECL_EXTERNAL (decl) || !targetm.binds_local_p (decl))
> > - && !mips_nomips16_decl_p (decl)
> > + && (compression_mode & MASK_MICROMIPS) == 0
> > + && (compression_mode & MASK_MIPS16) == 0
> > && const_call_insn_operand (XEXP (DECL_RTL (decl), 0), VOIDmode))
> > return false;
>
> should be something like:
>
> /* Direct Js are only possible to functions that use the same ISA encoding.
> There is no JX counterpart of JALX. */
> if (decl
> && const_call_insn_operand (XEXP (DECL_RTL (decl), 0), VOIDmode)
> && mips_call_may_need_jalx_p (decl))
> return false;
>
> with:
>
> /* Return true if a call to DECL may need to use JALX. */
>
> static bool
> mips_call_may_need_jalx_p (decl)
> {
> /* If the current translation unit would use a different mode for DECL,
> assume that the call needs JALX. */
> if (mips_get_compress_mode (decl) != TARGET_COMPRESSION)
> return true;
>
> /* mips_get_compress_mode is always accurate for locally-binding
> functions in the current translation unit. */
> if (!DECL_EXTERNAL (decl) && targetm.binds_local_p (decl))
> return false;
>
> /* PLTs use the standard encoding, so calls that might go via PLTs
> must use JALX. */
> if (TARGET_COMPRESSED
> && TARGET_ABICALLS_PIC0
> && DECL_EXTERNAL (decl)
> && !targetm.binds_local_p (decl))
> return true;
>
> /* When -minterlink-compressed is in effect, assume that functions
> could use a different encoding mode unless an attribute explicitly
> tells us otherwise. */
> if (TARGET_INTERLINK_COMPRESSED)
> {
> if (!TARGET_COMPRESSION && mips_get_compress_off_flags (decl) == 0)
> return true;
> if (TARGET_COMPRESSION && mips_get_compress_on_flags (decl) == 0)
> return true;
> }
>
> return false;
> }
>
> The TARGET_ABICALLS_PIC0 case should deal with the gnu-user.h comment above.
How about instead of complicating this we simply add support for
microMIPS encoding in the PLT? I think I should be able to squeeze out
some time next week to dust off and retest the binutils patch I've had
pending far too long now. This way we won't have to maintain separate
cases where tail calls may or may not be made via the PLT.
Note that we need that support sooner or later anyway due to the prospect
of pure-microMIPS processors.
Maciej
