Re: [COMMITTED PATCH v2] gdb: mips: Fix the handling of complex type of function return value

[Richard Sandiford <richard.sandiford@arm.com>]

"Maciej W. Rozycki" <macro@orcam.me.uk> writes:
> From: Youling Tang <tangyouling@loongson.cn>
>
> $ objdump -d outputs/gdb.base/varargs/varargs
> 00000001200012e8 <find_max_float_real>:
> ...
>    1200013b8:   c7c10000        lwc1    $f1,0(s8)
>    1200013bc:   c7c00004        lwc1    $f0,4(s8)
>    1200013c0:   46000886        mov.s   $f2,$f1
>    1200013c4:   46000046        mov.s   $f1,$f0
>    1200013c8:   46001006        mov.s   $f0,$f2
>    1200013cc:   46000886        mov.s   $f2,$f1
>    1200013d0:   03c0e825        move    sp,s8
>    1200013d4:   dfbe0038        ld      s8,56(sp)
>    1200013d8:   67bd0080        daddiu  sp,sp,128
>    1200013dc:   03e00008        jr      ra
>    1200013e0:   00000000        nop
>
> From the above disassembly, we can see that when the return value of the
> function is a complex type and len <= 2 * MIPS64_REGSIZE, the return value
> will be passed through $f0 and $f2, so fix the corresponding processing
> in mips_n32n64_return_value().
>
> $ make check RUNTESTFLAGS='GDB=../gdb gdb.base/varargs.exp --outdir=test'
>
> Before applying the patch:
>  FAIL: gdb.base/varargs.exp: print find_max_float_real(4, fc1, fc2, fc3, fc4)
>  FAIL: gdb.base/varargs.exp: print find_max_double_real(4, dc1, dc2, dc3, dc4)
>
>  # of expected passes            9
>  # of unexpected failures        2
>
> After applying the patch:
>  # of expected passes            11
>
> This also fixes:
>  FAIL: gdb.base/callfuncs.exp: call inferior func with struct - returns float _Complex
>
> Signed-off-by: Youling Tang <tangyouling@loongson.cn>
> Co-Authored-By: Maciej W. Rozycki <macro@orcam.me.uk>
> ---
> Hi Youling,
>
>> >   This has to be double-checked, because as I recall we have an ABI bug in
>> > GCC in this area.  Which is also the reason why the relevant test cases
>> > have not been fixed in 15+ years now (I've been aware of this issue).
>> > 
>> >   OTOH, if things have been like this for so long, then I suppose they need
>> > to stay as they are.  In any case I think this does have to be thoroughly
>> > understood and documented.
>> Thanks for your pointing out.If GCC's processing does not follow the ABI
>> call parameter specification, then this will be a GCC bug.GDB will remain
>> as is, without relevant modifications.
>
>  I have now carefully reviewed this issue and while we do have a bug in 
> this area (more on this later), your change is conceptually a move in the 
> right direction except for these issues:
>
> - $f0/$f2 may only be used if we have an FPU, so the `tdep->mips_fpu_type' 
>   check has to qualify complex types just as it does floating-point types,
>
> - single complex types occupy lower halves of $f0/$f2 only.
>
> You haven't stated exactly how you verified your change, but it does not 
> fix:
>
> FAIL: gdb.base/varargs.exp: print find_max_float_real(4, fc1, fc2, fc3, fc4)
>
> in my environment because of the latter issue.
>
>  I went ahead and fixed these issues (in addition to code formatting), and 
> added explanatory notes including real/imaginary part usage in particular. 
> I have verified the updated change with the `mips64-linux-gnu' big-endian 
> target and the n64 ABI using remote `gdbserver' with my Malta 5Kc MIPS64r1 
> system.  I have mentioned the gdb.base/callfuncs.exp progression in the 
> change description too.
>
>  I cannot see you listed in MAINTAINERS, but your original change only had 
> a legally insignificant amount of changes, so I have now committed this 
> updated version.  However for the sake of any future submissions you may 
> make can you please explain your current FSF copyright assignment status?
>
>  Now as to the bug mentioned earlier on it's with `long double' rather 
> than complex data types.  Considering this program:
>
> typedef struct
>   {
>     long double d;
>   }
> ldouble_t;
>
> long double
> v (long double d)
> {
>   return d;
> }
>
> ldouble_t
> s (long double d)
> {
>   return (ldouble_t) { .d = d };
> }
>
> we get this assembly:
>
> 	.file	1 "ldouble.c"
> 	.section .mdebug.abi64
> 	.previous
> 	.nan	legacy
> 	.module	fp=64
> 	.module	oddspreg
> 	.module	arch=mips3
> 	.abicalls
> 	.text
> 	.align	2
> 	.align	3
> 	.globl	v
> 	.set	nomips16
> 	.set	nomicromips
> 	.ent	v
> 	.type	v, @function
> v:
> 	.frame	$sp,0,$31		# vars= 0, regs= 0/0, args= 0, gp= 0
> 	.mask	0x00000000,0
> 	.fmask	0x00000000,0
> 	.set	noreorder
> 	.set	nomacro
> 	dmfc1	$4,$f12
> 	dmfc1	$5,$f13
> 	dmtc1	$4,$f0
> 	dmtc1	$5,$f2
> 	jr	$31
> 	nop
>
> 	.set	macro
> 	.set	reorder
> 	.end	v
> 	.size	v, .-v
> 	.align	2
> 	.align	3
> 	.globl	s
> 	.set	nomips16
> 	.set	nomicromips
> 	.ent	s
> 	.type	s, @function
> s:
> 	.frame	$sp,0,$31		# vars= 0, regs= 0/0, args= 0, gp= 0
> 	.mask	0x00000000,0
> 	.fmask	0x00000000,0
> 	.set	noreorder
> 	.set	nomacro
> 	dmfc1	$2,$f12
> 	dmfc1	$3,$f13
> 	dmtc1	$2,$f0
> 	nop
> 	dmtc1	$3,$f1
> 	jr	$31
> 	nop
>
> 	.set	macro
> 	.set	reorder
> 	.end	s
> 	.size	s, .-s
> 	.ident	"GCC: (GNU) 12.0.1 20220129 (experimental)"
> 	.section	.note.GNU-stack,"",@progbits
>
> so while `v' correctly returns the result in $f0/$f2, `s' returns it in 
> $f0/$f1 instead.  I'll see if I can discuss this with the GCC community 
> once Stage 1 has opened, likely in a couple of weeks' time.
>
>  Additionally I found another issue, which I think is an ABI bug too.  
> Considering this program:
>
> #include <complex.h>
>
> typedef struct
>   {
>     float complex c;
>   }
> complexf_t;
>
> typedef struct
>   {
>     double complex c;
>   }
> complex_t;
>
> float complex
> vf (float complex c)
> {
>   return c;
> }
>
> complexf_t
> sf (float complex c)
> {
>   return (complexf_t) { .c = c };
> }
>
> double complex
> v (double complex c)
> {
>   return c;
> }
>
> complex_t
> s (double complex c)
> {
>   return (complex_t) { .c = c };
> }
>
> we get this assembly:
>
> 	.file	1 "complex.c"
> 	.section .mdebug.abi64
> 	.previous
> 	.nan	legacy
> 	.module	fp=64
> 	.module	oddspreg
> 	.module	arch=mips3
> 	.abicalls
> 	.text
> 	.align	2
> 	.align	3
> 	.globl	vf
> 	.set	nomips16
> 	.set	nomicromips
> 	.ent	vf
> 	.type	vf, @function
> vf:
> 	.frame	$sp,0,$31		# vars= 0, regs= 0/0, args= 0, gp= 0
> 	.mask	0x00000000,0
> 	.fmask	0x00000000,0
> 	.set	noreorder
> 	.set	nomacro
> 	mov.s	$f2,$f13
> 	jr	$31
> 	mov.s	$f0,$f12
>
> 	.set	macro
> 	.set	reorder
> 	.end	vf
> 	.size	vf, .-vf
> 	.align	2
> 	.align	3
> 	.globl	sf
> 	.set	nomips16
> 	.set	nomicromips
> 	.ent	sf
> 	.type	sf, @function
> sf:
> 	.frame	$sp,16,$31		# vars= 16, regs= 0/0, args= 0, gp= 0
> 	.mask	0x00000000,0
> 	.fmask	0x00000000,0
> 	.set	noreorder
> 	.set	nomacro
> 	daddiu	$sp,$sp,-16
> 	swc1	$f12,0($sp)
> 	lwu	$4,0($sp)
> 	mfc1	$3,$f13
> 	daddiu	$sp,$sp,16
> 	dsll	$3,$3,32
> 	jr	$31
> 	or	$2,$4,$3
>
> 	.set	macro
> 	.set	reorder
> 	.end	sf
> 	.size	sf, .-sf
> 	.align	2
> 	.align	3
> 	.globl	v
> 	.set	nomips16
> 	.set	nomicromips
> 	.ent	v
> 	.type	v, @function
> v:
> 	.frame	$sp,0,$31		# vars= 0, regs= 0/0, args= 0, gp= 0
> 	.mask	0x00000000,0
> 	.fmask	0x00000000,0
> 	.set	noreorder
> 	.set	nomacro
> 	mov.d	$f2,$f13
> 	jr	$31
> 	mov.d	$f0,$f12
>
> 	.set	macro
> 	.set	reorder
> 	.end	v
> 	.size	v, .-v
> 	.align	2
> 	.align	3
> 	.globl	s
> 	.set	nomips16
> 	.set	nomicromips
> 	.ent	s
> 	.type	s, @function
> s:
> 	.frame	$sp,0,$31		# vars= 0, regs= 0/0, args= 0, gp= 0
> 	.mask	0x00000000,0
> 	.fmask	0x00000000,0
> 	.set	noreorder
> 	.set	nomacro
> 	dmfc1	$2,$f12
> 	dmfc1	$3,$f13
> 	jr	$31
> 	nop
>
> 	.set	macro
> 	.set	reorder
> 	.end	s
> 	.size	s, .-s
> 	.ident	"GCC: (GNU) 12.0.1 20220129 (experimental)"
> 	.section	.note.GNU-stack,"",@progbits
>
> again `v' and `vf' correctly return the result in $f0/$f2, but `s' holds 
> it in $2/$3 and especially `sf' goes through the hoops to squeeze it into 
> $3 (and the caller then has to unsqueeze it).  My understanding of MIPSpro 
> documentation is in both cases the result shall go into $f0/$f2.  I think 
> this is something to take with GCC as well.
>
>  Richard, would you by any chance know/remember what IRIX/MIPSpro did 
> here?

No, sorry.  I remember there were some cases where we had to choose
between compatibility with MIPSpro and backwards compatibility with
GCC (which did something different and thus presumably “wrong”).
Since GCC de facto the ABI for GNU systems, it wasn't always an
easy choice.  But that might not overlap with this case.

Richard

>
>   Maciej
> ---
>  gdb/mips-tdep.c |   34 ++++++++++++++++++++++++----------
>  1 file changed, 24 insertions(+), 10 deletions(-)
>
> gdb-tangyouling-mips-n64-complex-return.diff
> Index: binutils-gdb/gdb/mips-tdep.c
> ===================================================================
> --- binutils-gdb.orig/gdb/mips-tdep.c
> +++ binutils-gdb/gdb/mips-tdep.c
> @@ -5217,30 +5217,44 @@ mips_n32n64_return_value (struct gdbarch
>       that all composite results be handled by conversion to implicit first
>       parameters.  The MIPS/SGI Fortran implementation has always made a
>       specific exception to return COMPLEX results in the floating point
> -     registers.]  */
> +     registers.]
> +
> +     From MIPSpro Assembly Language Programmer's Guide, Document Number:
> +     007-2418-004
> +
> +              Software
> +     Register Name(from
> +     Name     fgregdef.h) Use and Linkage
> +     -----------------------------------------------------------------
> +     $f0, $f2 fv0, fv1    Hold results of floating-point type function
> +                          ($f0) and complex type function ($f0 has the
> +                          real part, $f2 has the imaginary part.)  */
>  
>    if (TYPE_LENGTH (type) > 2 * MIPS64_REGSIZE)
>      return RETURN_VALUE_STRUCT_CONVENTION;
> -  else if (type->code () == TYPE_CODE_FLT
> -	   && TYPE_LENGTH (type) == 16
> +  else if ((type->code () == TYPE_CODE_COMPLEX
> +	    || (type->code () == TYPE_CODE_FLT && TYPE_LENGTH (type) == 16))
>  	   && tdep->mips_fpu_type != MIPS_FPU_NONE)
>      {
> -      /* A 128-bit floating-point value fills both $f0 and $f2.  The
> -	 two registers are used in the same as memory order, so the
> -	 eight bytes with the lower memory address are in $f0.  */
> +      /* A complex value of up to 128 bits in width as well as a 128-bit
> +	 floating-point value goes in both $f0 and $f2.  A single complex
> +	 value is held in the lower halves only of the respective registers.
> +	 The two registers are used in the same as memory order, so the
> +	 bytes with the lower memory address are in $f0.  */
>        if (mips_debug)
>  	gdb_printf (gdb_stderr, "Return float in $f0 and $f2\n");
>        mips_xfer_register (gdbarch, regcache,
>  			  (gdbarch_num_regs (gdbarch)
>  			   + mips_regnum (gdbarch)->fp0),
> -			  8, gdbarch_byte_order (gdbarch),
> +			  TYPE_LENGTH (type) / 2, gdbarch_byte_order (gdbarch),
>  			  readbuf, writebuf, 0);
>        mips_xfer_register (gdbarch, regcache,
>  			  (gdbarch_num_regs (gdbarch)
>  			   + mips_regnum (gdbarch)->fp0 + 2),
> -			  8, gdbarch_byte_order (gdbarch),
> -			  readbuf ? readbuf + 8 : readbuf,
> -			  writebuf ? writebuf + 8 : writebuf, 0);
> +			  TYPE_LENGTH (type) / 2, gdbarch_byte_order (gdbarch),
> +			  readbuf ? readbuf + TYPE_LENGTH (type) / 2 : readbuf,
> +			  (writebuf
> +			   ? writebuf + TYPE_LENGTH (type) / 2 : writebuf), 0);
>        return RETURN_VALUE_REGISTER_CONVENTION;
>      }
>    else if (type->code () == TYPE_CODE_FLT