[PATCHv2, rs6000] Add minimum __float128 built-in support required for glibc

[Bill Schmidt]

Thanks, I'll make these changes and re-spin.  Not sure what was up
with my tabs...

> On Jun 23, 2016, at 6:49 PM, Segher Boessenkool <segher@kernel.crashing.org> wrote:
> 
> Hi Bill,
> 
> Some little things about the patch...
> 
> On Thu, Jun 23, 2016 at 04:44:27PM -0500, Bill Schmidt wrote:
>> We no longer have a half-clever implementation to construct an infinity
>> inside vector registers, or the full-clever one that Segher proposed in
>> response. :)  We can try to add that support later if desired.
> 
> For posterity:
> 
> Use   vspltisw A,N ; vsrw B,A,A ; vslo D,B,A   to create in D:
> N=-16  ffff_0000_0000_0000_0000_0000_0000_0000  (ieee128 -Inf)
> N=-17  7fff_0000_0000_0000_0000_0000_0000_0000  (ieee128 +Inf)

For the latter, it was N=-15 IIRC?  Anyway -17 is an illegal value for vspltisw.

> 
>> @@ -35569,6 +35639,7 @@ static struct rs6000_opt_mask const rs6000_builtin
>>   { "hard-dfp",		 RS6000_BTM_DFP,	false, false },
>>   { "hard-float",	 RS6000_BTM_HARD_FLOAT,	false, false },
>>   { "long-double-128",	 RS6000_BTM_LDBL128,	false, false },
>> +  { "float128",          RS6000_BTM_FLOAT128,   false, false },
> 
> The previous entries use tabs for indentation.
> 
>> --- gcc/config/rs6000/rs6000.h	(revision 237619)
>> +++ gcc/config/rs6000/rs6000.h	(working copy)
>> @@ -2689,6 +2689,7 @@ extern int frame_pointer_needed;
>> #define RS6000_BTM_HARD_FLOAT	MASK_SOFT_FLOAT	/* Hardware floating point.  */
>> #define RS6000_BTM_LDBL128	MASK_MULTIPLE	/* 128-bit long double.  */
>> #define RS6000_BTM_64BIT	MASK_64BIT	/* 64-bit addressing.  */
>> +#define RS6000_BTM_FLOAT128     MASK_P9_VECTOR  /* IEEE 128-bit float.  */
> 
> Here, too.
> 
>> @@ -2705,7 +2706,8 @@ extern int frame_pointer_needed;
>> 				 | RS6000_BTM_CELL			\
>> 				 | RS6000_BTM_DFP			\
>> 				 | RS6000_BTM_HARD_FLOAT		\
>> -				 | RS6000_BTM_LDBL128)
>> +				 | RS6000_BTM_LDBL128                   \
>> +				 | RS6000_BTM_FLOAT128)
> 
> And here.  And more later.  Let's try to stick to one style, at least
> locally.
> 
>> --- gcc/config/rs6000/rs6000.md	(revision 237619)
>> +++ gcc/config/rs6000/rs6000.md	(working copy)
>> @@ -13326,7 +13326,25 @@
>>    "xssqrtqp %0,%1"
>>   [(set_attr "type" "vecdiv")])
>> 
>> -(define_insn "copysign<mode>3"
>> +(define_expand "copysign<mode>3"
>> +  [(use (match_operand:IEEE128 0 "altivec_register_operand" ""))
>> +   (use (match_operand:IEEE128 1 "altivec_register_operand" ""))
>> +   (use (match_operand:IEEE128 2 "altivec_register_operand" ""))]
> 
> The "" is not needed.
> 
>> +  "FLOAT128_IEEE_P (<MODE>mode)"
>> +{
>> +  if (TARGET_FLOAT128_HW)
>> +    emit_insn (gen_copysign<mode>3_hard (operands[0], operands[1],
>> +    	                                 operands[2]));
> 
> Tabbing here...
> 
>> +  else
>> +    {
>> +      rtx tmp = gen_reg_rtx (<MODE>mode);
>> +      emit_insn (gen_copysign<mode>3_soft (operands[0], operands[1],
>> +      		                           operands[2], tmp));
> 
> ... and here is completely broken.
> 
>> @@ -13336,6 +13354,18 @@
>>    "xscpsgnqp %0,%2,%1"
>>   [(set_attr "type" "vecsimple")])
>> 
>> +(define_insn "copysign<mode>3_soft"
>> +  [(set (match_operand:IEEE128 0 "altivec_register_operand" "=v")
>> +	(unspec:IEEE128
>> +	 [(match_operand:IEEE128 1 "altivec_register_operand" "v")
>> +	  (match_operand:IEEE128 2 "altivec_register_operand" "v")
>> +	  (match_operand:IEEE128 3 "altivec_register_operand" "+v")]
>> +	 UNSPEC_COPYSIGN))]
>> +  "!TARGET_FLOAT128_HW && FLOAT128_IEEE_P (<MODE>mode)"
>> +   "xscpsgndp %x3,%x2,%x1\n\txxpermdi %x0,%x3,%x1,1"
> 
> Two machine insns in a template should be separated by \; not \n\t .
> 
>> +Additional built-in functions are available for the 64-bit PowerPC
>> +family of processors, for efficient use of 128-bit floating point
>> +(@code{__float128}) values.
>> +
>> +The following floating-point built-in functions are always available.  All
>> +of them implement the function that is part of the name.
> 
> "Always"?  Not just with -mfloat128?  And it needs VMX?

Sorry, pasto from the 386 docs.  I meant to change that, sorry.

Bill

> 
> 
> Segher
>