Re: lib/51530: ldexp(2.0,INT_MAX) still zero on mips64el

[David Holland <dholland-bugs%netbsd.org@localhost>]

The following reply was made to PR lib/51530; it has been noted by GNATS.

From: David Holland <dholland-bugs%netbsd.org@localhost>
To: gnats-bugs%NetBSD.org@localhost
Cc: 
Subject: Re: lib/51530: ldexp(2.0,INT_MAX) still zero on mips64el
Date: Thu, 6 Oct 2016 03:42:14 +0000

 On Thu, Oct 06, 2016 at 01:30:01AM +0000, coypu%SDF.ORG@localhost wrote:
  >  So:
  >  
  >  MIPS has its own implementation of ldexp: lib/libc/arch/mips/gen/ldexp.S
  >  
  >  
  >       71 	addu	t1, t1, a2		# scale exponent
  >       72 	sll	v0, a2, 20		# position N for addition
  >       73 	bge	t1, DEXP_INF, 8f	# overflow?
  >       74 	addu	v0, v0, v1		# multiply by (2**N)
  >       75 	ble	t1, zero, 4f		# underflow?
  >  
  >  
  >  The exponent will never underflow, because addu is 'add unsigned'.
 
 That's not the problem - the problem is that because the bias
 correction hasn't been applied to the exponent from the float
 argument, INT_MAX added to that exponent becomes negative so it goes
 to the underflow case (which produces 0) instead of the overflow
 case.
 
 One way to fix this is to change the bge to bgeu, but then INT_MIN
 will also be treated as an overflow and that's still wrong.
 
 It needs to apply the bias correction before doing those comparisons.
 (Or compare to bounds that have the bias correction built into them,
 but that will take more instructions.)
 
 It should be sufficient to apply it to t1 after line 71... but then
 the underflow/denormal logic, and anything else subsequent using the
 value of t1, needs to either take it off again or be adjusted to
 match, which is not entirely trivial.
 
 It might be expedient to check for large-magnitude negative values of
 the N argument (ones that will never produce a valid result no matter
 the number applied to, like say < -32768) and return zero directly.
 
 something like
    slti t4, a2, -32768
    bnez t4, 5f
 and
 5:
    mtc1 zero, $f1
    mtc1 zero, $f0
    j ra
 
 except it might want to return -0 for some cases. There's code like
 this at 7:, but it wants the sign of the x argument in t0 which won't
 have happened yet.
 
 (also, this file is too large to be using numbered scratch labels for
 large jumps. those should get names.)
 
 -- 
 David A. Holland
 dholland%netbsd.org@localhost