[Ada] Implement part of System.Fat_Gen more efficiently

[Ada] Implement part of System.Fat_Gen more efficiently

[Pierre-Marie de Rodat]

[-- Attachment #1: Type: text/plain, Size: 1030 bytes --]

This reimplements the low-level primitives of System.Fat_Gen, i.e.
Copy_Sign, Decompose and Scaling more efficiently, by using the same
direct approach as the existing implementation of Valid; only the
normalization of denormalized numbers and the gradual underflow are
still left to the hardware.

This also speeds up Pred and Succ for denormalized numbers.

Tested on x86_64-pc-linux-gnu, committed on trunk

gcc/ada/

	* libgnat/s-fatgen.ads (Compose): Add pragma Inline.
	(Copy_Sign): Likewise.
	(Exponent): Likewise.
	(Fraction): Likewise.
	* libgnat/s-fatgen.adb: Remove with clause for System, add
	with and use clauses for System.Unsigned_Types.
	Add pragma Warnings (Off) for non-static constants.
	Remove precomputed tables of powers of radix and add a few
	constants describing the floating-point format.
	(Gradual_Scaling): Delete.
	(Copy_Sign): Reimplement directly.
	(Decompose): Likewise.
	(Scaling): Likewise.
	(Pred): Speed up.
	(Succ): Likewise.
	(Truncation): Tidy up.
	(Valid): Move constants to library level.

[-- Attachment #2: patch.diff.gz --]
[-- Type: application/gzip, Size: 7173 bytes --]