split_complex_* flaw

[David Edelsohn]

>>>>> Ian Lance Taylor writes:

Ian> As you know, I originally wrote the PPC64 Linux ABI to be calling
Ian> convention compatible with the 64-bit AIX ABI.  I think that any
Ian> differences between the calling convention on AIX and Linux are the
Ian> result of independent evolution, or are mistakes.

Ian> The AIX documentation I was working from said this:

Ian> - Other scalar values, such as FORTRAN complex numbers, are mapped
Ian> to the number of doublewords required by their size.

Ian> I copied that directly into the PPC64 Linux ABI.  That says pretty
Ian> clearly to me that complex float is represented as a single
Ian> doubleword, since that is all that the size requires.

Ian> So, did AIX change, or was the ABI documentation incorrect?

	The AIX ABI documentation was written for 32-bit AIX.  This
confusion was an oversight.

	I have always interpreted that statement to mean what I know is
correct, reading "their size" to mean the size of a pair of GPRs, not the
type size.  Complex float is a pair of floats, passed in a pair of FPRs
(always doublewords), mapping to two GPRs (in 64-bit mode), so the size is
two doublewords.  I now realize the statement is ambiguous.

	Packing the PSA in 64-bit mode is not efficient and not what IBM
XLC does.

David