Further fixes on top of 9e2f316ed. Writing some doco for gcc on
inline assembly woke me up to some more errors.
- dgemv_kernel_4x4 asm did not mention *ap as a memory input, and
*y is both read and write.
- sasum_kernel_32 and casum_kernel_16 did not use %x for a vsx insn
operand, a problem if the "=f" sum output was ever allocated a vsx
reg in the altivec set. This might be possible with inlining and
future gcc optimisation.
Quoting patch author amodra from #1078
Lots of issues here.
- The vsx regs weren't listed as clobbered.
- Poor choice of vsx regs, which along with the lack of clobbers led to
trashing v0..v21 and fr14..fr23. Ideally you'd let gcc choose all
temp vsx regs, but asms currently have a limit of 30 i/o parms.
- Other regs were clobbered unnecessarily, seemingly in an attempt to
clobber inputs, with gcc-7 complaining about the clobber of r2.
(Changed inputs should be also listed as outputs or as an i/o.)
- "r" constraint used instead of "b" for gprs used in insns where the
r0 encoding means zero rather than r0.
- There were unused asm inputs too.
- All memory was clobbered rather than hooking up memory outputs with
proper memory constraints, and that and the lack of proper memory
input constraints meant the asms needed to be volatile and their
containing function noinline.
- Some parameters were being passed unnecessarily via memory.
- When a copy of a
Matt Brown
Alan Modra
Martin Kroeker
Werner Saar