This is the mail archive of the
mailing list for the GCC project.
Re: Bootstrap failure on sh-linux target
Roger Sayle <email@example.com> wrote:
> It almost certainly is papering over the real bug. The original
> "safe_from_p (target, exp1, 1)" test is removed from that clause
> because that test is currently performed as the very first step
> in expand_operands, which is called immediately after. The only
> functional differences are:
> (i) that in the original case we'll call expand_expr with a
> target that is a new gen_reg_rtx, and currently we call with with
> a target of NULL_RTX.
> (ii) The remaining code in the MIN_EXPR/MAX_EXPR will actually
> return the result in the requested target, if it isn't a memory.
> Both differences should improve the middle-end's RTL generation.
I think you are right. I've found that the bootstrap failure went
away with -fno-unit-at-a-time option. So it seems that the real
problem is in the other part. Perhaps the difference in rtl
expression reveals a latent bug. I'll take a look at the rtl dumps
Thanks for the explanation.