https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87104
pipcet at gmail dot com changed:
What |Removed |Added
----------------------------------------------------------------------------
Attachment #44605|0 |1
is obsolete| |
Attachment #44606|0 |1
is obsolete| |
--- Comment #11 from pipcet at gmail dot com ---
Created attachment 44617
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=44617&action=edit
WIP patch
I'm having partial success with this patch, which does two things:
1. Canonicalize to the easier-to-read (X & A) == B form
2. Emit optimized lea-test sequence for (X & A) == B
(1) is straightforward; we need a new routine to test that an integer
masks a contiguous range of bit positions.
(2) is straightforward except for there being three cases: 32-bit,
64-bit, and mixed. Emacs uses the mixed case, where the lower 3 bits
of a 64 bit value are tested.
The strange thing is that code like
int h17(long int i)
{
if ((i & 12) == 12)
return 1;
return 0;
}
does not work. I see this intermediate RTL:
(insn 7 6 8 2 (set (reg:CCZ 17 flags)
(compare:CCZ (and:DI (not:DI (reg/v:DI 86 [ i ]))
(const_int 12 [0xc]))
(const_int 0 [0]))) "h17.c":4 15 {*cmpdi_1}
(expr_list:REG_DEAD (reg:DI 88)
Surely we should be dealing with a canonical form instead? Who's
generating this non-canonical expression, and why?
Is it legal to do something like
{
operands[2] = negate_rtx (operands[2]);
}
in a define_split? That would avoid the need to generate complicated
RTL exprs from C.
Anyway, this all needs more work and much more testing, but it's a start.
