RE: [PATCH 2/2] tree-scalar-evolution: Handle toggle recurrences

> -----Original Message-----
> From: Jeffrey Law <[email protected]>
> Sent: Thursday, August 6, 2026 11:22 AM
> To: Liu, Hongtao <[email protected]>; [email protected]
> Cc: [email protected]
> Subject: Re: [PATCH 2/2] tree-scalar-evolution: Handle toggle recurrences
> 
> 
> 
> On 8/4/2026 8:45 PM, liuhongt wrote:
> > Recognize NEGATE_EXPR, BIT_NOT_EXPR and BIT_XOR_EXPR recurrences.
> > Fold constant iteration counts and use a conditional for symbolic counts.
> >
> > Require may_be_zero to be false, since otherwise niter does not
> > necessarily give the latch count.  Keep trapping signed negations
> > unchanged.  This replaces the old invariant bit-operation helper;
> > idempotent AND and IOR updates are handled by the preceding patch.
> >
> > gcc/ChangeLog:
> >
> > 	PR middle-end/124460
> > 	PR middle-end/114502
> > 	PR middle-end/112104
> > 	PR middle-end/98909
> > 	* tree-scalar-evolution.cc
> > 	(analyze_and_compute_bitop_with_inv_effect): Remove.
> > 	(compute_toggle_loop_value): New function.
> > 	(final_value_replacement_loop): Use it.
> >
> > gcc/testsuite/ChangeLog:
> >
> > 	PR middle-end/124460
> > 	PR middle-end/114502
> > 	PR middle-end/112104
> > 	PR middle-end/98909
> > 	* gcc.dg/tree-ssa/pr124460-4.c: New test.
> > 	* gcc.dg/tree-ssa/pr124460-5.c: Likewise.
> > 	* gcc.target/i386/pr105735-1.c: Adjust scan count.
> > 	* gcc.target/i386/pr105735-3.c: Likewise.
> > 	* gcc.target/i386/pr92080-12.c: Keep the XOR operand variant.
> This will need slight readjustment as I think we're better off with Rachit's patch
> for the idempotent cases.
> 
> It's been eons, but there was a paper, I think from the team at Rice, "Beyond
> Induction Variables" or something like that which described a bunch of these
> cases.  Did you happen to review that paper to see if there are any other
> toggle-like cases we should support?
The patch replaces the old function with a new one, git diff looks not easy to read. It's better to just look at the new function directly after applying the patch.
I haven't read any papers on this — those cases all came from bug reports in Bugzilla. I do notice there are opportunities for GCC to optimize when building itself.
> Jeff