Re: [PATCH 3/5] [ifcvt] optimize x=c ? (y AND z) : y by RISC-V Zicond like insns

[Jeff Law <jeffreyalaw@gmail.com>]

On 12/5/23 01:12, Fei Gao wrote:
> Take the following case for example.
> 
> CFLAGS: -march=rv64gc_zbb_zicond -mabi=lp64d -O2
> 
> long
> test_AND_ceqz (long x, long y, long z, long c)
> {
>    if (c)
>      x = y & z;
>    else
>      x = y;
>    return x;
> }
> 
> Before patch:
> 
> and	a2,a1,a2
> czero.eqz	a0,a2,a3
> czero.nez	a3,a1,a3
> or	a0,a3,a0
> ret
> 
> After patch:
> and	a0,a1,a2
> czero.nez	a1,a1,a3
> or	a0,a1,a0
FWIW I was having a conversation with Andrew W. a little while ago and 
his preference was to change the IOR into an ADD.  We have a better 
chance of using a compressed variant as c.add allows the full set of 
registers while c.or only allows a subset of registers.  Given one of 
the two input registers in that IOR will always have the value zero, it 
should be equivalent.

Given this is target independent code we have to be careful, but I'm not 
immediately aware of a target where using ADD would be worse than IOR 
here.   Anyway, feel free to submit a patch for that minor change, it'll 
need to queue for gcc-15, but should be non-controversial at that time.

The same trick can be done in the conditional move expander within riscv.cc.




> 
> Co-authored-by: Xiao Zeng<zengxiao@eswincomputing.com>
> 
> gcc/ChangeLog:
> 
> 	* ifcvt.cc (noce_cond_zero_binary_op_supported): Add support for AND.
>          (noce_bbs_ok_for_cond_zero_arith): Likewise.
>          (noce_try_cond_zero_arith): Likewise.
> 
> gcc/testsuite/ChangeLog:
> 
> 	* gcc.target/riscv/zicond_ifcvt_opt.c: add TCs for AND.
This looks fine.  I'll need to adjust the matches in the test file 
because we're not supporting SUBREGs.  I can do that easy enough.

I 3-stage bootstrapped and regression tested this on x86_64 as well as 
regression tested it in rv64gc.

I'll push it to the trunk shortly.

jeff