[Bug target/125765] `a ? CST0 : (CST0|CST1)` can be handled better for riscv

[Milan.Tripkovic@rt-rk.com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=125765

Milan Tripkovic <Milan.Tripkovic@rt-rk.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |Milan.Tripkovic@rt-rk.com

--- Comment #3 from Milan Tripkovic <Milan.Tripkovic@rt-rk.com> ---
Created attachment 64919
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=64919&action=edit
patchv1

Hello, I added a patch that extends noce_try_cmove_arith and generate correct
ASM for current problem. However, it contains a hard-coded variable,
cost_ignore, because of a costing issue I encountered.
Without variable the optimization doesn't apply because the compiler thinks
this is too expensive at the moment. During the ce1 pass, the cost is
calculated based on RTL, and the compiler can't see future optimizations (like
the combine pass) that will later simplify the RTL, get rid of redundant
instructions, and eventually generate much better assembly.

FOR code:

>int src(int c, int b, int a) {
>  int t = (a ? 0xff10 : 0xff00);
>  return t;
>}

trunk rv64gcb_zicond has cost 20 in ce1

RTL after ce1                         ASM

movdi_64bit                      li a0,65536
mvconst_internal                 addi a0,a0,-208
mvconst_internal                 czero.eqz a0,a0,a2
czero.eqz.did                    li a5,65536
czero.nez.didi                   addi a5,a5,-256
adddi3                           czero.nez a2,a5,a2
movdi_64bit                      add a0,a2,a0
                                 ret


trunk rv64gcb  has cost 20 in ce1

RTL after ce1                         ASM

movdi_64bit                      li a0,65536
mvconst_internal                 addi a0,a0,-256
branchdi                         beq a2,zero,.L2
mvconst_internal                 li a0,65536
movdi_64bit                      addi    a0,a0,-208
                                 ret

rv64gcb_zicond with my change has cost 32 in ce1                                

RTL after ce1                         ASM

movdi_64bit                      li a5,48               
movdi_64bit                      czero.nez a0,a5,a2
movdi_64bit                      li a5,65536
czero.eqz.didi                   addi a5,a5,-208
czero.nez.didi                   xor a0,a0,a5
adddi3                           ret
movdi_64bit
adddi3
xordi3
movdi_64bit
movdi_64bit

in 2nd example my RTL has cost 60 but generates better code.

>long f(long a)
>{
>    long t = 1234567890123;
>    if (a) return t;
>    return t+12;
>}
trunk rv64gcb_zicond

f:
        bne     a0,zero,.L6
        li      a0,18837504
        addi    a0,a0,507
        slli    a0,a0,16
        addi    a0,a0,1239
        ret
.L6:
        li      a0,18837504
        addi    a0,a0,507
        slli    a0,a0,16
        addi    a0,a0,1227
        ret

rv64gcb_zicond with my change

f:
        li      a5,28
        czero.nez       a0,a5,a0
        li      a5,18837504
        addi    a5,a5,507
        slli    a5,a5,16
        addi    a5,a5,1227
        xor     a0,a0,a5
        ret

I have question about this problem is this right way to optimize out or should
i have different approach.