Re: [PATCH] RISC-V: Revert the convert from vmv.s.x to vmv.v.i

[Jeff Law <jeffreyalaw@gmail.com>]

On 2/19/24 21:15, Alexandre Oliva wrote:
> This backport is the first of two required for the pr111935 testcase,
> already backported to gcc-13, to pass on riscv64-elf and riscv32-elf.
> The V_VLS mode iterator, used in the original patch, is not available in
> gcc-13, and I thought that would be too much to backport (and maybe so
> are these two patches, WDYT?), so I changed it to V, to match the
> preexisting gcc-13 pattern.  Regstrapped on x86_64-linux-gnu, along with
> other backports, and tested manually on riscv64-elf.  Ok to install?
> 
> From: Lehua Ding <lehua.ding@rivai.ai>
> 
> Hi,
> 
> This patch revert the convert from vmv.s.x to vmv.v.i and add new pattern
> optimize the special case when the scalar operand is zero.
> 
> Currently, the broadcast pattern where the scalar operand is a imm
> will be converted to vmv.v.i from vmv.s.x and the mask operand will be
> converted from 00..01 to 11..11. There are some advantages and
> disadvantages before and after the conversion after discussing
> with Juzhe offline and we chose not to do this transform.
> 
> Before:
> 
>    Advantages: The vsetvli info required by vmv.s.x has better compatibility since
>    vmv.s.x only required SEW and VLEN be zero or one. That mean there
>    is more opportunities to combine with other vsetlv infos in vsetvl pass.
> 
>    Disadvantages: For non-zero scalar imm, one more `li rd, imm` instruction
>    will be needed.
> 
> After:
> 
>    Advantages: No need `li rd, imm` instruction since vmv.v.i support imm operand.
> 
>    Disadvantages: Like before's advantages. Worse compatibility leads to more
>    vsetvl instrunctions need.
> 
> Consider the bellow C code and asm after autovec.
> there is an extra insn (vsetivli zero, 1, e32, m1, ta, ma)
> after converted vmv.s.x to vmv.v.i.
> 
> ```
> int foo1(int* restrict a, int* restrict b, int *restrict c, int n) {
>      int sum = 0;
>      for (int i = 0; i < n; i++)
>        sum += a[i] * b[i];
> 
>      return sum;
> }
> ```
> 
> asm (Before):
> 
> ```
> foo1:
>          ble     a3,zero,.L7
>          vsetvli a2,zero,e32,m1,ta,ma
>          vmv.v.i v1,0
> .L6:
>          vsetvli a5,a3,e32,m1,tu,ma
>          slli    a4,a5,2
>          sub     a3,a3,a5
>          vle32.v v2,0(a0)
>          vle32.v v3,0(a1)
>          add     a0,a0,a4
>          add     a1,a1,a4
>          vmacc.vv        v1,v3,v2
>          bne     a3,zero,.L6
>          vsetvli a2,zero,e32,m1,ta,ma
>          vmv.s.x v2,zero
>          vredsum.vs      v1,v1,v2
>          vmv.x.s a0,v1
>          ret
> .L7:
>          li      a0,0
>          ret
> ```
> 
> asm (After):
> 
> ```
> foo1:
>          ble     a3,zero,.L4
>          vsetvli a2,zero,e32,m1,ta,ma
>          vmv.v.i v1,0
> .L3:
>          vsetvli a5,a3,e32,m1,tu,ma
>          slli    a4,a5,2
>          sub     a3,a3,a5
>          vle32.v v2,0(a0)
>          vle32.v v3,0(a1)
>          add     a0,a0,a4
>          add     a1,a1,a4
>          vmacc.vv        v1,v3,v2
>          bne     a3,zero,.L3
>          vsetivli        zero,1,e32,m1,ta,ma
>          vmv.v.i v2,0
>          vsetvli a2,zero,e32,m1,ta,ma
>          vredsum.vs      v1,v1,v2
>          vmv.x.s a0,v1
>          ret
> .L4:
>          li      a0,0
>          ret
> ```
> 
> Best,
> Lehua
> 
> Co-Authored-By: Ju-Zhe Zhong <juzhe.zhong@rivai.ai>
> 
> gcc/ChangeLog:
> 
> 	* config/riscv/predicates.md (vector_const_0_operand): New.
> 	* config/riscv/vector.md (*pred_broadcast<mode>_zero): Ditto.
> 
> gcc/testsuite/ChangeLog:
> 
> 	* gcc.target/riscv/rvv/base/scalar_move-5.c: Update.
> 	* gcc.target/riscv/rvv/base/scalar_move-6.c: Ditto.
I wouldn't backport this.  Vector isn't something that's really expected 
to work with gcc-13.  Yea, you can do a bit of intrinsics, but that's it.

Jeff