[llvm] [VPlan] Fold reverse(reverse(x)) -> x in simplifyRecipe() (PR #196900)

[Luke Lau via llvm-commits]

https://github.com/lukel97 commented:

Unfortunately this causes some regressions on RISC-V `-march=rva23u64 -O3` with the following C snippet:

```c++
#include <stdint.h>
#include <stddef.h>

void repro(uint16_t *dst, const uint16_t *src, size_t n) {
  for (size_t i = n; i > 0; --i)
    dst[i - 1] = src[i - 1];
}
```

The codegen before:

```asm
.LBB0_4:                                # %vector.body.preheader
	li	a4, 0
	li	a6, 1
	mv	a5, a2
.LBB0_5:                                # %vector.body
                                        # =>This Inner Loop Header: Depth=1
	vsetvli	a7, a5, e8, m1, ta, ma
	not	a3, a4
	add	t0, a2, a3
	sub	t1, a6, a7
	sh1add	a3, t0, a1
	sh1add	a3, t1, a3
	vle16.v	v8, (a3)
	sub	a5, a5, a7
	sh1add	a3, t0, a0
	sh1add	a3, t1, a3
	vse16.v	v8, (a3)
	add	a4, a4, a7
	bnez	a5, .LBB0_5
```

And after:

```asm
.LBB0_4:                                # %vector.ph
	li	a3, 0
	li	a4, 2
	sub	a4, a4, a5
	vsetvli	a5, zero, e32, m4, ta, ma
	vid.v	v8
	addi	a6, a6, -1
	vsetvli	zero, zero, e16, m2, ta, ma
	vid.v	v12
	vrsub.vx	v12, v12, a6
	add	a6, a1, a4
	add	a7, a0, a4
	vsetvli	zero, zero, e8, m1, ta, ma
	vmv.v.i	v14, 0
	mv	a4, a2
.LBB0_5:                                # %vector.body
                                        # =>This Inner Loop Header: Depth=1
	vsetvli	a5, a4, e8, m1, ta, ma
	not	a1, a3
	vsetvli	a0, zero, e32, m4, ta, ma
	vmsltu.vx	v0, v8, a5
	add	a1, a1, a2
	vsetvli	zero, zero, e8, m1, ta, ma
	vmerge.vim	v15, v14, 1, v0
	vrgatherei16.vv	v16, v15, v12
	vmsne.vi	v0, v16, 0
	sh1add	a0, a1, a6
	vle16.v	v16, (a0), v0.t
	sub	a4, a4, a5
	sh1add	a0, a1, a7
	vse16.v	v16, (a0), v0.t
	add	a3, a3, a5
	bnez	a4, .LBB0_5
```

It looks like we reverse the header mask now which prevents the loads from being converted to vp.loads. I'm not sure what the right fix for this, I'll get back to you on this.

https://github.com/llvm/llvm-project/pull/196900