[llvm] [CodeGen][PreISelIntrinsicLowering] Add VP-based lowering for memcpy/memmove/memset (PR #165585)

[David Del Río via llvm-commits]

dadra-oc wrote:

As it stands, the expansion is triggered when the target supports scalable vectors and either the new option flag is enabled or the corresponding libcall is set to nullptr in TLI.

I could modify that condition to check if the target machine is RISC-V. Another option would be to introduce a virtual method in TM that returns a target-specific preference which in the case of RISC-V could be the value of  a command-line flag.

Let me know which approach you'd prefer.

Regarding the performance evaluation, I do not have any RISC-V hardware that I could use right now and I am basically using gem5 for evaluating small cases and qemu for validation. Running llvm tests-suite or SPEC takes foreverer in gem5. 

However, I can give you some numbers for the small tests I have used. For this tests I mesured the number of simulated cycles for the region of code where the call is. On the case of the memmove, I have evaluated with an overlap of half of the elements to be copied and triggering both cases backward and forward copy. 

The gem5 model I have used is set to have vector registers of 512b and size in the table is number of int32_t to be copied. I hope these number at least gives an idea of what could be the actual performance impact.

memcpy:
| Size       | Libcall    | VP-Expansion |
|----------|----------|----------|
| 2  | 328000  | 17000   |
| 4  | 326500 | 17000   |
| 8  |  327000  | 17000  |
| 16  | 328500 | 17000 |
| 32  | 331000  | 33500  |
| 64  |  427000 | 60500 |
| 128  |  555500 | 89000 |

memmove (Forward):
| Size       | Libcall    | VP-Expansion |
|----------|----------|----------|
| 2  | 66500  | 24000  |
| 4  | 174000 | 24000  |
| 8  |  174500  | 24500 |
| 16  | 175500| 24500 |
| 32  | 177500  | 38000  |
| 64  | 181500 | 64500 |
| 128  | 189500 | 89000 |

memmove (Backward):
| Size       | Libcall    | VP-Expansion |
|----------|----------|----------|
| 2  | 114500  | 19000  |
| 4  | 236500 | 19000  |
| 8  |  266500  | 19500 |
| 16  | 236500| 21000 |
| 32  | 238500  | 34500  |
| 64  | 242500 | 60500 |
| 128  | 248500 | 96000 |

memset:
| Size       | Libcall    | VP-Expansion |
|----------|----------|----------|
| 2  | 90000  | 17000  |
| 4  | 98500 | 19000  |
| 8  |  106000  | 19500 |
| 16  | 98000 | 17500 |
| 32  | 114000  | 30500  |
| 64  | 129500 | 56500 |
| 128  | 195500 | 107000 |




https://github.com/llvm/llvm-project/pull/165585