| Issue |
240939
|
| Summary |
Post-RA scheduler undoes the optimized pre-RA schedule
|
| Labels |
backend:RISC-V,
mi-sched
|
| Assignees |
|
| Reporter |
pgodeq
|
The pre-RA schedule at https://godbolt.org/z/sfcxTnY1d , correctly inserts independent instructions between load and it's use, given that the load latency of RISC-V generic CPU is 4 (as per debug logs from machine scheduler).
```
xor a0, a0, s2
zext.b a0, a0
sh2add a0, a0, s3
-> lw a0, 0(a0) // load
srli a1, s2, 8
-> xor s2, a0, a1 // use
addi s1, s1, -1
sw s2, 0(s0)
```
However, the post-RA hoists these instructions before load, which incurs performance penalty as load and it's use are scheduled consecutively. https://godbolt.org/z/jT38hfnoz
```
xor a0, a0, s2
srli a1, s2, 8 // hoisted here.
zext.b a0, a0
addi s1, s1, -1
sh2add a0, a0, s3
-> lw a0, 0(a0) // load
-> xor s2, a0, a1 // use
sw s2, 0(s0)
```
Could someone please help and share if this can be corrected or I am missing something ?
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs