| Issue |
60571
|
| Summary |
Suboptimal code generation for 256-bit integer addition containing constant bits in arm64
|
| Labels |
backend:AArch64
|
| Assignees |
|
| Reporter |
aqjune
|
Hello all,
It seems llc is producing suboptimal code for 256-bit int add when target is arm64.
https://godbolt.org/z/MfEMGKE8n
The command I used is `llc --mtriple=arm64-unknown-unknown -mcpu=neoverse-n1 -O3` (which is also used in the godbolt instance).
The above code evaluates [x3;x2;x1;x0] + [y3;y2;y1;0x1] and stores the results into the memory.
The generated code is
```
adds x8, x1, x4
adcs x9, x2, x5
adc x10, x3, x6
adds x11, x0, #1
adcs x8, x8, xzr
stp x11, x8, [x7]
adcs x8, x9, xzr
cinc x9, x10, hs
stp x8, x9, [x7, #16]
ret
```
.. which doesn't quite seem optimal.
I think it can be further reduced to:
```
adds x15, x0, #1
adcs x16, x1, x2
adcs x17, x3, x4
adc x18, x5, x6
stp x15, x16, [x30]
stp x17, x18, [x30, #16]
```
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
