1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
|
Using STNT1B to implement MOVNTDQ
As the title says. ASIMD only has STNP which doesn't match semantics. Using STNT1B to match semantics.
Although execution latency is not great on A715 for this instruction. (It gets a bit better on Neoverse-V2).
Found with a hot-loop in d3dcore.dll in Proton doing a non-temporal memcpy. Consuming about 1.5% CPU time.
```
1caf3a020 movdqa xmm4, xmmword [r10+rax]
1caf3a026 movntdq xmmword [rcx+rax], xmm4
1caf3a02b add rax, 0x10
1caf3a02f cmp rdx, rax
1caf3a032 ja 0x1caf3a020
```
rdx is 0x1500 in the hot loop that I found
```
0x00007ffac21cc384: adr x0, 0x7ffac21cc380
0x00007ffac21cc388: str x0, [x28, #184]
0x00007ffac21cc38c: ldr q20, [x4, x14, sxtx]
0x00007ffac21cc390: str q20, [x4, x5, sxtx]
=> 0x00007ffac21cc394: add x4, x4, #0x10
0x00007ffac21cc398: sub x26, x6, x4
0x00007ffac21cc39c: eor x27, x6, x4
0x00007ffac21cc3a0: cmp x6, x4
0x00007ffac21cc3a4: cfinv
0x00007ffac21cc3a8: cset x20, cc // cc = lo, ul, last
0x00007ffac21cc3ac: csel x20, x20, xzr, ne // ne = any
0x00007ffac21cc3b0: cbnz x20, 0x7ffac21cc38c
0x00007ffac21cc3b4: b 0x7ffac21cc3f0
0x00007ffac21cc3b8: blr x0
```
|
