Re: RFC: Petition Intel/AMD to add POPF_IF insn

[Denys Vlasenko]

On 08/18/2016 07:24 PM, Linus Torvalds wrote:
> That said, your numbers really aren't very convincing. If popf really
> is just 10 cycles on modern Intel hardware, it's already fast enough
> that I really don't think it matters.

It's 20 cycles. I was wrong in my email, I forgot that the insn count
also counts "push %ebx" insns.

Since I already made a mistake, let me double-check.

200 million iterations of this loop execute under 17 seconds:

  400100:	b8 00 c2 eb 0b       	mov    $0xbebc200,%eax # 1000*1000*1000
  400105:	9c                   	pushfq
  400106:	5b                   	pop    %rbx
  400107:	90                   	nop
....
0000000000400140 <loop>:
  400140:	53                   	push   %rbx
  400141:	9d                   	popfq
  400142:	53                   	push   %rbx
  400143:	9d                   	popfq
  400144:	53                   	push   %rbx
  400145:	9d                   	popfq
  400146:	53                   	push   %rbx
  400147:	9d                   	popfq
  400148:	53                   	push   %rbx
  400149:	9d                   	popfq
  40014a:	53                   	push   %rbx
  40014b:	9d                   	popfq
  40014c:	53                   	push   %rbx
  40014d:	9d                   	popfq
  40014e:	53                   	push   %rbx
  40014f:	9d                   	popfq
  400150:	53                   	push   %rbx
  400151:	9d                   	popfq
  400152:	53                   	push   %rbx
  400153:	9d                   	popfq
  400154:	53                   	push   %rbx
  400155:	9d                   	popfq
  400156:	53                   	push   %rbx
  400157:	9d                   	popfq
  400158:	53                   	push   %rbx
  400159:	9d                   	popfq
  40015a:	53                   	push   %rbx
  40015b:	9d                   	popfq
  40015c:	ff c8                	dec    %eax
  40015e:	75 e0                	jne    400140 <loop>

The loop is exactly 32 bytes, aligned.
There are 14 POPFs. Other insns are very fast.

No perf, just "time taskset 1 ./test".
My CPU frequency hovers around 3500 MHz when loaded.

17 seconds is 17*3500 million cycles.
17*3500 million cycles / 200*14 million cycles = 21.25

Thus, one POPF in CPL3 is ~20 cycles on Skylake.