Re: In PG12, query with float calculations is slower than PG11 - Mailing list pgsql-hackers

Hi,

On 2020-02-06 11:03:51 -0500, Tom Lane wrote:
> Andres seems to be of the opinion that the compiler should be willing
> to ignore the semantic requirements of the C standard in order
> to rearrange the code back into the cheaper order.  That sounds like
> wishful thinking to me ... even if it actually works on his compiler,
> it certainly isn't going to work for everyone.

Sorry, but, uh, what are you talking about?  Please tell me which single
standards violation I'm advocating for?

I was asking about the inlining bit because the first email of the topic
explained that as the problem, which I don't believe can be the full
explanation - and it turns out it isn't. As Amit Langote's followup
email explained, there's the whole issue of the order of checks being
inverted - which is clearly bad. And wholly unrelated to inlining.

And I asked about __isinf() being used because there are issues with
accidentally ending up with the non-intrinsic version of isinf() when
not using gcc, due to badly written standard library headers.


> The patch looks unduly invasive to me, but I think that it might be
> right that we should go back to a macro-based implementation, because
> otherwise we don't have a good way to be certain that the function
> parameter won't get evaluated first.

I'd first like to see some actual evidence of this being a problem,
rather than just the order of the checks.


> (Another reason to do so is so that the file/line numbers generated
> for the error reports go back to being at least a little bit useful.)
> We could use local variables within the macro to avoid double evals,
> if anyone thinks that's actually important --- I don't.

I don't think that's necessarily a good idea. In fact, I think we should
probably do the exact opposite, and move the error messages further out
of line. All these otherwise very small functions having their own
ereports makes them much bigger. Our low code density, and the resulting
rate of itlb misses, is pretty significant cost (cf [1]).

master:
   text       data        bss        dec        hex    filename
  36124         44         65      36233       8d89    float.o
error messages moved out of line:
   text       data        bss        dec        hex    filename
  32883         44         65      32992       80e0    float.o

Taking int4pl as an example - solely because it is simpler assembly to
look at - we get:

master:
   0x00000000004ac190 <+0>:    mov    0x30(%rdi),%rax
   0x00000000004ac194 <+4>:    add    0x20(%rdi),%eax
   0x00000000004ac197 <+7>:    jo     0x4ac19c <int4pl+12>
   0x00000000004ac199 <+9>:    cltq
   0x00000000004ac19b <+11>:    retq
   0x00000000004ac19c <+12>:    push   %rbp
   0x00000000004ac19d <+13>:    lea    0x1a02c4(%rip),%rsi        # 0x64c468
   0x00000000004ac1a4 <+20>:    xor    %r8d,%r8d
   0x00000000004ac1a7 <+23>:    lea    0x265da1(%rip),%rcx        # 0x711f4f <__func__.26823>
   0x00000000004ac1ae <+30>:    mov    $0x30b,%edx
   0x00000000004ac1b3 <+35>:    mov    $0x14,%edi
   0x00000000004ac1b8 <+40>:    callq  0x586060 <errstart>
   0x00000000004ac1bd <+45>:    lea    0x147e0e(%rip),%rdi        # 0x5f3fd2
   0x00000000004ac1c4 <+52>:    xor    %eax,%eax
   0x00000000004ac1c6 <+54>:    callq  0x5896a0 <errmsg>
   0x00000000004ac1cb <+59>:    mov    $0x3000082,%edi
   0x00000000004ac1d0 <+64>:    mov    %eax,%ebp
   0x00000000004ac1d2 <+66>:    callq  0x589540 <errcode>
   0x00000000004ac1d7 <+71>:    mov    %eax,%edi
   0x00000000004ac1d9 <+73>:    mov    %ebp,%esi
   0x00000000004ac1db <+75>:    xor    %eax,%eax
   0x00000000004ac1dd <+77>:    callq  0x588fb0 <errfinish>

out-of-line error:
   0x00000000004b04e0 <+0>:    mov    0x30(%rdi),%rax
   0x00000000004b04e4 <+4>:    add    0x20(%rdi),%eax
   0x00000000004b04e7 <+7>:    jo     0x4b04ec <int4pl+12>
   0x00000000004b04e9 <+9>:    cltq
   0x00000000004b04eb <+11>:    retq
   0x00000000004b04ec <+12>:    push   %rax
   0x00000000004b04ed <+13>:    callq  0x115e17 <out_of_range_err>

With the out-of-line error, we can fit multiple of these functions into one
cache line. With the inline error, not even one.

Greetings,

Andres Freund

[1] https://twitter.com/AndresFreundTec/status/1214305610172289024