newSeqWith is 3 times slower than fill

[mratsim]

Test case:

import math, algorithm, sequtils, times

const
  dz = 0.01
  z = 100
  spaceSteps = int(z / dz)
  timeSteps = 50000

let start = cpuTime()

# var foo = newSeq[float](timeSteps * spaceSteps)
# foo.fill(1.0) # 3.629437

var foo = newSeqWith(timeSteps * spaceSteps, 1.0) # 10.429099

let stop = cpuTime()
echo stop - start

cc @narimiran

[skilchen]

Its not fill that is that much faster. IMHO the slowness of newSeqWith is due to the fact that it is implemented as a template. This causes first an allocation of the sequence and then a copying assignment to the target value.
If newSeqWith is changed to a generic procedure, it is as fast as your version using fill:

proc newSeqWith*[T](len: int, init: T): seq[T] =
  result = newSeq[T](len)
  for i in 0 ..< len:
    result[i] = init

I don't know, why newSeqWith is implemented as a template and not as a generic procedure.

[mratsim]

It's implemented as a template to allow inline proc arguments like "rand" to fill a tensor with random values for example. In your case the whole seq will be filled with the same "init" value.

[cooldome]

New timings:
newSeqWith: 6.849931 sec
fill: 6.800397999999999 sec

[skilchen]

mratsim wrote:
It's implemented as a template to allow inline proc arguments like "rand" to fill a tensor with random values for example. In your case the whole seq will be filled with the same "init" value.

ah, i see. thanks @mratsim for the explanation. this feature is too clever for me ...

[narimiran]

New timings:
newSeqWith: 6.849931 sec
fill: 6.800397999999999 sec

I've just tried it on the latest Nim devel (0.19.9) and:

• in debug mode, newSeqWith is ~30% slower
• in release mode, newSeqWith is ~2x slower

[krux02]

newSeqWith uses a for loop, where fill uses a while loop

[data-man]

Is time for #7572? ;-)

[krux02]

I think those optimizations should stay in the realm of the backend (C(++)) compiler for now. They do this for years, it is well tested.

[data-man]

I want to control the optimization of a particular loop, not rely on some strange heuristic.

[ringabout]

Tested on latest devel version on windows.

With ARC/ORC, I got the similar results for both operations(8.83 vs 8.225).

[timotheecour]

any of these changes would make it identical to fill:

• replacing newSeqWith with a proc
• using --gc:arc|orc
• replacing result by move(result) in last line of the template

the last point is the most useful because it applies to other cases as well (where a template is needed) and doesn't require one to change the gc

=> #18168