On Sat, Mar 4, 2017, at 04:00 PM, Bart Schaefer wrote:
> Although the following might help in some cases -- (v->start < 0) can't
> be true very often or arrlen() would have shown up nearly as much as
> arrlen_le(), but maybe this shaves a few unnecessary recounts?
> 
> Anyway changing from arrlen_le() to arrlen_gt() removes a couple of
> tail calls, so this might get faster, and will also reveal how many
> of those arrlen_le() are coming through here.

Time is still the same, 25 s. The zrealloc-patch that was pushed helped,
before: 37 s, after: the 25 s. Hashes are 17 s. Attached callgrind
result from with-patch current-head run:

29,965,124,064  arrlen_le [/usr/local/bin/zsh-5.3.1-dev-0_O2_bopt]
29,965,124,049  arrlen_gt [/usr/local/bin/zsh-5.3.1-dev-0_O2_bopt]
 6,700,764,538  itype_end [/usr/local/bin/zsh-5.3.1-dev-0_O2_bopt]
 6,668,004,434  arrlen [/usr/local/bin/zsh-5.3.1-dev-0_O2_bopt]
 4,552,766,774  _platform_memmove$VARIANT$Nehalem
 [/usr/lib/system/libsystem_platform.dylib]
 2,677,827,871  stringsubst [/usr/local/bin/zsh-5.3.1-dev-0_O2_bopt]

-- 
Sebastian Gniazdowski
psprint3@fastmail.com