Would've had results sooner - but I had to run an errand.

I added HTML::Template::JIT to the tests. I also added CGI::Ex::Template::XS which I wrote last August - I haven't released it yet because there are two memory leaks in some obscure TT filter operations - and because I haven't seen that much interest. The results are below - I've tried to compact the thing down.

The abreviations are as before but with the additions of
CTX - CGI::Ex::Template::XS
HTJ - HTML::Template::JIT

TT_str        313/s     --   -5% -60% -64% -69% -78% -80% -80% -83% -8
+4% -84% -87% -87% -88% -91% -91% -92% -95% -95%
TTX_str       329/s     5%    -- -58% -62% -67% -77% -79% -79% -82% -8
+3% -83% -86% -87% -88% -90% -91% -91% -95% -95%
TT_compile    778/s   148%  136%   -- -11% -23% -45% -50% -50% -57% -6
+0% -60% -67% -69% -71% -77% -78% -80% -87% -89%
TTX_compile   872/s   178%  165%  12%   -- -14% -39% -44% -44% -51% -5
+5% -55% -63% -65% -68% -74% -76% -77% -86% -87%
HTE_str      1009/s   222%  207%  30%  16%   -- -29% -35% -35% -44% -4
+8% -48% -57% -60% -63% -70% -72% -74% -84% -85%
HT_str       1420/s   354%  332%  83%  63%  41%   --  -9%  -9% -21% -2
+6% -27% -39% -43% -47% -58% -61% -63% -77% -80%
HTE_mem      1556/s   397%  373% 100%  78%  54%  10%   --  -0% -13% -1
+9% -20% -34% -38% -42% -54% -57% -59% -75% -78%
CET_str      1562/s   399%  375% 101%  79%  55%  10%   0%   -- -13% -1
+9% -19% -33% -37% -42% -54% -57% -59% -75% -77%
TextTemplate 1790/s   472%  444% 130% 105%  77%  26%  15%  15%   --  -
+7%  -8% -23% -28% -34% -47% -50% -53% -71% -74%
CTX_str      1930/s   516%  487% 148% 121%  91%  36%  24%  24%   8%   
+--  -0% -17% -23% -28% -43% -47% -50% -69% -72%
HT_compile   1938/s   519%  489% 149% 122%  92%  37%  25%  24%   8%   
+0%   -- -17% -22% -28% -43% -46% -49% -69% -72%
TT_mem       2339/s   647%  611% 201% 168% 132%  65%  50%  50%  31%  2
+1%  21%   --  -6% -13% -31% -35% -39% -62% -66%
CET_compile  2497/s   698%  659% 221% 186% 148%  76%  61%  60%  40%  2
+9%  29%   7%   --  -7% -26% -31% -35% -60% -64%
HT_mem       2696/s   761%  719% 247% 209% 167%  90%  73%  73%  51%  4
+0%  39%  15%   8%   -- -20% -25% -30% -56% -61%
TTX_mem      3377/s   979%  927% 334% 287% 235% 138% 117% 116%  89%  7
+5%  74%  44%  35%  25%   --  -7% -12% -45% -51%
CTX_compile  3612/s  1054%  998% 364% 314% 258% 154% 132% 131% 102%  8
+7%  86%  54%  45%  34%   7%   --  -6% -42% -48%
CET_mem      3836/s  1125% 1066% 393% 340% 280% 170% 147% 146% 114%  9
+9%  98%  64%  54%  42%  14%   6%   -- -38% -45%
HTJ_compile  6193/s  1878% 1782% 696% 610% 514% 336% 298% 296% 246% 22
+1% 219% 165% 148% 130%  83%  71%  61%   -- -11%
CTX_mem      6934/s  2115% 2007% 791% 695% 587% 388% 346% 344% 287% 25
+9% 258% 196% 178% 157% 105%  92%  81%  12%   --
[download]

So - H::T::JIT is very very very fast. Faster than just about anything else. It also is nice because it doesn't have to stay resident in memory to give a speed boost. You just compile the template to c and run it from scratch each time (unlike the CTX_mem test that was faster - but had to keep the compiled template in memory such as you would do with mod_perl). The downside to H::T::JIT is that there is no expression support (I hate to fan the expression/no-expression fire - but most people I work or talk with use expressions in their templates - but we're talking a couple dozen people at most).

Someday - I'd really love to have a module that compiled the templates down to Parrot bytecode - once - and then executed them - or even better JITed them and then executed them. I think Parrot based templates would fly because you could avoid the costly process of jumping from Perl to C and back - you'd stay in the Parrot interpreter.

If we don't do (or even if we did do) the parrot version, then at minimum I'd love to see a "common compiled template" format that was able to suit all of the various template engines. I think my parsed optree comes pretty close to being suitable - but it certainly isn't commonly used. With a little more massaging, my current format could be stored language agnostic as JSON or some other low level type. This would make life easier for the Jemplate guys - and it would also help out the Python guys who are currently trying to make version of TT in python. We could all share a common compiled template between the languages and various template syntaxes and create "the one true executor" in each language - possibly with an extensible c based library used by each language. OK - now I'm rambling.

Not that you'd be interested (but if you are) - it would be relatively easy to create HTML::Template::TT based on HTML::Template and the parse_expr and play_expr methods of CGI::Ex::Template that would offer TT style expression support in HTML::Templates. It could even go a little farther and compile the HTML::Template to the opcode tree that I use - at which point it could be cached as a CGI::Ex::Template pre-compiled template. I have though about this recently as I've been maintaining some HTML::Template::Expr templates. Just a thought.