RE: Ways and Build Tags for Optimisation
I'd like to include three different flavours of my HBase package, with these combinations of optimisation and profiling:
release: optimised (with -O -fvia-C) profiling: optimised and profiled quick: neither
The release and profiling flavours take forever to compile, so it's useful to have a "quick" flavour to speed up the development cycle. The trouble is, any Haskell program using the release flavour of HBase also compiles very slowly, even if it itself does not use -O. If nothing else, this may be due to the huge interface file for the root import:
$ ls -Al HBase.*hi -rw-r--r-- 1 ashley ashley 2117554 May 28 04:04 HBase.hi -rw-r--r-- 1 ashley ashley 2119865 May 28 08:15 HBase.p_hi -rw-r--r-- 1 ashley ashley 72669 May 28 16:20 HBase.q_hi
Wow :-) There is some ancient support for "user-ways" in the compiler, but as far as I know it doesn't currently work. The only ways that the compiler knows about are the built-in ones (profiling, ticky, parallel etc.). I think the best way to achieve what you want is to have two completely separate packages, say hbase and hbase_q. The hbase package can have profiling support, but the hbase_q package need not. Cheers, Simon
In article
<9584A4A864BD8548932F2F88EB30D1C60D17DC7F@TVP-MSG-01.europe.corp.microso
ft.com>,
"Simon Marlow"
-rw-r--r-- 1 ashley ashley 2117554 May 28 04:04 HBase.hi -rw-r--r-- 1 ashley ashley 2119865 May 28 08:15 HBase.p hi -rw-r--r-- 1 ashley ashley 72669 May 28 16:20 HBase.q hi
Wow :-)
It looks like the problem is very data-heavy Unicode property files. For instance, Org.Org.Semantic.HBase.Text.UnicodeNames exports just one value: getCharacterName :: Char -> String Inside the module is an "Array Char String" created from a "[(Char,String)]" that is a long list of Unicode character names. The file is automatically generated from a downloaded data file. For instance:
getCharacterName '\x189F' "MONGOLIAN LETTER MANCHU ALI GALI DDHA"
For some reason, even though only getCharacterName is exported, when optimisation is switched on, the interface file balloons a thousandfold: $ ls -l UnicodeNames.*hi -rw-r--r-- 1 ashley ashley 5854480 May 28 02:49 UnicodeNames.hi -rw-r--r-- 1 ashley ashley 5854497 May 28 06:56 UnicodeNames.p_hi -rw-r--r-- 1 ashley ashley 2385 May 28 15:59 UnicodeNames.q_hi What's the best way to stop this? Is it reasonable to simply switch off profiling just for these few files? Also, I'd like to make all that data disappear when a binary program that doesn't use it is stripped; currently it doesn't. Any ideas? -- Ashley Yakeley, Seattle WA
participants (2)
-
Ashley Yakeley -
Simon Marlow