search

danvk / source-map-explorer

Analyze and debug space usage through source maps
Apache License 2.0
3.82k stars 100 forks source link

gzip size of modules should be calculated in bulk #177

Closed connorjclark closed 4 years ago

connorjclark commented 4 years ago

impl: https://github.com/danvk/source-map-explorer/blob/f408a40a46a07f01b7b10c7a49ca19133bfe2c40/src/explore.ts#L265-L280

My concern is that calculating the gzip size of each mapping in isolation misses out on compression savings from common patterns. I expect this means that --gzip overestimates the gzip size of each module. Though I'm not sure what mergeRanges does, perhaps it is grouping everything?

(came up in https://github.com/GoogleChrome/lighthouse/issues/10476)

nikolay-borzov commented 4 years ago

I expected this kind of issue. Getting 100% precise module gzip size seems almost impossible - a module is just part of a file. And gzip applies to the whole file. I still wonder what use of knowing the gzip size of a module when it's bundled with other modules. Merging range purpose is to collect the biggest string that is part of a module. I haven't tested it enough but I think in most cases it should include the whole module code. You can see actual bundle gzip size in totalBytes property

connorjclark commented 4 years ago

I still wonder what use of knowing the gzip size of a module when it's bundled with other modules.

Our specific use case is we want to minify the download time for a bundle. Because of gzip, a large JSON module is probably fewer bytes than an equivalently sized JS module. It's nice to get a rough sense (understanding that exact numbers are impossible) so that we know where to invest engineering effort.

Merging range purpose is to collect the biggest string that is part of a module. I haven't tested it enough but I think in most cases it should include the whole module code.

if this is true then I think this tool is close enough. I see a comment saying it combines continuous mappings of the same source - but I think mappings tend to be continuous for an entire module so that sounds good to me.