This is very impressive work!

Are you sure you would want to support both int and float pixel types?

I think the float support is enough.

How should the API be handled?
I think it would make the most sense to keep the pixels as float all the way until the document container, and then that can decide what to do with them.

You are right, this is the only painless way. As for API break - the library is in beta status. API is changed frequently. The latest master is not compatible with the latest numerical version because of new features and removing legacy. It is OK.

Also there are some structures whose name doesn't make sense now, like typed_int and int_int_cache. These also always store pixel sizes so they now use pixel_t. But they are part of the API surface so renaming them would be breaking.

containing_block_context::typed_int is internal type. It can be renamed to typed_pixel for example.
int_int_cache - the same. Can be renamed to ```pixel_cache`` for example.
These names don't make sense now, agree.

It is interesting, how many tests will fail after switching to the floats?

I think the float support is enough.

Cool, should I still keep the pixel_t alias then, or just use float everywhere? I guess it's maybe worth keeping. It's neater, and somebody might want to change it to double in the future, or something.

As for API break - the library is in beta status. API is changed frequently.

Oh, alright, I didn't realize that.

It is interesting, how many tests will fail after switching to the floats?

Yeah, so there are 242 failures right now.
In some of them it looks like there is some extra anti-aliasing. Rather than a clean edge there is a row of pixels that is between the two colors. Which would make sense with floats, but I haven't updated the document containers yet to pixel_t everywhere, so I'm not sure.
There are also tests where text appears to be at a noticeably different place vertically. That's a bit strange, I'll have to investigate it further.

I guess it's maybe worth keeping. It's neater, and somebody might want to change it to double in the future, or something.

You are right, using pixel_t is the best solution.

Yeah, so there are 242 failures right now.

It is good result 😀 I've expected more failed tests. Anyway the container_cairo and litehtml code contains some workarounds for int rounding issues.

@DRKV333, please let me know when you finish with this PR. Don't worry about failed tests, I'll update them by myself.

Cool, thanks. I would still like to at least look through all the test failures, to make sure I didn't break anything. I can probably finish that this weekend.

There is one problem I've found, where border drawing seems to be a bit off, for instance in border-top-width-036.htm:

So, in a lot of test failures, the difference from the expected image is basically imperceptible. Some of these happen because of lines of text being shifted down slightly, due to less rounding error in line height calculations.

Any test that has list markers is also falling. Makers appear a bit smaller now. This is because their size is a third of the line height, which is usually a fraction. The exact size of the marker doesn't really matter anyway, so this should be fine.

Here's a summary of test failures worth looking at:

background-repeat-applies-to-008.htm
border-bottom-width-014.htm
border-bottom-width-036.htm
border-bottom-width-047.htm
border-top-width-014.htm
border-top-width-036.htm
border-top-width-047.htm
c43-center-000.htm
margin-bottom-067.htm
margin-bottom-068.htm
margin-bottom-113.htm
margin-left-067.htm
margin-left-068.htm
padding-left-036.htm
padding-left-047.htm
padding-right-047.htm

These are actually failing, in that the output contains "red pixels". This happens because of weirdness in fractional pixel rendering. e.g. in border-bottom-width-014.htm there is a 1pt red background that is supposed to be fully covered by a 1pt solid black border. That 1pt ends up being 1.333333px, which renders as 1 line of solid pixels, and then a line of 1/3 translucent pixels. The translucent pixels then alpha-blend instead of covering each other, so there is a bit of dark red on the page.

After going trough the spec it looks like this behavior is not correct, and at least borders have to be rounded in a particular way. (https://www.w3.org/TR/css-values-4/#lengths) So I'll still need to fix this. This doesn't apply to heights though, so that test will be even more wrong. In this case implementations get to decide what to do, so I guess just not rounding and letting the container deal with the fractions is still correct. Though in practice browsers do seem to round things to whole pixels, but the exact behavior varies.

c414-flt-wrap-000.htm

See #416, this is failing, thought not directly because of this PR.

border-spacing-047.htm
c544-valgn-002.htm
c544-valgn-003.htm
c544-valgn-004.htm
text-indent-091.htm
text-indent-092.htm
white-space-collapsing-003.htm

These tests use the "Ahem" font. They now have noticeable spaces between characters that weren't there before. I think this is actually correct, it looks similar in other browsers.

font-matching-rule-012.htm

This test says that the size for scalable fonts should be rounded to the nearest whole pixel. I did not find this in the standard though. All that says is that fonts are typically rounded to the nearest whole pixel. (https://www.w3.org/TR/css-fonts-4/#font-style-matching, point 4.4) That is true, that's how other browsers do it, but not rounding should be fine as well.

Ok, I think I'm ready. I've ended up adding some rounding, right before draw operations. This is the way others do it, and it does fix a lot of issues with things being blurry. It also resolves the "red pixel visible" test failures, and reduces the number of failed tests to 375, from the 472 it was before.