As reported by Peter Breitenlohner:

I think this is a very bad idea because ENV is used to specify a startup
file to be read by some/all shells.

Thanks to Khaled for spotting it.

Now, it looks like this breaks Cygwin again.  Have to figure out what's
going on.

Looks like Cygwin / MSVC declare it that way, and it still works on
MinGW32 cross.

This reverts commit 0a492357.

Enables code on more compilers.

Based on fontconfig patch from Raimund Steger.

Sample crash:

0  _hb_graphite2_shaper_face_data_destroy (data=0xffffffffffffffff)
    at ../../src/hb-graphite2.cc:129
1  0x00007ffff4271d7d in hb_graphite2_shaper_face_data_ensure (
    face=<optimized out>) at ../../src/hb-shaper-list.hh:35
2  hb_shape_plan_plan (shaper_list=<optimized out>, num_user_features=0,
    user_features=0x0, shape_plan=0xf7b490) at ../../src/hb-shaper-list.hh:35
3  hb_shape_plan_create (face=<optimized out>, props=<optimized out>,
    user_features=0x0, num_user_features=0, shaper_list=<optimized out>)
    at ../../src/hb-shape-plan.cc:108
4  0x00007ffff4272c93 in hb_shape_plan_create_cached (face=0x10cf2b0,
    props=0x11980d8, user_features=0x0, num_user_features=<optimized out>,
    shaper_list=0x0) at ../../src/hb-shape-plan.cc:283

For example: u0c9a u0ccd u0c9a u0ccd with Lohit.  See:

https://bugs.freedesktop.org/show_bug.cgi?id=59118

If in a MarkPos table, a base has no anchor for a particular mark class,
return NULL such that the subsequent subtables get a chance at it.

Test case:
hb-shape ./EBGaramond12-Regular.otf ἂ --features="ss20","smcp"

The previous name was clashing with harfbuzz.old.  There are systems
that need to link both...

Clash-free now again.

Apparently helps with MSVC compilation.

Not sure if this is the most desired behavior.  It's the most easily
defined though.

Previous logic was just wrong.

As reported by Jonathan Kew.

All passing now.

Fixes test-set.c

Fails now.  Fixing.