WIP Fix some warnings about float conversion (-Wfloat-conversion)

[rbuj]

Test: marco.make.log (master), marco/make.log (PR)

$ CFLAGS="-O3 -fomit-frame-pointer -malign-double -fstrict-aliasing -ffast-math -Wconversion -Wunused-macros -Wunused-parameter" ./autogen.sh --prefix=/usr --enable-compile-warnings=maximum && make &> make.log && sudo make install

Comparing make logs

$ diff <(./warnings-file.sh marco.make.log) <(./warnings-file.sh marco/make.log)
3c3
< -Wconversion 978
---
> -Wconversion 951
5c5
< -Wfloat-conversion 97
---
> -Wfloat-conversion 18

warnings-file.sh

FILE=$1
for warning in $(grep warning $FILE | grep -Po '\[\-W.*\]' | sed 's/[][]//g' | sort -u)
do
 type=${warning#"-W"}
 num=$(echo "grep $type $FILE | wc -l" | sh)
 echo "$warning $num"
done

Testing scaled surfaces: save scaled surfaces to png files "/home/robert/SC_TIME " + TIMESTAMP + ".png"

• Run mate-appearance-properties

• Select Theme tab

  static cairo_surface_t *
  scale_surface (cairo_surface_t *surface,
             gdouble          old_width,
             gdouble          old_height,
             gdouble          new_width,
             gdouble          new_height,
             gboolean         vertical_stripes,
             gboolean         horizontal_stripes)

  
  diff --git a/src/ui/theme.c b/src/ui/theme.c
  index be3e87a..6c7a055 100644
  --- a/src/ui/theme.c
  +++ b/src/ui/theme.c
  @@ -140,6 +140,10 @@ scale_surface (cairo_surface_t *surface,
  cairo_pattern_set_extend (cairo_get_source (cr), CAIRO_EXTEND_PAD);
  
  cairo_paint (cr);

• char *file = g_strdup_printf ("/home/robert/SC_TIME%lu.png", g_get_monotonic_time ());

• cairo_surface_write_to_png (surface, file);

• g_free (file);

• cairo_destroy (cr);

  return scaled;

  
  Testing Alt+Tab thumbnails: save window screenshots to png files "/home/robert/SC_TIME " + TIMESTAMP + ".png"

• Press Alt+Tab

  static cairo_surface_t *
  get_window_surface (MetaWindow *window)

  
  diff --git a/src/core/screen.c b/src/core/screen.c
  index 1218232..a821be5 100644
  diff --git a/src/core/screen.c b/src/core/screen.c
  index 1218232..6a300ff 100644
  --- a/src/core/screen.c
  +++ b/src/core/screen.c
  @@ -1265,6 +1265,10 @@ get_window_surface (MetaWindow *window)
  cairo_destroy (cr);
  cairo_surface_destroy (surface);

• char *file = g_strdup_printf ("/home/robert/SC_TIME%lu.png", g_get_monotonic_time ());

• cairo_surface_write_to_png (scaled, file);

• g_free (file);

• return scaled; }

benchmark: master vs PR (there is no significant difference)

$ marco-theme-viewer BlueMenta

### master

![Screenshot at 2020-08-09 21-53-07](https://user-images.githubusercontent.com/10171411/89740719-8e3ea180-da8b-11ea-9e9d-45fbeec80c6c.png)

### PR

![Screenshot at 2020-08-09 21-54-37](https://user-images.githubusercontent.com/10171411/89740723-9696dc80-da8b-11ea-9ff9-0499c86a80ce.png)

[raveit65]

@rbuj We have a conflict...

[rbuj]

@raveit65 done. I fixed the merge conflict.

[rbuj]

@muktupavels We talked about this some time ago in another PR #476, I thought you participated in the conversation, but I was wrong. All I'm telling you is from memory:

At that moment of writing the other PR, I was focused on the performance after seeing an unusual comment on the code.

Doing it that way introduced another warning, From memory, do not convert directly the returned value of a function i.e. do the conversion using a variable, because of mismatch on datatypes, since round returned double and it was stored on a non double variable.

In other parts of the code I think I remember that they use the same principle, rounding up or down depending on whether the value is closer to one of the ends, i.e. using the round operation. I think it's done using macros to avoid the round function of math.h. What’s more, I think there is also a scaling macro, which cannot be reused by the data type, i.e. wide or sign, which is not much different from what this PR proposes.

I will see where what I told you comes from... perhaps from other projects.

./atril/backend/dvi/mdvi-lib/fontmap.c:#define DROUND(x)    ((x) >= 0 ? floor((x) + 0.5) : ceil((x) - 0.5))
./atril/backend/dvi/mdvi-lib/common.h:#define SFROUND(x)    (int)((x) >= 0 ? floor((x) + 0.5) : ceil((x) + 0.5))

Edit: I saw it on mate-desktop/libmate-desktop/mate-colorsel.c :)

#define SCALE(i) (i / 65535.)
#define UNSCALE(d) ((guint16)(d * 65535 + 0.5))

[muktupavels]

You asked for review and you got one...

[rbuj]

I'm very grateful to you, I will work to include the improvements you have asked me.

[rbuj]

@muktupavels I added the text below on the commit body:

- round on unscale (scale_factor * max_value + 0.5)
- do not require to invert float: cairo_scale (cr, 1/ratio, 1/ratio)
- use specific operations based on datatype: sqrtf for float
- no mixing single precision with double precision numbers

[muktupavels]

@rbuj Now imagine that it was not your work but was done by somebody else. You have found that this commit has introduced regression. Does that message gives anything useful to understand why change was made? Did it fix something?

Adding 0.5 can easily introduce regression because original author might have relied on truncating. Obviously you will not notice any regression in solid_picture()... Getting 32768 instead of 32767 for 0.5 changes nothing. If you would have looked how that function is used than you would notice that it is used only in 3 places, one is only if debugging redraws. You could use unsigned short for r, g, b parameters...

Also I still think this should be done in more smaller commits.

[rbuj]

Ok, I close this MR and I'll do small MRs, the 1st #638