Fix baseline alignment when using usetex.

anntzer · anntzer · commit ccabb61a3dae · 2020-04-14T10:11:30.000+02:00
Previously we inferred the baseline of usetex strings with the baseline
of the character with the lowest baseline, which is wrong in the
presence of indices (see usetex_baseline_test.py).  There was an option
to use the "preview" latex package to fix that, but that was never made
the default and likely cannot as the package is GPL.

Instead I can infer the baseline position by "reverse-engineering"
the starting instructions of the dvi stream (this was done without
consulting the approach in "preview").  The results can be checked
using usetex_baseline_test.py (preview=False now looks the same as
preview=True; likewise, the output of test_usetex was changed.

The text.latex.preview rc can now be deprecated (in a future PR).

The test was also updated to include the strings of the
usetex_baseline_test example (which may go away once text.latex.preview
is removed).
diff --git a/lib/matplotlib/backends/backend_ps.py b/lib/matplotlib/backends/backend_ps.py
@@ -484,8 +484,7 @@ def draw_tex(self, gc, x, y, s, prop, angle, ismath='TeX!', mtext=None):
                 r'\psfrag{%s}[Bl][Bl][1][%f]{\fontsize{%f}{%f}%s}' % (
                     thetext, angle, fontsize, fontsize*1.25, tex))
         else:
-            # Stick to the bottom alignment, but this may give incorrect
-            # baseline some times.
+            # Stick to the bottom alignment.
             pos = _nums_to_str(x-corr, y-bl)
             self.psfrag.append(
                 r'\psfrag{%s}[bl][bl][1][%f]{\fontsize{%f}{%f}%s}' % (
diff --git a/lib/matplotlib/dviread.py b/lib/matplotlib/dviread.py
@@ -274,6 +274,9 @@ def _output(self):
             maxx = max(maxx, x + w)
             maxy = max(maxy, y + e)
             maxy_pure = max(maxy_pure, y)
+        if self._baseline_v is not None:
+            maxy_pure = self._baseline_v  # This should normally be the case.
+            self._baseline_v = None
 
         if not self.text and not self.boxes:  # Avoid infs/nans from inf+/-inf.
             return Page(text=[], boxes=[], width=0, height=0, descent=0)
@@ -304,9 +307,24 @@ def _read(self):
         Read one page from the file. Return True if successful,
         False if there were no more pages.
         """
+        # Pages appear to start with the sequence
+        #   bop (begin of page)
+        #   xxx comment
+        #   down
+        #   push
+        #     down, down
+        #     push
+        #       down (possibly multiple)
+        #       push  <=  here, v is the baseline position.
+        #         etc.
+        # (dviasm is useful to explore this structure.)
+        self._baseline_v = None
         while True:
             byte = self.file.read(1)[0]
             self._dtable[byte](self, byte)
+            if (self._baseline_v is None
+                    and len(getattr(self, "stack", [])) == 3):
+                self._baseline_v = self.v
             if byte == 140:                         # end of page
                 return True
             if self.state is _dvistate.post_post:   # end of file
diff --git a/lib/matplotlib/tests/baseline_images/test_usetex/test_usetex.pdf b/lib/matplotlib/tests/baseline_images/test_usetex/test_usetex.pdf
diff --git a/lib/matplotlib/tests/baseline_images/test_usetex/test_usetex.png b/lib/matplotlib/tests/baseline_images/test_usetex/test_usetex.png
diff --git a/lib/matplotlib/tests/test_usetex.py b/lib/matplotlib/tests/test_usetex.py
@@ -12,24 +12,32 @@
     pytestmark = pytest.mark.skip('Missing TeX of Ghostscript or dvipng')
 
 
-@image_comparison(baseline_images=['test_usetex'],
-                  extensions=['pdf', 'png'],
-                  tol={'aarch64': 2.868, 'x86_64': 2.868}.get(
-                      platform.machine(), 0.3
-                      ))
+@image_comparison(
+    baseline_images=['test_usetex'],
+    extensions=['pdf', 'png'],
+    style="mpl20")
 def test_usetex():
     mpl.rcParams['text.usetex'] = True
     fig = plt.figure()
     ax = fig.add_subplot(111)
-    ax.text(0.1, 0.2,
+    kwargs = {"verticalalignment": "baseline", "size": 24,
+              "bbox": dict(pad=0, edgecolor="k", facecolor="none")}
+    ax.text(0.2, 0.7,
             # the \LaTeX macro exercises character sizing and placement,
             # \left[ ... \right\} draw some variable-height characters,
             # \sqrt and \frac draw horizontal rules, \mathrm changes the font
             r'\LaTeX\ $\left[\int\limits_e^{2e}'
             r'\sqrt\frac{\log^3 x}{x}\,\mathrm{d}x \right\}$',
-            fontsize=24)
-    ax.set_xticks([])
-    ax.set_yticks([])
+            **kwargs)
+    ax.text(0.2, 0.3, "lg", **kwargs)
+    ax.text(0.4, 0.3, r"$\frac{1}{2}\pi$", **kwargs)
+    ax.text(0.6, 0.3, "$p^{3^A}$", **kwargs)
+    ax.text(0.8, 0.3, "$p_{3_2}$", **kwargs)
+    for x in {t.get_position()[0] for t in ax.texts}:
+        ax.axvline(x)
+    for y in {t.get_position()[1] for t in ax.texts}:
+        ax.axhline(y)
+    ax.set_axis_off()
 
 
 @check_figures_equal()
diff --git a/lib/matplotlib/texmanager.py b/lib/matplotlib/texmanager.py
@@ -411,7 +411,7 @@ def get_text_width_height_descent(self, tex, fontsize, renderer=None):
             return width, height + depth, depth
 
         else:
-            # use dviread. It sometimes returns a wrong descent.
+            # use dviread.
             dvifile = self.make_dvi(tex, fontsize)
             with dviread.Dvi(dvifile, 72 * dpi_fraction) as dvi:
                 page, = dvi
