Directly support ^% in pgf.

Support for `^` is inspired from the support for `_` as implemented
in underscore.sty, but much simpler because 1) we don't care about
hyphenation, 2) we don't care about substituting a real underscore by
a rule box when the current font lacks an underscore (there's no real
way out if the font has no caret, anyways), and 3) we don't care about
the active character ending up in tex `\commands` because the character
is only made active in the context of matplotlib-provided strings, not
globally.

Author: anntzer

lib/matplotlib/backends/backend_pgf.py

@@ -85,9 +85,6 @@ def _get_preamble():
 
 _NO_ESCAPE = r"(?<!\\)(?:\\\\)*"
 _split_math = re.compile(_NO_ESCAPE + r"\$").split
-_replace_escapetext = functools.partial(
-    # When the next character is an unescaped % or ^, insert a backslash.
-    re.compile(_NO_ESCAPE + "(?=[%^])").sub, "\\\\")
 _replace_mathdefault = functools.partial(
     # Replace \mathdefault (when not preceded by an escape) by empty string.
     re.compile(_NO_ESCAPE + r"(\\mathdefault)").sub, "")
@@ -106,8 +103,6 @@ def _tex_escape(text):
     This distinguishes text-mode and math-mode by replacing the math separator
     ``$`` with ``\(\displaystyle %s\)``. Escaped math separators (``\$``)
     are ignored.
-
-    The following characters are escaped in text segments: ``^%``
     """
     # Sometimes, matplotlib adds the unknown command \mathdefault.
     # Not using \mathnormal instead since this looks odd for the latex cm font.
@@ -116,11 +111,7 @@ def _tex_escape(text):
     # split text into normaltext and inline math parts
     parts = _split_math(text)
     for i, s in enumerate(parts):
-        if not i % 2:
-            # textmode replacements
-            s = _replace_escapetext(s)
-        else:
-            # mathmode replacements
+        if i % 2:  # mathmode replacements
             s = r"\(\displaystyle %s\)" % s
         parts[i] = s
     return "".join(parts)
@@ -168,7 +159,17 @@ def _escape_and_apply_props(s, prop):
         commands.append(r"\bfseries")
 
     commands.append(r"\selectfont")
-    return "".join(commands) + " " + _tex_escape(s)
+    return (
+        "{"
+        + "".join(commands)
+        + r"\catcode`\^=\active\def^{\ifmmode\sp\else\^{}\fi}"
+        # It should normally be enough to set the catcode of % to 12 ("normal
+        # character"); this works on TeXLive 2021 but not on 2018, so we just
+        # make it active too.
+        + r"\catcode`\%=\active\def%{\%}"
+        + _tex_escape(s)
+        + "}"
+    )
 
 
 def _metadata_to_str(key, value):
@@ -357,7 +358,11 @@ def _get_box_metrics(self, tex):
         """
         # This method gets wrapped in __init__ for per-instance caching.
         self._stdin_writeln(  # Send textbox to TeX & request metrics typeout.
-            r"\sbox0{%s}\typeout{\the\wd0,\the\ht0,\the\dp0}" % tex)
+            # \sbox doesn't handle catcode assignments inside its argument,
+            # so repeat the assignment of the catcode of "^" and "%" outside.
+            r"{\catcode`\^=\active\catcode`\%%=\active\sbox0{%s}"
+            r"\typeout{\the\wd0,\the\ht0,\the\dp0}}"
+            % tex)
         try:
             answer = self._expect_prompt()
         except LatexError as err:

lib/matplotlib/tests/test_backend_pgf.py

@@ -35,19 +35,22 @@ def compare_figure(fname, savefig_kwargs={}, tol=0):
 
 @pytest.mark.parametrize('plain_text, escaped_text', [
     (r'quad_sum: $\sum x_i^2$', r'quad_sum: \(\displaystyle \sum x_i^2\)'),
-    ('% not a comment', r'\% not a comment'),
-    ('^not', r'\^not'),
 ])
 def test_tex_escape(plain_text, escaped_text):
     assert _tex_escape(plain_text) == escaped_text
 
 
 @needs_pgf_xelatex
+@needs_ghostscript
 @pytest.mark.backend('pgf')
 def test_tex_special_chars(tmp_path):
     fig = plt.figure()
-    fig.text(.5, .5, "_^ $a_b^c$")
-    fig.savefig(tmp_path / "test.pdf")  # Should not error.
+    fig.text(.5, .5, "%_^ $a_b^c$")
+    buf = BytesIO()
+    fig.savefig(buf, format="png", backend="pgf")
+    buf.seek(0)
+    t = plt.imread(buf)
+    assert not (t == 1).all()  # The leading "%" didn't eat up everything.
 
 
 def create_figure():