In LogTransform, clip after log, not before.

When clipping before taking the log, the clip value in the output array cannot be beyond log(epsilon) ~ -300. When clipping after taking the log, the clip value can be much further (see comments regarding choice of value). This allows setting the default nonpos mode back to "clip". Some test images had to be updated as this effectively moves the position of the "clip point", slightly changing the slope of the line that goes to that point and thus the rasterization (by a tiny amount).
matplotlib · tacaswell · Oct 25, 2017 · Oct 19, 2017 · Oct 22, 2017 · Oct 22, 2017
commit 07a9450b407ee5dc89b82e388840d510b009ef3e
diff --git a/lib/matplotlib/axes/_axes.py b/lib/matplotlib/axes/_axes.py
@@ -1618,6 +1618,7 @@ def semilogx(self, *args, **kwargs):
             self.cla()
         d = {'basex': kwargs.pop('basex', 10),
              'subsx': kwargs.pop('subsx', None),
+             'nonposx': kwargs.pop('nonposx', 'clip'),
              }
 
         self.set_xscale('log', **d)
@@ -1673,6 +1674,7 @@ def semilogy(self, *args, **kwargs):
             self.cla()
         d = {'basey': kwargs.pop('basey', 10),
              'subsy': kwargs.pop('subsy', None),
+             'nonposy': kwargs.pop('nonposy', 'mask'),
              }
         self.set_yscale('log', **d)
         b = self._hold
@@ -2851,11 +2853,6 @@ def errorbar(self, x, y, yerr=None, xerr=None,
             Valid kwargs for the marker properties are
 
             %(Line2D)s
-
-        Notes
-        -----
-        Error bars with negative values will not be shown when plotted on a
-        logarithmic axis.
         """
         kwargs = cbook.normalize_kwargs(kwargs, _alias_map)
         # anything that comes in as 'None', drop so the default thing

diff --git a/lib/matplotlib/axes/_base.py b/lib/matplotlib/axes/_base.py
@@ -2970,10 +2970,8 @@ def set_xscale(self, value, **kwargs):
 
         matplotlib.scale.LogisticTransform : logit transform
         """
-        # If the scale is being set to log, mask nonposx to prevent headaches
-        # around zero
-        if value.lower() == 'log' and 'nonposx' not in kwargs:
-            kwargs['nonposx'] = 'mask'
+        if value.lower() == "log":  # Avoid headaches around zero.
+            kwargs.setdefault("nonposx", "clip")
 
         g = self.get_shared_x_axes()
         for ax in g.get_siblings(self):
@@ -3292,10 +3290,8 @@ def set_yscale(self, value, **kwargs):
 
         matplotlib.scale.LogisticTransform : logit transform
         """
-        # If the scale is being set to log, mask nonposy to prevent headaches
-        # around zero
-        if value.lower() == 'log' and 'nonposy' not in kwargs:
-            kwargs['nonposy'] = 'mask'
+        if value.lower() == "log":  # Avoid headaches around zero.
+            kwargs.setdefault("nonposy", "clip")
 
         g = self.get_shared_y_axes()
         for ax in g.get_siblings(self):

diff --git a/lib/matplotlib/scale.py b/lib/matplotlib/scale.py
@@ -92,15 +92,24 @@ class LogTransformBase(Transform):
 
     def __init__(self, nonpos):
         Transform.__init__(self)
-        if nonpos == 'mask':
-            self._fill_value = np.nan
-        else:
-            self._fill_value = 1e-300
+        self._clip = {"clip": True, "mask": False}[nonpos]
 
     def transform_non_affine(self, a):
-        with np.errstate(invalid="ignore"):
-            a = np.where(a <= 0, self._fill_value, a)
-        return np.divide(np.log(a, out=a), np.log(self.base), out=a)
+        with np.errstate(divide="ignore", invalid="ignore"):
+            out = np.log(a)
+        out /= np.log(self.base)
+        if self._clip:
+            # SVG spec says that conforming viewers must support values up
+            # to 3.4e38 (C float); however experiments suggest that Inkscape
+            # (which uses cairo for rendering) runs into cairo's 24-bit limit
+            # (which is apparently shared by Agg).
+            # Ghostscript (used for pdf rendering appears to overflow even
+            # earlier, with the max value around 2 ** 15 for the tests to pass.
+            # On the other hand, in practice, we want to clip beyond
+            #     np.log10(np.nextafter(0, 1)) ~ -323
+            # so 1000 seems safe.
+            out[a <= 0] = -1000
+        return out
 
 
 class InvertedLogTransformBase(Transform):
@@ -432,18 +441,17 @@ class LogitTransform(Transform):
 
     def __init__(self, nonpos):
         Transform.__init__(self)
-        if nonpos == 'mask':
-            self._fill_value = np.nan
-        else:
-            self._fill_value = 1e-300
         self._nonpos = nonpos
+        self._clip = {"clip": True, "mask": False}[nonpos]
 
     def transform_non_affine(self, a):
         """logit transform (base 10), masked or clipped"""
-        with np.errstate(invalid="ignore"):
-            a = np.select(
-                [a <= 0, a >= 1], [self._fill_value, 1 - self._fill_value], a)
-        return np.log10(a / (1 - a))
+        with np.errstate(divide="ignore", invalid="ignore"):
+            out = np.log10(a / (1 - a))
+        if self._clip:  # See LogTransform for choice of clip value.
+            out[a <= 0] = -1000
+            out[1 <= a] = 1000
+        return out
 
     def inverted(self):
         return LogisticTransform(self._nonpos)

diff --git a/lib/matplotlib/tests/baseline_images/test_axes/log_scales.png b/lib/matplotlib/tests/baseline_images/test_axes/log_scales.png
diff --git a/lib/matplotlib/tests/baseline_images/test_axes/log_scales.svg b/lib/matplotlib/tests/baseline_images/test_axes/log_scales.svg
diff --git a/lib/matplotlib/tests/baseline_images/test_scale/logscale_mask.png b/lib/matplotlib/tests/baseline_images/test_scale/logscale_mask.png