Merge pull request numba#8852 from stuartarchibald/wip/slp_disabled_b…

…y_default

Disable SLP vectorisation due to miscompilations.

docs/source/reference/envvars.rst

@@ -312,8 +312,10 @@ Compilation options
 .. envvar:: NUMBA_SLP_VECTORIZE
 
    If set to non-zero, enable LLVM superword-level parallelism vectorization.
+   Note that use of this feature has occasionally resulted in LLVM producing
+   miscompilations, hence it is off by default.
 
-   *Default value:* 1
+   *Default value:* 0
 
 .. envvar:: NUMBA_ENABLE_AVX
 

numba/core/config.py

@@ -305,8 +305,9 @@ def optional_str(x):
         LOOP_VECTORIZE = _readenv("NUMBA_LOOP_VECTORIZE", int,
                                   not (IS_WIN32 and IS_32BITS))
 
-        # Switch on  superword-level parallelism vectorization, default is on.
-        SLP_VECTORIZE = _readenv("NUMBA_SLP_VECTORIZE", int, 1)
+        # Enable superword-level parallelism vectorization, default is off
+        # since #8705 (miscompilation).
+        SLP_VECTORIZE = _readenv("NUMBA_SLP_VECTORIZE", int, 0)
 
         # Force dump of generated assembly
         DUMP_ASSEMBLY = _readenv("NUMBA_DUMP_ASSEMBLY", int, DEBUG)

numba/tests/test_vectorization.py

@@ -1,8 +1,8 @@
 import platform
 import numpy as np
 from numba import types
-from unittest import TestCase, skipIf
-from numba.tests.support import override_env_config
+import unittest
+from numba.tests.support import override_env_config, TestCase
 from numba.core.compiler import compile_isolated, Flags
 from numba.core.cpu_options import FastMathOptions
 
@@ -13,7 +13,7 @@
     llvm.set_option("", "--debug-only=loop-vectorize")
 
 
-@skipIf(platform.machine() != 'x86_64', 'x86_64 only test')
+@unittest.skipIf(platform.machine() != 'x86_64', 'x86_64 only test')
 class TestVectorization(TestCase):
     """
     Tests to assert that code which should vectorize does indeed vectorize
@@ -40,6 +40,9 @@ def do_sum(x):
         self.assertIn("vector.body", llvm_ir)
         self.assertIn("llvm.loop.isvectorized", llvm_ir)
 
+    # SLP is off by default due to miscompilations, see #8705. Put this into a
+    # subprocess to isolate any potential issues.
+    @TestCase.run_test_in_subprocess(envvars={'NUMBA_SLP_VECTORIZE': '1'})
     def test_slp(self):
         # Sample translated from:
         # https://www.llvm.org/docs/Vectorizers.html#the-slp-vectorizer
@@ -71,3 +74,7 @@ def sum_sqrt_list(lst):
                               fastmath=True)
         self.assertIn("vector.body", llvm_ir)
         self.assertIn("llvm.loop.isvectorized", llvm_ir)
+
+
+if __name__ == '__main__':
+    unittest.main()