Subzero: Crosstest test_arith properly tests i8/i16.

crosstest/crosstest.cfg

 [simple_loop]
 driver: simple_loop_main.c
 test: simple_loop.c
 
 [mem_intrin]
 driver: mem_intrin_main.cpp
 test: mem_intrin.cpp
 flags: --sz=-fmem-intrin-opt
 
 [test_arith]
 driver: test_arith_main.cpp
-test: test_arith.cpp test_arith_frem.ll test_arith_sqrt.ll test_arith_fabs.ll
+# Note: test_arith.ll is auto-generated from test_arith.cpp - switch the "test:"
+# lines and then rename the resulting test_arith.*.pnacl.ll to test_arith.ll.
+# Then test_arith.ll is hand-modified so that i8 and i16 functions actually use
+# the i8/i16 form of the arithmetic instruction, to counteract C++'s integral

[John, 2016/05/25 17:26:28]

It's not C++'s rule as the spec say types **can** be promoted. This seems more
like a pnacl-specific issue.

+# promotion rules.
+#
+#test: test_arith.cpp test_arith_frem.ll test_arith_sqrt.ll test_arith_fabs.ll
+test: test_arith.ll test_arith_frem.ll test_arith_sqrt.ll test_arith_fabs.ll
 
 [test_bitmanip]
 driver: test_bitmanip_main.cpp
 test: test_bitmanip.cpp test_bitmanip_intrin.ll
 
 [test_calling_conv]
 driver: test_calling_conv_main.cpp
 test: test_calling_conv.cpp
 
 [test_cast]
@@ skipping 27 matching lines @@
 # bitcode patterns.
 flags: --clang-opt=0
 
 [test_sync_atomic]
 driver: test_sync_atomic_main.cpp
 test: test_sync_atomic.cpp
 
 [test_vector_ops]
 driver: test_vector_ops_main.cpp
 test: test_vector_ops.cpp test_vector_ops_ll.ll