test failed on SPARC

[tankf33der]

test failed on SPARC

[bakpakin]

The runtime error from UBSan can probably be ignored - we rely on (/ 0 0) == qNaN, and that should hold true on most architectures, and should for SPARC.

Janet uses the nanboxing technique internally, and relies on NaNs generated by floating point operations to be of a certain form that is different than the artificially created ones that are actually tagged pointers (quiet NaNs vs. signaling, but's it's really more about the specific bits in the signalling NaNs) . There are a bunch of fiddly macros in janet.h for dealing with this.

If I had to guess, the table has length 2, so src/core/table.c, line 135 is not properly guarding the
insertion of a NaN key in the table.

    if (janet_checktype(key, JANET_NUMBER) && isnan(janet_unwrap_number(key))) return;

Some things to try:

• Compile again with -DJANET_NO_NANBOX. This should hopefully work.
• Assert that 0.0 / 0.0 is really producing a NaN number type.

Sample diff:

diff --git a/Makefile b/Makefile
index e04662a..5b75117 100644
--- a/Makefile
+++ b/Makefile
@@ -35,6 +35,7 @@ MANPATH?=$(PREFIX)/share/man/man1/
 DEBUGGER=gdb
 
 CFLAGS=-std=c99 -Wall -Wextra -Isrc/include -fpic -O2 -fvisibility=hidden \
+          -DJANET_NO_NANBOX -fsanitize=undefined \
           -DJANET_BUILD=$(JANET_BUILD)
 LDFLAGS=-rdynamic
 
diff --git a/src/boot/system_test.c b/src/boot/system_test.c
index cf991f9..1b03a7d 100644
--- a/src/boot/system_test.c
+++ b/src/boot/system_test.c
@@ -48,5 +48,7 @@ int system_test() {
     assert(janet_equals(janet_cstringv("a string."), janet_cstringv("a string.")));
     assert(janet_equals(janet_csymbolv("sym"), janet_csymbolv("sym")));
 
+    assert(JANET_NUMBER == janet_type(janet_wrap_number(0.0 / 0.0)));
+
     return 0;
 }

[tankf33der]

-DJANET_NO_NANBOX helped.