GitLab

- x86 extended float mantissa size includes the explicit one now
- fixed broken qnan/snan for x86 extended float
- simplification/tuning

This fixes opt/combo6.c.

Before this change the .so file would not mention libm as a dependency,
so users had to explicitely link -lm when linking -lfirm.

Add some documentation about this fact.

The tarval module has a maximally supported precision. Assert that no
modes with higher precision are created, otherwise the tarval module
will silently drop the higher bits on calculation. TODO: Change tarval
module to dynamically use smaller/bigger representations depending on
the mode.

Removed mode_LLs, mode_LLu, mode_Q for now as they are bigger than the
limit.

This could lead to endless inlining.

This fixes gcc/pr22379.c. Yeah.

Fixes valgrind error

This fixes opt/loadstore2.c.

The file wasn't really meant to be included (it lacks the proper includes to be
useful on its own), the correct file to include is irnode.h.

Anyway the guards don't hurt either...

Happened in llvm/2003-05-23-TransparentUnion.c test case

Avoids signed left shift overflow, which is undefined behavior.

Silences clang -fsanitize=undefined.
Not sure why clang complains here.

If scheduled-before block is not a cf-predecessor, the function
ultimately computed a NaN and returned false. However, such blocks
should be aligned.

Initially, clang -fsanitize=undefined found the NaN, which is certainly
not expected by original programmer. The DELTA approach still looks
fishy to me.