Summary:
Currently CMake doesn't build builtins for AArch64 and if one does this anyway
it's likely that at least `__multc3`, `__floatditf` and `__floatunditf` will be
missing. There is actually more builtins to add, but these come from
different libc implementations, thus providing them makes compiler-rt for
AArch64 good enough at least for basic usage.
Builtins implementation were originally taken from FreeBSD project:
* [[ https://reviews.freebsd.org/D2173 | __multc3 ]]
* [[ https://reviews.freebsd.org/D2174 | __floatditf and __floatunditf ]]
Until they have been tested to find mistakes in `__float*` functions.
`__floatditf` was based on `__floatsitf`, which had the same mistakes
(fixed it in r243746).
Version of the builtins in this patch are fixed and complemented with basic
tests. Additionally they were tested via GCC's torture (this is what revealed
these issues).
P.S. Ed (author of FreeBSD patches) asked for feedback on the list some time ago (here [[ http://lists.cs.uiuc.edu/pipermail/llvmdev/2015-March/084064.html | here ]])
and got no response, but it seems to be worth adding these builtins as is and
extracting common part later.
Reviewers: howard.hinnant, t.p.northover, jmolloy, enefaim, rengolin, zatrazz
Subscribers: asl, emaste, samsonov, aemerson, llvm-commits, rengolin
Differential Revision: http://reviews.llvm.org/D11679
llvm-svn: 245296
Negative numbers were handled properly initially, but got broken
during addressing review, so none of them did actually work. Issues:
* Wrong negation.
* Wrong exponent calculation.
llvm-svn: 243746
Summary:
This patch fixes incorrect truncation when the input wider value is
exactly 2^dstBits. For that value, the overflow to infinity is not
correctly handled. The fix is to replace a strict '>' with '>='.
Currently,
__truncdfsf2(340282366900000000000000000000000000000.0) returns infinity
__truncdfsf2(340282366920938463463374607431768211456.0) returns 0
__truncdfsf2(400000000000000000000000000000000000000.0) returns infinity
Likewise, __truncdfhf2 and __truncsfhf2 (and consequently gnu_f2h_ieee)
are discontinuous at 65536.0.
This patch adds tests for all three cases, along with adding a missing
header include to fp_test.h.
Reviewers: joerg, ab, srhines
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D10594
llvm-svn: 240450
Like we do for the various __*tf* tests, check that long double is the
128bit type we expect directly in the header. The latter is now used by
unrelated tests (__*hf* since r237161), and those tests will break for
no reason if uint128_t doesn't exist, and long double isn't fp128.
llvm-svn: 239630
Mostly uninteresting, except:
- in __extendXfYf2, when checking if the number is normal, the old
code relied on the unsignedness of src_rep_t, which is a problem
when sizeof(src_rep_t) < sizeof(int): the result gets promoted to
int, the signedness of which breaks the comparison.
I added an explicit cast; it shouldn't affect other types.
- we can't pass __fp16, so src_t and src_rep_t are the same.
- the gnu_*_ieee symbols are simply duplicated definitions, as aliases
are problematic on mach-o (where only weak aliases are supported;
that's not what we want).
Differential Revision: http://reviews.llvm.org/D9693
llvm-svn: 237161
Summary:
This makes their local declarations match their definitions for ARM targets,
where they have a different calling convention.
This really only affects functions that use floating point types (since the
runtime functions use soft-float, and some targets may default to hard-float)
but it seemed good to make it uniform and do the int-only ones too.
Differential Revision: http://reviews.llvm.org/D9062
llvm-svn: 235722
80bit Intel/PPC long double is excluded due to lacking support
for the abstraction. Consistently provide saturation logic.
Extend to long double on 128bit IEEE extended platforms.
Initial patch with test cases from GuanHong Liu.
Reviewed by Steve Canon.
Differential Revision: http://reviews.llvm.org/D2804
llvm-svn: 231965
The clear_cache and enable_execute_stack tests attempt to memcpy the definition
of a function into a buffer before executing the function. The problem with
this approach is that on some targets (ARM with thumb mode compilation, MIPS
with MIPS16 codegen or uMIPS), you would use a pointer which is incorrect (it
would be off-by-one) due to the ISA selection being encoded into the address.
This ensures that the function address is retrieved correctly in all cases.
llvm-svn: 225215
Josh Gao