There were a few minor bugs in the rounding behavior and Inf/NaN
handling for the f80 __addxf3 and __subtf3 functions.
This change updates the original generic implementation to correctly
handle f80 floats, including the explicit integer bit.
Some SPARC CPUs (particularly old and/or embedded ones) only has atomic
TAS instruction available (`ldstub`). This adds support for emitting
that instruction in the spinlock.
* goals
- zig as linker for object files generated by other compilers
- zig-specific runtime features for eventual standardisation
* changes
- missing routines are marked with `missing`
- structure inspired by libgcc docs, but improved order and wording
- rename misspelled functions
- reorder and rephrase compiler_rt.zig to reflect documentation
- potential decimal float or fixed-point arithmetic support:
* 'Decimal float library routines' ca. 120 functions
* 'Fixed-point fractional library routines' ca. 300 functions
thanks to @Vexu for multiple reviews and @scheibo for review
Get rid of `std.math.F80Repr`. Instead of trying to match the memory
layout of f80, we treat it as a value, same as the other floating point
types. The functions `make_f80` and `break_f80` are introduced to
compose an f80 value out of its parts, and the inverse operation.
stage2 LLVM backend: fix pointer to zero length array tripping LLVM
assertion. It now checks for when the element type is a zero-bit type
and lowers such thing the same way that pointers to other zero-bit types
are lowered.
Both stage1 and stage2 LLVM backends are adjusted so that f80 is lowered
as x86_fp80 on x86_64 and i386 architectures, and identical to a u80 on
others. LLVM constants are lowered in a less hacky way now that #10860
is fixed, by using the expression `(exp << 64) | fraction` using llvm
constants.
Sema is improved to handle c_longdouble by recursively handling it
correctly for whatever the float bit width is. In both stage1 and
stage2.
- approach by Hacker's Delight with wrapping subtraction
- performance expected to be similar to addo
- tests with all relevant combinations of min,max with -1,0,+1 and all
combinations of sequences +-1,2,4..,max
- approach by Hacker's Delight with wrapping addition
- ca. 1.10x perf over the standard approach on my laptop
- tests with all combinations of min,max with -1,0,+1 and combinations of
sequences +-1,2,4..,max
We're going to remove the first parameter from this function in the
future. Stage2 already ignores the first parameter. So we put an `@as`
in here to make it work for both.
- use usize to decide if register size is big enough to store
multiplication result or if division is necessary
- multiplication routine with check of integer bounds
- wrapping multipliation and division routine from Hacker's Delight
Before this commit, compiling an empty main with Stage 2 on macOS x86_64 results in
```
../stage2/bin/zig build-exe -ODebug -fLLVM empty_main.zig
error: sub-compilation of compiler_rt failed
[...]/zig/stage2/lib/zig/std/special/compiler_rt/os_version_check.zig:26:10: error: TODO: Sema.zirStructInit for runtime-known struct values
```
By assigning the value to a variable we can sidestep the issue for now.
This allows stage2 to build more of compiler-rt.
I also changed `-%` to `-` for comptime ints in the div and mul
implementations of compiler-rt. This is clearer code and also happens to
work around a bug in stage2.
This improves readability as well as compatibility with stage2. Most of
compiler-rt is now enabled for stage2 with just a few functions disabled
(until stage2 passes more behavior tests).
- neg can only overflow, if a == MIN
- case `-0` is properly handled by hardware, so overflow check by comparing
`a == MIN` is sufficient
- tests: MIN, MIN+1, MIN+4, -42, -7, -1, 0, 1, 7..
See #1290
- abs can only overflow, if a == MIN
- comparing the sign change from wrapping addition is branchless
- tests: MIN, MIN+1,..MIN+4, -42, -7, -1, 0, 1, 7..
See #1290
- adds __cmpsi2, __cmpdi2, __cmpti2
- adds __ucmpsi2, __ucmpdi2, __ucmpti2
- use 2 if statements with 2 temporaries and a constant
- tests: MIN, MIN+1, MIN/2, -1, 0, 1, MAX/2, MAX-1, MAX if applicable
See #1290
- use negXi2.zig to prevent confusion with negXf2.zig
- used for size optimized builds and machines without carry instruction
- tests: special cases 0, -INT_MIN
* use divTrunc range and shift with constant offsets
See #1290
- each byte gets masked, shifted and combined
- use boring masks instead of comptime for readability
- tests: bit patterns with reverse operation, if applicable
See #1290
- use Bit Twiddling Hacks: Compute parity in parallel
- test cases derived from popcount.zig
- tests: compare naive approach 10_000 times with random numbers created
from naive seed 42
- compiler_rt.zig: sort by LLVM builtin order and add comments to improve structure
See #1290
- apply simpler approach than LLVM for __popcountdi2
taken from The Art of Computer Programming and generalized
- rename popcountdi2.zig to popcount.zig
- test cases derived from popcountdi2_test.zig
- tests: compare naive approach 10_000 times with
random numbers created from naive seed 42
See #1290
The BPF target does not support mutable global variables. Mark the BPF
target as a target that does not support atomic variables in order to
avoid including the global spinlock table provided in compiler_rt.
LLVM and compiler-rt must agree on how the parameters are passed, it
turns out that in LLVM13 something changed and broke the test case for
AArch64 systems.
It has nothing to do with fma at all.
Closes#9900
* work around a stage1 miscompilation leading to the wrong integer
comparison predicate being emitted.
* fix the bug of not annotating callsites with the calling convention
of the callee, leading to undefined behavior.
* add the `nobuiltin` attribute when building freestanding libc or
compiler_rt libraries to prevent e.g. memcpy from being "optimized"
into a call to itself.
* compiler-rt: change a call to be comptime to make the generated LLVM
IR simpler and easier to study.
I still can't enable the widening tests due to the compiler-rt compare
function being miscompiled in some not-yet-diagnosed way.
