- cbe: Implement linksection support, to support TLS when not linking libc
- cbe: Support under-aligned variables / struct fields
- cbe: Support packed structs (in the C definition of packed)
- windows: Fix regression with x86 _tls_array
- compiler_rt: Add 128-bit atomics to compiler_rt
- tests: Re-enable threadlocal tests on cbe+windows, and llvm+x86
- tests: Re-enable f80 tests that now pass
- ci: change windows ci to run the CBE behaviour tests with -lc, to match how the compiler is bootstrapped
- update zig1.wasm
A followup to 5bd69c655d9e04102c8a64ced1215c9d69f4f03f .
For compatibility purposes clang defines __GNUC__. This broke logic
where clang14 __has_builtin returns false but `|| defined(__GNUC__)`
made it true.
CI did not catch this because it uses clang15. However, bootstrapping
with latest apple compiler is clang14. We also want to support older
clang.
- Add cpuid / getXCR0 functions for the cbe to use instead of asm blocks
- Don't cast between 128 bit types during truncation
- Fixup truncation to use functions for shifts / adds
- Fixup float casts for undefined values
- Add test for 128 bit integer truncation
- Fix zig_clz_u128 not respecting the bits argument. This was crashing the compile-rt addxf3 tests with the cbe
- Instead of redering a negation for negative 128 bit int literals, render the literal as twos complement. This allows
rendering int representations of floats correctly (specifically f80).
- Add .StaticInitializer to ValueRenderLocation to indicate that the emitted values
must be constant expressions (no function calls, struct casting).
- Add new path for special float types (nan, inf) that works in constant expressions
- Implement windows.teb() using a syscall for .stage2_c because x64 MSVC
doesn't support any kind of inline asm
- Map the __builtin_nan(f|l)? functions to nan(f|l)? on msvc
- MSVC throws C2099 when initializing a struct with cast syntax
in a global initializer. Added zig_as_init_ to handle this case,
and generate it only in static initializers for > 64 bit ints.
- Change float initialization to emit the integer representation
in global initializers to avoid C2099 caused by calling nan.
MSVC can't explicitly cast a struct to a typedef of itself (ie. f128 to i128). Added a
set of macros to handle float casting, and to not produce a cast for this specific
case on MSVC. A better approach would probably be to know if the cast is redundant
and not do it.
- Forward declare int builtins, so the definitions aren't assumed incorrectly
- Add define to handle MSVC not support static const in function parameter array lengths
- Fixup several spots where int128 support was assumed.
- Support zig_align
- Support zig_export
- Stub out some missing non-builtin functions
- Added StringLiteral to automatically split string literals when they get to 16380 in size,
which is the maxmimum pre-concatenation string literal size on MSVC.
In general the C backend should lower to human-maintainable C code
whenever possible. Directly using C types that one would use when
writing C code is one part of the strategy.
The concern with including stdint.h is C89 compatibility. Well, we can
just check the C std lib version before deciding to include that header.
