I made sure to create a merge conflict with the code attached to ticket #229
(which comments this code out due to changes in the DMD 'Type' type hierarchy)
so that if it gets committed as-is there will be a human looking at it.
The switch allows the optimizer and inliner to run on all modules at once and opens the door for template instantiation improvements that should lower compile time and executable size.
- See through typedefs and enums in passByVal
- Don't depend on TypeFunction::parameters->dim being the actual number of
parameters; it contains unexpanded tuples as single elements.
- Implement x86-64 extern(C), hopefully correctly.
- Tried to be a bit smarter about extern(D) while I was there.
Interestingly, this code seems to be generating more efficient code than
gcc and llvm-gcc in some edge cases, like returning a `{ [7 x i8] }` loaded from
a stack slot from an extern(C) function. (gcc generates 7 1-byte loads, while
this code generates a 4-byte, a 2-byte and a 1-byte load)
I also added some changes to make sure structs being returned from functions or
passed in as parameters are stored in memory where the rest of the backend seems
to expect them to be. These should be removed when support for first-class
aggregates improves.
According to the spec and C ABIs, the contents of padding is undefined.
This was breaking dstress creal_01,02,03,07,08,09 and 20 on x86-64.
The code was apparently added to fix creal_13,14 and 15, which explicitly
compare padding and assert if it's different. I'd argue those tests are broken.
(Also, on x86-64 they *also* fail)
The tests this fixes, on the other hand, are basic arithmetic.
functions.
There's no need to waste cycles with extern(D), which we get to define
ourselves. Fixes tests/mini/asm8.d. (Since the asm abiret code already assumed
{xmm0, xmm1} returns)
llvm::OStream provides all std::ostream functionality (by holding a
std::ostream* internally), but
* doesn't include <iostream>, avoiding per-file overhead.
* allows the stream pointer to be null, and the (inlined) operators do nothing
when that's the case. (This also allows removal of the ofstream("/dev/null")
hack Logger used when disabled, which presumably wasn't very portable)
- Moved main() into its own file gen/main.cpp
- Fixed basic cross compilation
- removed the option for setting OS
- added support for llc's mattr, mcpu and mtriple switches
- added basic ABI abstraction for return value rewrites, it's not perfect and will probably be completely rewritten once I get to handling parameter rewrites as well.
- x86-64 extern(C) abi for cfloat returns now match (llvm-)gcc.
Note: For a backward compatible interface, use the new bin/ldmd script. It
supports all old options while passing on anything it doesn't recognize.
Some changes caused by this:
* -debug and -version are now -d-debug and -d-version due to a conflict with
standard LLVM options.
* All "flag" options now allow an optional =true/=1/=false/=0 suffix.
* Some "hidden debug switches" starting with "--" were renamed because LLVM
doesn't care about the number of dashes, so they were conflicting with other
options (such as -c).
The new versions start with "-hidden-debug-" instead of "--"
* --help works, but has a non-zero exit code. This breaks some Tango scripts
which use it to test for compiler existence. See tango.patch.
Some changes not (directly) caused by this;
* (-enable/-disable)-FOO options are now available for pre- and postconditions.
* -march is used instead of -m (like other LLVM programs), but -m is an alias
for it.
* -defaultlib, -debuglib, -d-debug and -d-version allow comma-separated values.
The effect should be identical to specifying the same option multiple times.
I decided against allowing these for some other options because paths might
contain commas on some systems.
* -fPIC is removed in favor of the standard LLVM option -relocation-model=pic
Bug:
* If -run is specified as the last argument in DFLAGS, no error is generated.
(Not very serious IMHO)
The assembler was miscompiling "add" (specifically, the "add reg/mem, imm"
variations).
The change that caused this seems to have been made because without it, some
"add"s didn't compile at all.
This patch reverts the previous change, and makes sure assembler operands are
remapped correctly even though the input operands auto-generated due to
updating operations aren't explicitly used.
I think that technically, using "*m0" instead of "*0" allows LLVM to pick
between using the same memory as output 0 and using a new memory location.
(So far I haven't been able to construct a testcase that actually breaks
because of this, though)
