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)
The first was that unless otherwise specified, inputs are assumed to be in
registers even if they specify a "matching output" that's in memory.
While testing that fix, I also ran into a bug causing the generated "matching
output" for any input was always the first one instead of the correct one.
CFGSimplification pass.
The pass was originally added as a workaround for a LLVM 2.4 bug where debug
info would reference labels that have been optimized away.
While emitting a template instantiation only once is good for compile times
and binary sizes, it doesn't work with linkonce linkage as inlined function
bodies could be discarded. Since we don't want to inhibit inlining, templates
are reverted to the previous behavior, where an instantiation is emitted for
each module using it.
In the future, a custom inlining pass may allow us to switch back to
common/weak linkage and reenable smart template instance emission.
Applied patch from ticket #129 to compile against latest LLVM. Thanks Frits van Bommel.
Fixed implicit return by asm block at the end of a function on x86-32. Other architectures will produce an error at the moment. Adding support for new targets is fairly simple.
Fixed return calling convention for complex numbers, ST and ST(1) were switched around.
Added some testcases.
I've run a dstress test and there are no regressions. However, the runtime does not seem to compile with symbolic debug information. -O3 -release -inline works well and is what I used for the dstress run. Tango does not compile, a small workaround is needed in tango.io.digest.Digest.Digest.hexDigest. See ticket #206 .
Fixed align N; in asm blocks.
Fixed inreg parameter passing on x86 for ref/out params.
Removed support for lazy initialization of function local static variables, I have no idea why I ever implemented this, it's not in the D spec, and DMD doesn't support it :P
Some of the global variable related changes might cause minor regressions, but they should be easily fixable.
