GitLab

default STOLEN_X86_REGS to 4 (workaround because GHC doesn't normally
define STOLEN_X86_REGS when compiling plan .c source)

Fix up STM when compiling unregisterised.  There were a few wibbles
with the stack layout.

Instead of defining NO_REGS when IN_STG_CODE==0, define
NO_GLOBAL_REG_DECLS instead.

This means that in non-STG code we can still get at the values of
REG_R1 & co., even though the global register decls are turned off.
This is necessary because we sometimes need to set up different stack
layouts depending on REG_R1.

Make it so that global register declarations are turned off if
NO_GLOBAL_REG_DECLS is defined.

Improve a comment

slight simplification

Fix the unreg test in the rule for building AutoApply$(_way).cmm

Add way debug_u

Fix bug(s) in the register allocator: if a virtual register is both in
memory and in a register (perhaps because it was recently loaded from
a spill slot), and the current instruction writes it, we're supposed
to invalidate the memory slot.  That wasn't happening properly.

This fixes two problems noticed when using -prof -fasm: 10queens in the
testsuite gives the wrong answer, and nofib/spectral/hartel/ida fails
with a 'head []' message.  Interesting bug to track down!

-package Cabal is required whenever $(bootstrapped) is yes, not only
for bootstrapped compiles with GHCi.

Fix a bug in mk_switch.

Track removal of Distribution.Compat.Error

Track removal of Distribution.Compat.Error.

Restore splitting at ':' for the -i option, which I broke recently.

Yet another STM fix: don't attempt to kick any threads which are ThreadRelocated.

fix System.IO.Error breakage

Numerous bug fixes to the STM code, mostly from a debugging session
with Tim Harris.  The test that flushed out all the bugs will shortly
be added to the test suite.

Issue an error when the stdcall calling convention is requested on
non-x86 platforms.

import System.IO.Error

Move _srtd and _closure_tbl to read-only data instead of text
(these things contain pointers to other symbols, so they can't go into
.text for position-independent code).

PowerPC Linux: Fix a slight ABI bug in genCCall
Double and I64 arguments on the stack were not properly aligned.

Cleanup and additional comments for the Mac OS/AIX/powerpc64-linux adjustor
code.

ignore OPTIONS_anything_else when looking for OPTIONS_GHC/OPTIONS pragmas.

Rename OPTIONS to OPTIONS_GHC.

Make GHC accept OPTIONS_GHC.  OPTIONS is also accepted, for now.

fix for parsing OPTIONS pragmas: OPTIONS should be followed by a
non-identifier character.

wibble: got my endifs mixed up

When way=u, turn off split objects by setting SplitObjs=NO, rather
than testing for way=u directly.  Fixes a build wibble.

Fix $(StripLibraries) code after ld changes yesterday.

Support foreign import "wrapper" with > 6 arguments on PowerPC Linux (32-bit).
This calling convention looks deceptively simple at first, but it turns out
to be hideously complex...

createDirectoryIfMissing is added to Compat.Directory and is used in ghc-pkg.
The mingw32_HOST_OS is replaced with mingw32_TARGET_OS. I don't know why but
prior the last commit the tool was working with mingw32_HOST_OS fine but not
it isn't. Maybe I miss something. Simon, could you check whether the patch is
fine?

Fix for AIX-like ABIs.

Refactored and cleaned up ld-related tests. Only tested mildly under Linux.

Handle foreign import wrapper properly for MacOS X, powerpc64-linux and AIX.
Only Mac OS X tested so far.
Pass information about argument types from DsForeign to createAdjustor encoded
as a string ('i' for integers, 'f' for floats, 'd' for doubles and 'l' for
long [64bit] integers).

Add *.S to $(C_SRCS).  Seems to do the right thing, including  mkdependC.

Obscure bugfix affecting foreign import "wrapper" with non-word-sized
argument types (eg. Double) on Windows.  The list of arguments types
returned by dsFExport was the boxed types, rather than the unboxed
types, so dsFExportDynamic couldn't get the correct sizes of the
argument types to the stub function.

It's more correct now, but not totally correct (see the comment for
details).

Noticed by: Wolfgang Thaller.

Remove copied definitions for joinFileName, splitFileName:
System.FilePath is in libghccompat now.

Work around problems caused by limit on the length of the command line
in ld, which shows up when building OpenGL on Windows with SplitObjs=YES.

We now pass the names of the input files to ld via a linker script.
This is (probably) only supported by GNU ld, so we now have to detect
GNU ld in the configure script (yawn) and back off to the old method
if we don't have it.

Modified the mangler for powerpc64.  This is the last missing piece for supporting
registerised builds on PPC64.

The current state of the mangler is temporary.  We're discussing a better way to take
advantage of the PPC64 ABI's unique features to provide a better way of storing info
tables.

I don't think I've changed anything that affects other platforms.

Fixed this 6.4 TODO item listed on the wiki:

 PowerPC Linux (32bit): Fix GHCi FFI calls for arguments that are not passed on the stack (ByteCodeFFI).

Separated the code for Darwin and Linux (for PowerPC only).  Rewrote the Linux version to account for the differences in the ABIs.

All changes are inside #if powerpc_TARGET_ARCH && linux_TARGET_OS except:
  - import Data.List ( mapAccumL )      (used by my code)
  - small fix to a comment typo in Wolfgang's Darwin code
  - changed 'undefined' to a more meaningful 'error' message if
     mkMarshalCode is unimplemented

Ran regression tests.  It passes them all except for the ones that are broken because of the 'wrapper' problems currently being addressed.