4 years agox86: Reinstate hand-written unwind info for sysv.S
Richard Henderson [Wed, 10 Dec 2014 21:37:36 +0000 (13:37 -0800)] 
x86: Reinstate hand-written unwind info for sysv.S

4 years agox86: Expand FFI_GO_CLOSURE
Richard Henderson [Wed, 10 Dec 2014 17:43:58 +0000 (09:43 -0800)] 
x86: Expand FFI_GO_CLOSURE

If we're going to have to hand-write unwind info for darwin,
these macros make the job harder.

4 years agox86: More Darwin unwind fixups
Iain Sandoe [Tue, 25 Nov 2014 10:43:40 +0000 (11:43 +0100)] 
x86: More Darwin unwind fixups

EHFrame{N} IIRC is a special cue to ld64 that it should treat the unwind
in the object as "special/legacy" .. [these days everything is .cfi_xxxx
(except, cctools-as, as you noted)] .. without that much confusion arises
with ld64's atom-isation of the eh_frame section. labels are not needed for darwin ld64 >= 85.2.1 (i.e. darwin9,
xcode 3.1.4) to all intents and purposes, that's all that matters now,
since I think that anyone trying to build on 10.4/darwin8/xcode2.5 would
have to use a later ld64 (from odcctools) for other reasons.

4 years agox86: Reinstate hand-written unwind info for unix64.S
Richard Henderson [Tue, 25 Nov 2014 08:27:54 +0000 (09:27 +0100)] 
x86: Reinstate hand-written unwind info for unix64.S

One more try to get default Darwin to work.

4 years agox86: Avoid using gas local labels
Richard Henderson [Mon, 24 Nov 2014 15:26:50 +0000 (16:26 +0100)] 
x86: Avoid using gas local labels

Which are unsupported by Darwin cctools as.
Thankfully this doesn't uglify the source too much.

4 years agox86: Remove use of .cfi_escape
Richard Henderson [Mon, 24 Nov 2014 12:02:03 +0000 (13:02 +0100)] 
x86: Remove use of .cfi_escape

The unwind info isn't 100% correct at all points during the epilogue,
and not annotating is just as incorrect as the annotation.  This works
better on systems that do not support DW_OP_call_frame_cfa.

4 years agox86: Honor alignment of arguments
Richard Henderson [Mon, 24 Nov 2014 11:55:43 +0000 (12:55 +0100)] 
x86: Honor alignment of arguments

Darwin aligns long-double to 16, and thus all of the long double
tests were failing due to not honoring that.  We ought to be able
to devise a test case for GCC using __attribute__((aligned)) that
would have failed too.

4 years agox86: Use .balign not .align
Richard Henderson [Mon, 24 Nov 2014 10:24:02 +0000 (11:24 +0100)] 
x86: Use .balign not .align

The Apple assembler defaults to power of two alignment, rather than
byte alignment like everyone else.  Force byte alignment by using
the proper directive.

4 years agox86: Disable .org for Darwin
Richard Henderson [Mon, 24 Nov 2014 09:42:02 +0000 (10:42 +0100)] 
x86: Disable .org for Darwin

4 years agox86: Best guess at update for Darwin
Richard Henderson [Sat, 22 Nov 2014 19:02:43 +0000 (20:02 +0100)] 
x86: Best guess at update for Darwin

4 years agoMerge pull request #158 from rth7680/s390
Anthony Green [Sat, 20 Dec 2014 15:20:40 +0000 (10:20 -0500)] 
Merge pull request #158 from rth7680/s390


4 years agos390: Inline and tidy ffi_prep_args
Richard Henderson [Fri, 19 Dec 2014 16:38:17 +0000 (11:38 -0500)] 
s390: Inline and tidy ffi_prep_args

As per discussion with Ulrich Weigand, document the restrictions
on the code within ffi_call_int as we simultaneously prepare
stack frames for ffi_call_SYSV and the target function.

4 years agos390: Use pc-relative insns in 31-bit mode
Richard Henderson [Thu, 18 Dec 2014 21:21:07 +0000 (16:21 -0500)] 
s390: Use pc-relative insns in 31-bit mode

It's silly to stick to esa/390 features when the compiler won't.
Detect when brasl and larl are used by the compiler and then use
them in the assembly.

4 years agos390: Reorganize assembly
Richard Henderson [Thu, 18 Dec 2014 21:01:59 +0000 (16:01 -0500)] 
s390: Reorganize assembly

Avoid using ffi_prep_args as a callback; do all the work setting
up the frame within ffi_call_int directly.  Save fewer registers
in ffi_closure_SYSV.

4 years agos390: Avoid aliasing warnings
Richard Henderson [Thu, 18 Dec 2014 21:01:15 +0000 (16:01 -0500)] 
s390: Avoid aliasing warnings

4 years agos390: Kill trailing whitespace
Richard Henderson [Wed, 17 Dec 2014 18:24:03 +0000 (13:24 -0500)] 
s390: Kill trailing whitespace

4 years agos390: Go closure support
Dominik Vogt [Wed, 17 Dec 2014 18:20:51 +0000 (13:20 -0500)] 
s390: Go closure support

4 years agoMerge pull request #157 from rth7680/x86
Anthony Green [Fri, 12 Dec 2014 15:13:56 +0000 (10:13 -0500)] 
Merge pull request #157 from rth7680/x86

Two fixes for x86

4 years agoMerge pull request #156 from rth7680/sparc
Anthony Green [Fri, 12 Dec 2014 15:13:27 +0000 (10:13 -0500)] 
Merge pull request #156 from rth7680/sparc


4 years agox86: Handle void arguments as if an empty structure
Richard Henderson [Thu, 11 Dec 2014 22:16:00 +0000 (14:16 -0800)] 
x86: Handle void arguments as if an empty structure

Since libffi currently doesn't allow empty structures, libgo
currently maps them to ffi_type_void.  Given that we'll abort
on this case, handle it gracefully.

4 years agox86: Fix some unwind errors
Richard Henderson [Wed, 10 Dec 2014 21:25:14 +0000 (13:25 -0800)] 
x86: Fix some unwind errors

4 years agoFix for
James Greenhalgh [Sun, 7 Dec 2014 04:58:41 +0000 (23:58 -0500)] 
Fix for

4 years agoMerge pull request #151 from amodra/master
Anthony Green [Thu, 20 Nov 2014 02:13:50 +0000 (21:13 -0500)] 
Merge pull request #151 from amodra/master

powerpc: go closures for linux

4 years agopowerpc: Fix ffi_go_closure_linux64
Richard Henderson [Tue, 18 Nov 2014 02:26:58 +0000 (12:56 +1030)] 
powerpc: Fix ffi_go_closure_linux64

Unlike ffi_closure_LINUX64, this entry point is called normally,
so we already have the TOC in R2 and the closure in R11.

* powerpc/linux64_closure.S (ffi_closure_LINUX64): Remove a
register dependency chain.
(ffi_go_closure_linux64): Don't load r11 or r2.

4 years agoGO closures for powerpc linux
Alan Modra [Sun, 9 Nov 2014 23:12:31 +0000 (09:42 +1030)] 
GO closures for powerpc linux

Plus .cfi async unwind info, rearrangement of ffi_call_linux64 and
ffi_call_SYSV function params to avoid register copies, tweaks to

* src/powerpc/ffitarget.h (FFI_GO_CLOSURES): Define.
* src/powerpc/ffi.c (ffi_call_int): New function with extra
closure param, and args rearranged on ffi_call_linux64 and
ffi_call_SYSV calls, extracted from ..
(ffi_call_go, ffi_prep_go_closure): New functions.
* src/powerpc/ffi_linux64.c (ffi_prep_closure_loc_linux64): Make
hidden.  Only flush insn part of ELFv2 trampoline.  Don't shuffle
ELFv1 trampoline.
(ffi_closure_helper_LINUX64): Replace closure param with cif, fun,
user_data params.
* src/powerpc/ffi_powerpc.h (ffi_go_closure_sysv): Declare.
(ffi_go_closure_linux64): Declare.
(ffi_call_SYSV, fi_call_LINUX64): Update.
(ffi_prep_closure_loc_sysv, ffi_prep_closure_loc_linux64): Declare.
(ffi_closure_helper_SYSV, ffi_closure_helper_LINUX64): Update.
* src/powerpc/ffi_sysv.c (ASM_NEEDS_REGISTERS): Increase to 6.
(ffi_prep_closure_loc_sysv): Use bcl in trampoline, put data words
last, flush just the insn part.
(ffi_closure_helper_SYSV): Replace closure param with cif, fun and
user_data params.
* src/powerpc/linux64.S (ffi_call_LINUX64): Replace hand-written
.eh_frame with .cfi directives.  Adjust for changed param order.
Pass extra "closure" param to user function in static chain.  Add
.cfi directives to describe epilogue.  Don't provide traceback
table for ELFv2 or _CALL_LINUX.
* src/powerpc/linux64_closure.S (ffi_closure_LINUX64): Replace
hand-written .eh_frame with .cfi directives.  Adjust for changed
ffi_closure_helper_LINUX64 params.  Add .cfi directives to
describe epilogue.  Don't provide traceback table for ELFv2 or
(ffi_go_closure_linux64): New function.
* src/powerpc/sysv.S: Remove redundant .globl ffi_prep_args_SYSV.
(ffi_call_SYSV): Make hidden.  Replace hand-written .eh_frame with
.cfi directives.  Adjust for changed params.  Pass extra "closure"
param to user function in static chain.  Add .cfi directives to
describe epilogue.
* src/powerpc/ppc_closure.S (ffi_closure_SYSV): Make hidden.
Replace hand-written .eh_frame with .cfi directives.  Adjust for
changed ffi_closure_helper_SYSV params.  Add .cfi directives to
describe epilogue.  Don't just use nops in the dead __NO_FPRS__
(ffi_go_closure_sysv): New function.

4 years agoFix powerpc breakage from 6e8a4460
Alan Modra [Mon, 10 Nov 2014 03:20:05 +0000 (13:50 +1030)] 
Fix powerpc breakage from 6e8a4460

* src/powerpc/ffitarget.h: #error on unexpected FFI_TYPE_LAST.
(FFI_TYPE_UINT128): Define in terms of FFI_PPC_TYPE_LAST.

4 years agoMerge pull request #148 from gpakosz/dlmalloc-preprocessor-fix
Anthony Green [Wed, 19 Nov 2014 16:32:08 +0000 (11:32 -0500)] 
Merge pull request #148 from gpakosz/dlmalloc-preprocessor-fix

dlmalloc: change defined(i386) to defined(__i386__)

4 years agosparc: Define FFI_TARGET_SPECIFIC_VARIADIC for v9
Richard Henderson [Tue, 18 Nov 2014 13:07:00 +0000 (05:07 -0800)] 

This is a port of

aka GCC PR libffi/60073, to the rewritten Sparc codebase.
Supposedly, we should have seen failures with the existing testcase, but I hadn't.
Perhaps a gcc newer than 4.6.3 is required to see that...

4 years agodlmalloc: change defined(i386) to defined(__i386__)
Gregory Pakosz [Tue, 18 Nov 2014 08:37:07 +0000 (09:37 +0100)] 
dlmalloc: change defined(i386) to defined(__i386__)

When compiling with --std==c99, GCC and clang don't define i386 but __i386__

4 years agoMerge pull request #147 from rth7680/go/alpha
Anthony Green [Mon, 17 Nov 2014 14:07:07 +0000 (09:07 -0500)] 
Merge pull request #147 from rth7680/go/alpha

testsuite: Fix alpha static chain register name

4 years agotestsuite: Fix alpha static chain register name
Richard Henderson [Sun, 16 Nov 2014 11:12:23 +0000 (12:12 +0100)] 
testsuite: Fix alpha static chain register name

4 years agoMerge pull request #145 from rth7680/master
Anthony Green [Sat, 15 Nov 2014 12:31:41 +0000 (07:31 -0500)] 
Merge pull request #145 from rth7680/master

Configure and testsuite cleanups, v2

4 years agoMerge pull request #144 from atrosinenko/testsuite-fp-comparison-fix
Anthony Green [Sat, 15 Nov 2014 12:29:58 +0000 (07:29 -0500)] 
Merge pull request #144 from atrosinenko/testsuite-fp-comparison-fix

Floating point number comparison fix for testsuite

4 years agoMerge pull request #143 from adamkaplan/patch-1
Anthony Green [Sat, 15 Nov 2014 12:27:21 +0000 (07:27 -0500)] 
Merge pull request #143 from adamkaplan/patch-1

Update i386 & armv7 minimum iOS version to 7.0

4 years agox86: Fix typo in ffi_prep_go_closure
Richard Henderson [Fri, 14 Nov 2014 12:05:14 +0000 (13:05 +0100)] 
x86: Fix typo in ffi_prep_go_closure

Used the wrong register for THISCALL and FASTCALL.

4 years agotestsuite: Add trivial tests for Go closures
Richard Henderson [Fri, 14 Nov 2014 12:04:33 +0000 (13:04 +0100)] 
testsuite: Add trivial tests for Go closures

4 years agotestsuite: Move complex tests to their own subdirectory
Richard Henderson [Fri, 14 Nov 2014 10:00:14 +0000 (11:00 +0100)] 
testsuite: Move complex tests to their own subdirectory

It seems a bit silly to isolate them by globbing vs "*complex*"
when we can just as easily put them in their own subdirectory.

4 years agotestsuite: Use feature test rather than enumeration for complex
Richard Henderson [Fri, 14 Nov 2014 09:50:29 +0000 (10:50 +0100)] 
testsuite: Use feature test rather than enumeration for complex

4 years agotestsuite: Detect clang
Richard Henderson [Thu, 13 Nov 2014 17:06:10 +0000 (09:06 -0800)] 
testsuite: Detect clang

Clang doesn't like the -Wno-psabi argument that we want to pass to GCC.
Since clang is detected as GCC via __GNUC__, use ax_cv_c_compiler_vendor.

4 years agoconfigure: Move target source selection into
Richard Henderson [Thu, 13 Nov 2014 12:50:39 +0000 (13:50 +0100)] 
configure: Move target source selection into

This eliminates the AM_CONDITIONAL ugliness, which eliminates
just a bit of extra boilerplate for a new target.

At the same time, properly categorize the EXTRA_DIST files
into SOURCES and HEADERS, for the generation of ctags.

4 years agoTake a float absolute value using fabs() instead of abs().
Anatoly Trosinenko [Fri, 14 Nov 2014 11:21:35 +0000 (14:21 +0300)] 
Take a float absolute value using fabs() instead of abs().

Replace integer abs() by floating point fabs() in the approximate
equality check for float values.

4 years agoFix floating point number comparisons in testsuite/[123].c.
Anatoly Trosinenko [Fri, 14 Nov 2014 10:18:04 +0000 (13:18 +0300)] 
Fix floating point number comparisons in testsuite/[123].c.

Rewrite the checks for approximate equality of floating point return
values to be in the form "fabs(a - b) < EPS" instead of just
"a - b < EPS".

4 years agoUpdate i386 & armv7 minimum iOS version to 7.0
Adam [Thu, 13 Nov 2014 20:08:16 +0000 (15:08 -0500)] 
Update i386 & armv7 minimum iOS version to 7.0

iOS 5 (and any below 7.1) is deprecated. Apple doesn't distribute the supporting libraries for this platform anymore as of Xcode 6, so it causes a linker error.

4 years agoconfigure: Split out
Richard Henderson [Thu, 13 Nov 2014 11:42:54 +0000 (12:42 +0100)] 
configure: Split out

Split out the host case statement to a separate file, so that
we don't have to regenerate for changes therein.

4 years agopowerpc: Delete patch output
Richard Henderson [Thu, 13 Nov 2014 11:32:35 +0000 (12:32 +0100)] 
powerpc: Delete patch output

Clearly added by mistake.

4 years agoMerge branch 'rth7680-go-closure'
Anthony Green [Wed, 12 Nov 2014 13:08:57 +0000 (08:08 -0500)] 
Merge branch 'rth7680-go-closure'

4 years agoUpdate in preparation for next release
Anthony Green [Wed, 12 Nov 2014 13:04:51 +0000 (08:04 -0500)] 
Update in preparation for next release

4 years agoMerge branch 'go-closure' of into rth7680-go-closure
Richard Henderson [Wed, 12 Nov 2014 12:11:53 +0000 (07:11 -0500)] 
Merge branch 'go-closure' of into rth7680-go-closure


4 years agoFix for AArch64. Release as 3.2.1.
Anthony Green [Wed, 12 Nov 2014 12:00:59 +0000 (07:00 -0500)] 
Fix for AArch64. Release as 3.2.1.

4 years agox86: Work around clang bugs
Richard Henderson [Wed, 12 Nov 2014 11:58:58 +0000 (03:58 -0800)] 
x86: Work around clang bugs

4 years agosparc: Re-add abi compliant structure support
Richard Henderson [Sun, 26 Oct 2014 22:29:04 +0000 (15:29 -0700)] 
sparc: Re-add abi compliant structure support

The original code, removed in the "rewrite" patch, was incorrect for
large structures, and required dynamic allocation of a trampoline on
every ffi_call.

Instead, allocate a 4k entry table of all possible structure returns.
The table is 80k, but is read-only and dynamically paged, which ought
to be better than allocating the trampoline.

This is difficult to test with gcc.  One can only use -O0 at present.

4 years agosparc: Add support for Go closures
Richard Henderson [Sun, 26 Oct 2014 21:48:28 +0000 (14:48 -0700)] 
sparc: Add support for Go closures

4 years agosparc: Add support for complex types
Richard Henderson [Sun, 26 Oct 2014 20:16:03 +0000 (13:16 -0700)] 
sparc: Add support for complex types

4 years agosparc: Handle more cases of structure return directly
Richard Henderson [Sat, 25 Oct 2014 23:24:41 +0000 (16:24 -0700)] 
sparc: Handle more cases of structure return directly

4 years agosparc: Preprocess float point struct return
Richard Henderson [Sat, 25 Oct 2014 21:52:45 +0000 (14:52 -0700)] 
sparc: Preprocess float point struct return

We can eliminate recursion and speed structure return
by flattening a nested structure tree into a bitmask.

4 years agosparc: Rewrite everything
Richard Henderson [Fri, 24 Oct 2014 23:10:48 +0000 (16:10 -0700)] 
sparc: Rewrite everything

It's impossible to call between v8 and v9 ABIs, because of the stack bias
in the v9 ABI.  So let's not pretend it's just not implemented yet.  Split
the v9 code out to a separate file.

The register windows prevent ffi_call from setting up the entire stack
frame the assembly, but we needn't make an indirect call back to prep_args.

4 years agosparc: Tidy up symbols
Richard Henderson [Fri, 24 Oct 2014 18:45:53 +0000 (11:45 -0700)] 
sparc: Tidy up symbols

Assembly to use local labels, .type annotation, hidden annotation.
I do retain the _prefix for the symbols, but given that it wasn't
done consistently across all symbols, I doubt it's actually needed.

4 years agosparc: Eliminate long double ifdefs
Richard Henderson [Fri, 24 Oct 2014 18:11:53 +0000 (11:11 -0700)] 
sparc: Eliminate long double ifdefs

4 years agoarm: Add support for Go closures
Richard Henderson [Wed, 29 Oct 2014 18:38:42 +0000 (14:38 -0400)] 
arm: Add support for Go closures

4 years agoarm: Add argument space for the hidden struct return pointer
Richard Henderson [Tue, 21 Oct 2014 15:27:11 +0000 (11:27 -0400)] 
arm: Add argument space for the hidden struct return pointer

This should have been failing all along, but it's only
exposed by the complex_int test case.

4 years agoarm: Add support for complex types
Richard Henderson [Tue, 21 Oct 2014 15:26:59 +0000 (11:26 -0400)] 
arm: Add support for complex types

4 years agoarm: Remove internal FFI_TYPE constants
Richard Henderson [Mon, 20 Oct 2014 19:10:43 +0000 (15:10 -0400)] 
arm: Remove internal FFI_TYPE constants

These have been replaced by the contents of internal.h.

4 years agoarm: Rewrite ffi_closure
Richard Henderson [Fri, 17 Oct 2014 06:07:32 +0000 (02:07 -0400)] 
arm: Rewrite ffi_closure

Move the push of the argument registers into ffi_closure_SYSV,
reducing the size of the trampoline.

4 years agoarm: Rewrite ffi_call
Richard Henderson [Fri, 17 Oct 2014 05:27:16 +0000 (01:27 -0400)] 
arm: Rewrite ffi_call

Use the trick to allocate the stack frame for ffi_call_SYSV
within ffi_call itself.

4 years agoarm: Rewrite vfp_type_p
Richard Henderson [Fri, 17 Oct 2014 05:21:22 +0000 (01:21 -0400)] 
arm: Rewrite vfp_type_p

Do not modify the ffi_type.  Rearrange the tests so that we
quickly eliminate structures that cannot match.  Return an
encoded value of element count and base type.

4 years agoarm: Deref ffi_put_arg arguments
Richard Henderson [Fri, 17 Oct 2014 05:02:52 +0000 (01:02 -0400)] 
arm: Deref ffi_put_arg arguments

4 years agoarm: Deref ffi_align argument
Richard Henderson [Fri, 17 Oct 2014 04:53:21 +0000 (00:53 -0400)] 
arm: Deref ffi_align argument

4 years agoarm: Reindent arm/ffi.c
Richard Henderson [Wed, 15 Oct 2014 21:28:53 +0000 (17:28 -0400)] 
arm: Reindent arm/ffi.c

4 years agoalpha: Add support for Go closures
Richard Henderson [Sat, 18 Oct 2014 04:26:52 +0000 (21:26 -0700)] 
alpha: Add support for Go closures

4 years agoalpha: Add support for complex types
Richard Henderson [Sat, 18 Oct 2014 03:46:48 +0000 (20:46 -0700)] 
alpha: Add support for complex types

4 years agoalpha: Clean up conversion of float values
Richard Henderson [Fri, 17 Oct 2014 20:47:26 +0000 (13:47 -0700)] 
alpha: Clean up conversion of float values

Don't use "real" conversion to double, lest we raise
exceptions when passing signalling nans.

4 years agoalpha: Reorganize cif flags
Richard Henderson [Fri, 17 Oct 2014 19:33:53 +0000 (12:33 -0700)] 
alpha: Reorganize cif flags

Unties the backend from changes to FFI_TYPE_* constants, and allows
compilation to succeed after the addition of FFI_TYPE_COMPLEX.

Delete the hand-written unwind info.

4 years agoaarch64: Add support for Go closures
Richard Henderson [Thu, 23 Oct 2014 04:26:14 +0000 (00:26 -0400)] 
aarch64: Add support for Go closures

4 years agoaarch64: Move x8 out of call_context
Richard Henderson [Thu, 23 Oct 2014 03:48:12 +0000 (23:48 -0400)] 
aarch64: Move x8 out of call_context

Reduces stack size.  It was only used by the closure, and there
are available argument registers.

4 years agoaarch64: Add support for complex types
Richard Henderson [Thu, 23 Oct 2014 02:58:09 +0000 (22:58 -0400)] 
aarch64: Add support for complex types

4 years agoaarch64: Remove aarch64_flags
Richard Henderson [Thu, 23 Oct 2014 02:36:07 +0000 (22:36 -0400)] 
aarch64: Remove aarch64_flags

This field was useless from the start, since the normal flags
field is available for backend use.

4 years agoaarch64: Unify scalar fp and hfa handling
Richard Henderson [Thu, 23 Oct 2014 02:32:13 +0000 (22:32 -0400)] 
aarch64: Unify scalar fp and hfa handling

Since an HFA of a single element is exactly the same as scalar,
this tidies things up a bit.

4 years agoaarch64: Move return value handling into ffi_closure_SYSV
Richard Henderson [Thu, 23 Oct 2014 01:53:30 +0000 (21:53 -0400)] 
aarch64: Move return value handling into ffi_closure_SYSV

As with the change to ffi_call_SYSV, this avoids copying data
into a temporary buffer.

4 years agoaarch64: Move return value handling into ffi_call_SYSV
Richard Henderson [Wed, 22 Oct 2014 21:06:19 +0000 (17:06 -0400)] 
aarch64: Move return value handling into ffi_call_SYSV

This lets us pass return data directly to the caller of ffi_call
in most cases, rather than storing it into temporary storage first.

4 years agoaarch64: Merge prep_args with ffi_call
Richard Henderson [Wed, 22 Oct 2014 17:58:59 +0000 (13:58 -0400)] 
aarch64: Merge prep_args with ffi_call

Use the trick to allocate the stack frame for ffi_call_SYSV
within ffi_call itself.

4 years agoaarch64: Tidy up abi manipulation
Richard Henderson [Wed, 22 Oct 2014 16:52:07 +0000 (12:52 -0400)] 
aarch64: Tidy up abi manipulation

Avoid false abstraction, like get_x_addr.  Avoid recomputing data
about the type being manipulated.  Use NEON insns for HFA manipulation.

Note that some of the inline assembly will go away in a subsequent patch.

4 years agoaarch64: Treat void return as not passed in registers
Richard Henderson [Wed, 22 Oct 2014 16:33:59 +0000 (12:33 -0400)] 
aarch64: Treat void return as not passed in registers

This lets us do less post-processing when there's no return value.

4 years agoaarch64: Use correct return registers
Richard Henderson [Wed, 22 Oct 2014 02:49:05 +0000 (22:49 -0400)] 
aarch64: Use correct return registers

There are fewer return registers than argument registers.

4 years agoaarch64: Reduce the size of register_context
Richard Henderson [Wed, 22 Oct 2014 02:41:07 +0000 (22:41 -0400)] 
aarch64: Reduce the size of register_context

We don't need to store 32 general and vector registers.
Only 8 of each are used for parameter passing.

4 years agoaarch64: Simplify AARCH64_STACK_ALIGN
Richard Henderson [Tue, 21 Oct 2014 17:30:40 +0000 (13:30 -0400)] 
aarch64: Simplify AARCH64_STACK_ALIGN

The iOS abi doesn't require padding between arguments, but
that's not what AARCH64_STACK_ALIGN meant.  The hardware will
in fact trap if the SP register is not 16 byte aligned.

4 years agoaarch64: Always distinguish LONGDOUBLE
Richard Henderson [Tue, 21 Oct 2014 17:27:57 +0000 (13:27 -0400)] 
aarch64: Always distinguish LONGDOUBLE

Avoid if-deffery by forcing FFI_TYPE_LONGDOUBLE different
from FFI_TYPE_DOUBLE.  This will simply be unused on hosts
that define them identically.

4 years agoaarch64: Improve is_hfa
Richard Henderson [Tue, 21 Oct 2014 17:17:39 +0000 (13:17 -0400)] 
aarch64: Improve is_hfa

The set of functions get_homogeneous_type, element_count, and is_hfa
are all intertwined and recompute data.  Return a compound quantity
from is_hfa that contains all the data and avoids the recomputation.

4 years agoaarch64: Fix non-apple compilation
Richard Henderson [Tue, 21 Oct 2014 17:00:34 +0000 (13:00 -0400)] 
aarch64: Fix non-apple compilation

4 years agox86: Use win32 name mangling for fastcall functions
Richard Henderson [Thu, 6 Nov 2014 18:57:04 +0000 (10:57 -0800)] 
x86: Use win32 name mangling for fastcall functions

4 years agox86: Add support for Go closures
Richard Henderson [Wed, 5 Nov 2014 16:04:29 +0000 (17:04 +0100)] 
x86: Add support for Go closures

4 years agox86: Add support for Complex
Richard Henderson [Wed, 5 Nov 2014 15:34:41 +0000 (16:34 +0100)] 
x86: Add support for Complex

4 years agox86: Rewrite closures
Richard Henderson [Wed, 5 Nov 2014 09:15:25 +0000 (10:15 +0100)] 
x86: Rewrite closures

Move everything into sysv.S, removing win32.S and freebsd.S.
Handle all abis with a single ffi_closure_inner function.
Move complexity of the raw THISCALL trampoline into assembly
instead of the trampoline itself.
Only push the context for the REGISTER abi; let the rest
receive it in a register.

4 years agox86: Rewrite ffi_call
Richard Henderson [Sat, 1 Nov 2014 22:10:34 +0000 (15:10 -0700)] 
x86: Rewrite ffi_call

Decouple the assembly from FFI_TYPE_*.  Merge prep_args with ffi_call,
passing the frame and the stack to the assembly.

Note that this patch isn't really standalone, as this breaks closures.

4 years agox86: Convert to gas generated unwind info
Richard Henderson [Fri, 31 Oct 2014 19:07:02 +0000 (12:07 -0700)] 
x86: Convert to gas generated unwind info

4 years agox86: Force FFI_TYPE_LONGDOUBLE different from FFI_TYPE_DOUBLE
Richard Henderson [Thu, 30 Oct 2014 20:57:39 +0000 (13:57 -0700)] 

There are few abis that set double = long double.  Eliminate the
conditional compilation and let this code simply be unused there.

4 years agox86: Remove some conditional compilation
Richard Henderson [Thu, 30 Oct 2014 19:41:31 +0000 (12:41 -0700)] 
x86: Remove some conditional compilation

Removal of ifdefs made possible to due to ffi_abi unification.

4 years agox86: Tidy ffi_abi
Richard Henderson [Thu, 30 Oct 2014 19:13:31 +0000 (12:13 -0700)] 
x86: Tidy ffi_abi

The x86_64 unix port only handles one ABI; don't define all of the
other symbols.  The UNIX64 symbol retains the same value.

The i386 ports ought to have the same symbols, even if we can't yet
unify the values without incrementing the libffi soname.

4 years agotestsuite: Add two dg-do run markers
Richard Henderson [Fri, 7 Nov 2014 14:56:55 +0000 (06:56 -0800)] 
testsuite: Add two dg-do run markers

Caught by clang warning about unused -L parameter.

4 years agotestsuite: Fix return_complex2 vs excessive precision
Richard Henderson [Wed, 5 Nov 2014 15:33:44 +0000 (16:33 +0100)] 
testsuite: Fix return_complex2 vs excessive precision

Use the previously computed rc2 to validate, rather than
recomputing a floating point result with excess precision.

4 years agox86_64: Add support for complex types
Richard Henderson [Tue, 28 Oct 2014 18:21:50 +0000 (11:21 -0700)] 
x86_64: Add support for complex types