Enable MSVC EH for Win64 #1354

kinke · 2016-03-12T17:58:53Z

Enable @rainers #1168 for Win64 too and add x86 AppVeyor jobs.
Requires ldc-developers/druntime#65 and a post-3.8 LLVM (Rainer's work has uncovered a few LLVM bugs whose near-instant fixes didn't make it into 3.8 final).

kinke · 2016-03-13T00:35:22Z

What AppVeyor doesn't show is that dmd-testsuite (debug+release) finally passes, for both 32bit and 64bit MSVC targets! I.e. for the very first time for LDC (on Windows in general) in case you're wondering.
Awesome work Rainer, thanks again. And thanks to @majnemer and all the other LLVM guys. 👍

dnadlinger · 2016-03-13T01:13:38Z

Nice!

smolt · 2016-03-13T01:41:11Z

I second that 👍

Just noticed the Add You Reaction drop down. This deserves the party hat.

rainers · 2016-03-13T10:20:46Z

@kinke Thanks for pushing this.

And thanks to @majnemer and all the other LLVM guys

I second that, too.

kinke · 2016-03-13T14:11:11Z

To recap: the only (very minor) issues leading to test failures on Windows after this are:

~~enable and fix tests of std.experimental.allocator for Win64/Win32mscoff dlang/phobos#4034 (32-bit only)~~
Fix NaN payload unittest in errorfunction dlang/phobos#3976 (32-bit release only)
LDC IR test issue ~~IR tests fail on Windows for target x86 #1265 (32-bit only due to special mangling and calling convention)~~ 32-bit x86: Alignment issues for LL byval params #1356 (32-bit only)
A Win64-specific fibers unittest in core.thread not working with inlining (a problem of the test itself).

redstar · 2016-03-13T14:52:27Z

Very well done! @kinke I think you deserve the honor to push the merge button. :-)

…nctions with EH

…ret"

…nclet" token

…mpletely the same as C++

… as "noreturn" by the optimizer - do not inline functions into the cleanup funclets, it can break exception handling - even if no var given, retrieve the caught exception to pass it to _d_eh_enter_catch

…cleanup

…ow_bug.cgi?id=26833

Newest LLVM takes care of this for us.

kinke · 2016-03-17T09:11:05Z

dmd-testsuite is now enabled too for the AppVeyor jobs (it doesn't take that much longer). There are some minor path-related issues I don't have on my local system, probably due to a different set of GNU utils (I assume), e.g.:

... compilable/test9680.sh
Error: Could not create output directory 'C:C:/Program Files/Git/projects/ninja-ldc/dmd-testsuite-debug/compilable'.

Otherwise, there are only issues wrt. Dan's new runnable\ldc_cabi1.d (on my system too - I've previously used a dmd-testsuite from a couple of weeks ago). The x64 job hits LLVM assertions - apparently it doesn't like the align 16 attribute:

Wrong types for attribute: byval inalloca nest noalias nocapture nonnull readnone readonly signext sret zeroext dereferenceable(1) dereferenceable_or_null(1)
  call void (i32, ...) @cvfun(i32 3, %ldc_cabi1.F4 noalias nocapture align 16 %509), !dbg !913
Wrong types for attribute: byval inalloca nest noalias nocapture nonnull readnone readonly signext sret zeroext dereferenceable(1) dereferenceable_or_null(1)
  call void (i32, ...) @cvfun(i32 4, %ldc_cabi1.D4 noalias nocapture align 16 %536), !dbg !918
Wrong types for attribute: byval inalloca nest noalias nocapture nonnull readnone readonly signext sret zeroext dereferenceable(1) dereferenceable_or_null(1)
  call void (i32, ...) @cvfun(i32 5, %ldc_cabi1.DHFA2 noalias nocapture align 16 %552), !dbg !921
Wrong types for attribute: byval inalloca nest noalias nocapture nonnull readnone readonly signext sret zeroext dereferenceable(1) dereferenceable_or_null(1)
  call void (i32, ...) @cvfun(i32 6, %ldc_cabi1.DHFA2a noalias nocapture align 16 %568), !dbg !924
LLVM ERROR: Broken function found, compilation aborted!

The x86 job uncovers mangling issues:

ldc_cabi2.cpp.obj : error LNK2019: unresolved external symbol "signed char a" (?a@@3CA) referenced in function _cretb1
ldc_cabi2.cpp.obj : error LNK2019: unresolved external symbol "signed char b" (?b@@3CA) referenced in function _cretb2
ldc_cabi2.cpp.obj : error LNK2019: unresolved external symbol "signed char c" (?c@@3CA) referenced in function _cretf4
ldc_cabi2.cpp.obj : error LNK2019: unresolved external symbol "signed char d" (?d@@3CA) referenced in function _cretf4
ldc_cabi2.cpp.obj : error LNK2019: unresolved external symbol "unsigned int errors" (?errors@@3IA) referenced in function _cretb1

~~Oh and dmd-testsuite's exit code somehow isn't ignored, although I use ctest ... || echo .... That'll need to be fixed too.~~

smolt · 2016-03-17T15:18:15Z

The mangling error I think is a flaw in ldc_cabi2.cpp. Those vars should be extern "C".

kinke · 2016-03-17T18:59:01Z

Those vars should be extern "C".

That's right, I'll fix them, thanks for the hint. Edit: Why hasn't this hit us before then? Let's see how Travis reacts...
Now we hit some ldc_cabi1 failures for Win32.
The (current) Win64 issue only applies to variadic functions.

But ignore failures for now.

kinke · 2016-03-17T22:52:34Z

dmd-testsuite passes again on Win64, both on my box and for the AppVeyor x64 Debug job.

kinke · 2016-03-18T17:07:43Z

The Win32 ldc_cabi1 failures are also vararg related. Fix is on its way.

As for regular args due to LDC issue ldc-developers#1356. This fixes runnable/ldc_cabi1 and so dmd-testsuite passes again.

kinke · 2016-03-18T23:01:47Z

Alright, now all tests are enabled for AppVeyor, but lit test failures are ignored for x86 jobs because of align.d failing due to issue #1356.

The debug jobs are green. The release ones fail (and are allowed to fail)

on Win64 only because of the inline-incompatible core.thread unittest (Win64 specific test, hence no problem for all other platforms), and
on Win32 only due to @smolt's Fix NaN payload unittest in errorfunction dlang/phobos#3976 not having been merged yet.

This really elevates Windows MSVC to a first class target for LDC. As I mentioned earlier, it depends on a post-3.8 LLVM. @rainers even included a TLS alignment bugfix for Windows < 8.1. The recent addition of alignment attributes for IR arguments has uncovered quite a few alignment issues due to the optimizer apparently making good use of that information, so that's a good sign too.

smolt · 2016-03-18T23:24:33Z

This is awesome!

Enable MSVC EH for Win64

CyberShadow · 2016-04-19T10:41:57Z

appveyor.yml

+  - dmd2\windows\bin\dmd.exe --version
+  # Download & extract GNU make + utils (for dmd-testsuite)
+  - bash --version
+  - ps: Start-FileDownload 'https://dl.dropboxusercontent.com/s/d0cqmxf9arbjn27/make-3.81.7z?dl=0' -FileName 'make.7z'


Hey, just wanted to ask, what's the original source of this file?
(Working on a test command for Digger...)
Figured it out, GnuWin32

Yes, but it contains some additional utils. See http://wiki.dlang.org/Building_and_hacking_LDC_on_Windows_using_MSVC#Running_the_dmd-testsuite_tests.

kinke force-pushed the msvcEH branch 2 times, most recently from 4a68f45 to ae0f049 Compare March 12, 2016 21:51

rainers added 19 commits March 13, 2016 17:27

thunks don't need a personality

c970d5d

MSVC/x86 EH support using __CxxFrameHandler3 personality

aad2706

fix MSVC-EH again, disable inlining and tail call optimization for fu…

5dabf78

…nctions with EH

inside a cleanup block, convert "unreachable" to a branch to "cleanup…

73ab5f4

…ret"

fix build for LLVM < 3.8

141f099

avoid replacing library functions in funclet, it doesn't copy the "fu…

4b2e781

…nclet" token

remove setting NoBuiltin attribute, fixed in LLVM

a8d5c46

use catch-all and rethrow instead of cleanup pads, they don't work co…

5b22904

…mpletely the same as C++

- add a global condition to cleanup pads to avoid them being inferred…

858128a

… as "noreturn" by the optimizer - do not inline functions into the cleanup funclets, it can break exception handling - even if no var given, retrieve the caught exception to pass it to _d_eh_enter_catch

replace _d_skipCleanup var with callbacks _d_enter_cleanup, _d_leave_…

2a86b72

…cleanup

- add classinfo argument to _d_eh_enter_catch

8336890

eh unwinding: don't alloc data in parent frame, just use nullptr for now

d8bbc04

avoid outlining catch handler

18b1492

_d_enter_catch can throw, so it might have to be invoked

3adf19c

fix compilation after merge

1f41496

cleanup obsolete changes

03e7834

fix ordering of overloads

4a200b7

add debug info to thunks as a workaround for https://llvm.org/bugs/sh…

6f3adae

…ow_bug.cgi?id=26833

fix function name mangling for debug info

1131e6f

kinke force-pushed the msvcEH branch 2 times, most recently from 9c6f491 to ca1241f Compare March 16, 2016 18:19

MSVC: Don't truncate debug info names anymore

6a6f6be

Newest LLVM takes care of this for us.

kinke force-pushed the msvcEH branch 5 times, most recently from d67dd7d to 7ae0223 Compare March 17, 2016 01:24

kinke mentioned this pull request Mar 17, 2016

AppVeyor does not execute the dmd-testsuite #1338

Closed

kinke added 2 commits March 17, 2016 20:41

AppVeyor: Enable dmd-testsuite

2dca401

But ignore failures for now.

dmd-testsuite: Tiny mangling fix for runnable/ldc_cabi1.d

c5bedcb

kinke force-pushed the msvcEH branch from 989e314 to c5bedcb Compare March 17, 2016 19:41

Win64: Fix ICE when passing varargs > 64 bit

3e20e4c

kinke mentioned this pull request Mar 18, 2016

Instead of using the C++-linker, use the D compiler to link LDC2 and LDMD #1361

Closed

kinke force-pushed the msvcEH branch 2 times, most recently from 3c7f55a to 65c868a Compare March 18, 2016 19:51

Win32: Remove alignment attributes for byval varargs

97d6a12

As for regular args due to LDC issue ldc-developers#1356. This fixes runnable/ldc_cabi1 and so dmd-testsuite passes again.

kinke force-pushed the msvcEH branch from 65c868a to 97d6a12 Compare March 18, 2016 20:24

kinke mentioned this pull request Mar 18, 2016

fix TLS alignment for Windows 8.0 or below ldc-developers/druntime#66

Closed

druntime: fix TLS alignment for Windows 8.0 or below

155017e

kinke force-pushed the msvcEH branch from 8a2d030 to 155017e Compare March 19, 2016 08:56

kinke added a commit that referenced this pull request Mar 19, 2016

Merge pull request #1354 from kinke/msvcEH

1481dea

Enable MSVC EH for Win64

kinke merged commit 1481dea into ldc-developers:master Mar 19, 2016

kinke mentioned this pull request Mar 19, 2016

[WIP] support building for x86 with MSVC runtime #1168

Closed

kinke deleted the msvcEH branch March 19, 2016 10:17

CyberShadow reviewed Apr 19, 2016
View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Enable MSVC EH for Win64 #1354

Enable MSVC EH for Win64 #1354

kinke commented Mar 12, 2016

kinke commented Mar 13, 2016

dnadlinger commented Mar 13, 2016

smolt commented Mar 13, 2016

rainers commented Mar 13, 2016

kinke commented Mar 13, 2016

redstar commented Mar 13, 2016

kinke commented Mar 17, 2016

smolt commented Mar 17, 2016

kinke commented Mar 17, 2016

kinke commented Mar 17, 2016

kinke commented Mar 18, 2016

kinke commented Mar 18, 2016

smolt commented Mar 18, 2016

CyberShadow Apr 19, 2016 •

edited

Loading

kinke Apr 20, 2016

Enable MSVC EH for Win64 #1354

Enable MSVC EH for Win64 #1354

Conversation

kinke commented Mar 12, 2016

kinke commented Mar 13, 2016

dnadlinger commented Mar 13, 2016

smolt commented Mar 13, 2016

rainers commented Mar 13, 2016

kinke commented Mar 13, 2016

redstar commented Mar 13, 2016

kinke commented Mar 17, 2016

smolt commented Mar 17, 2016

kinke commented Mar 17, 2016

kinke commented Mar 17, 2016

kinke commented Mar 18, 2016

kinke commented Mar 18, 2016

smolt commented Mar 18, 2016

CyberShadow Apr 19, 2016 • edited Loading

Choose a reason for hiding this comment

kinke Apr 20, 2016

Choose a reason for hiding this comment

CyberShadow Apr 19, 2016 •

edited

Loading