Dragonegg miscompiles dlls/d3dx9_36/tests/math.c from Wine

[llvmbot]

Bugzilla Link	10640
Version	2.9
OS	Linux
Depends On	llvm/llvm-bugzilla-archive#10704
Blocks	llvm/llvm-bugzilla-archive#10682
Attachments	llvm preprocessed, gcc preprocessed, reduced testcase
Reporter	LLVM Bugzilla Contributor

Extended Description

Dragonegg can compile wine successfully, aside from http://bugs.winehq.org/show_bug.cgi?id=28050. Running its test suite, however, shows a few problems.

To reproduce, compile wine with:
$ CC=llvm-gcc ./configure
$ make

then run the test:
$ cd dlls/d3dx9_36/tests
$ make math.ok

this will fail:
llvm-gcc -c -I. -I. -I../../../include -I../../../include -DWINE_STRICT_PROTOTYPES -DWINE_NO_NAMELESS_EXTENSION -D_REENTRANT -fPIC -Wall -pipe -fno-strict-aliasing -Wdeclaration-after-statement -Wempty-body -Wstrict-prototypes -Wtype-limits -Wwrite-strings -Wpointer-arith -Wlogical-op -g -O2 -o math.o math.c
../../../tools/winegcc/winegcc -B../../../tools/winebuild --sysroot=../../.. -fasynchronous-unwind-tables asm.o core.o effect.o line.o math.o mesh.o shader.o surface.o texture.o rsrc.res testlist.o -o d3dx9_36_test.exe.so ../../../libs/port/libwine_port.a -ld3dx9 -ld3d9 -luser32 -lgdi32
../../../tools/runtest -q -P wine -M d3dx9_36.dll -T ../../.. -p d3dx9_36_test.exe.so math.c && touch math.ok
fixme:d3dx:D3DXQuaternionLn The quaternion (0.200000, 0.100000, 0.300000, 0.900000) has a norm <1. This should not happen. Windows returns a result anyway. This case is not implemented yet.
math.c:2278: Test failed: Got ffff, expected 7fff or 7fff for index 18.
math.c:2278: Test failed: Got 7fff, expected ffff or ffff for index 19.
make: *** [math.ok] Error 2

compiling math.c with gcc, or using -O0 instead of 01/02 works around the issue. I'll attach preprocessed source files from llvm-gcc and gcc (versions below), as well as a reduced down version of math.c that more clearly shows the problem. If more info is needed, please don't hesitate to ask.

austin@debian:~/wine-llvmgcc/dlls/crypt32/tests$ llvm-gcc --version
gcc-4.5 (Debian 4.5.3-3) 4.5.3
Copyright (C) 2010 Free Software Foundation, Inc.
This is free software; see the source for copying conditions. There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

austin@debian:~/wine-llvmgcc/dlls/crypt32/tests$ gcc --version
gcc (Debian 4.4.6-7) 4.4.6
Copyright (C) 2010 Free Software Foundation, Inc.
This is free software; see the source for copying conditions. There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

[llvmbot]

I noticed a maths test failing on my 64 bit machine because GCC generates code
that uses the floating point stack, while dragonegg is using SSE, resulting in
different results due to rounding being changed by this. I've opened by 10704
to track this issue.

[llvmbot]

mentioned in issue llvm/llvm-bugzilla-archive#10682

[llvmbot]

mentioned in issue llvm/llvm-bugzilla-archive#10704

[asl]

Dragonegg is obsolete