New subject: [Bug 46079] Support 32-bit ARM hard-float cross-toolchains with GNU binutils assembler that doesn 't implicitly assume '.fpu {vfp,neon}' (Yocto/Poky SDK)

3 Nov 2018


      https://bugs.winehq.org/show_bug.cgi?id=46079
Bug ID: 46079
           Summary: Support 32-bit ARM hard-float cross-toolchains with
                    GNU binutils assembler that doesn't implicitly assume
                    '.fpu {vfp,neon}' (Yocto/Poky SDK)
           Product: Wine
           Version: 3.19
          Hardware: x86-64
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P2
         Component: tools
          Assignee: wine-bugs@winehq.org
          Reporter: focht@gmx.net
      Distribution: ---
Hello folks,
after working around bug 46053 ("Support cross-toolchains that need target
sysroot explicitly passed via ' --sysroot' command-line option (Yocto/Poky
SDK)") by having 'winegcc' passing the target sysroot to the
cross-compiler/linker via '--sysroot', aarch64 Wine now cross compiles
successfully with a full target sysroot.
-> aarch64-poky-linux
-------------------------------------------------------------------------
Build Configuration:
BB_VERSION           = "1.38.0"
BUILD_SYS            = "x86_64-linux"
NATIVELSBSTRING      = "fedora-28"
TARGET_SYS           = "aarch64-poky-linux"
MACHINE              = "qemuarm64"
DISTRO               = "poky"
DISTRO_VERSION       = "2.5.1"
TUNE_FEATURES        = "aarch64"
TARGET_FPU           = ""
meta                 
meta-poky            
meta-yocto-bsp       = "sumo:d240b885f26e9b05c8db0364ab2ace9796709aad"
meta-oe              = "sumo:2bb21ef27c4b0c9d52d30b3b2c5a0160fd02b966"
meta-clang           = "sumo:1784b06e6b1a83748438b4bec7e079d735d5b608"
...
-------------------------------------------------------------------------
Building for 32-bit ARMv7 hard-float fails (using the same mechanics).
-> arm-poky-linux-gnueabi (cortexa9hf-vfpv3-poky-linux-gnueabi)
-------------------------------------------------------------------------
Build Configuration:
BB_VERSION           = "1.38.0"
BUILD_SYS            = "x86_64-linux"
NATIVELSBSTRING      = "fedora-28"
TARGET_SYS           = "arm-poky-linux-gnueabi"
MACHINE              = "qemuarma9"
DISTRO               = "poky"
DISTRO_VERSION       = "2.5.1"
TUNE_FEATURES        = "arm armv7a vfp vfpv3d16 vfpv3 callconvention-hard
cortexa9"
TARGET_FPU           = "hard"
meta                 
meta-poky            
meta-yocto-bsp       = "sumo:d240b885f26e9b05c8db0364ab2ace9796709aad"
meta-oe              = "sumo:2bb21ef27c4b0c9d52d30b3b2c5a0160fd02b966"
meta-clang           = "sumo:1784b06e6b1a83748438b4bec7e079d735d5b608"
...
-------------------------------------------------------------------------
--- snip ---
...
/home/focht/projects/wine/mainline-build-x86_64/tools/winegcc/winegcc -o
d2d1.dll.so \
  -B/home/focht/projects/wine/mainline-build-x86_64/tools/winebuild
--sysroot=../.. -b \
  arm-poky-linux-gnueabi -marm -mfloat-abi=hard -fasynchronous-unwind-tables
-shared /home/focht/projects/wine/mainline-src/dlls/d2d1/d2d1.spec \
  bitmap.o bitmap_render_target.o brush.o dc_render_target.o device.o effect.o
factory.o geometry.o \
  hwnd_render_target.o layer.o mesh.o state_block.o stroke.o
wic_render_target.o version.res \
  -ldwrite -ld3d10_1 ../../dlls/dxguid/libdxguid.a ../../dlls/uuid/libuuid.a
-lgdi32 -luser32 \
  -ladvapi32 ../../libs/port/libwine_port.a -Wb,-ddwrite -Wl,-O1
-Wl,--hash-style=gnu -Wl,--as-needed
d2d1.IYX82U.s: Assembler messages:
d2d1.IYX82U.s:291: Error: selected processor does not support `vpush {s0-s15}'
in ARM mode
d2d1.IYX82U.s:310: Error: selected processor does not support `vpush {s0-s15}'
in ARM mode
winebuild: arm-poky-linux-gnueabi-gcc failed with status 1
winegcc:
/home/focht/projects/wine/mainline-build-x86_64/tools/winebuild/winebuild
failed
make[1]: *** [Makefile:1440: d2d1.dll.so] Error 2
make[1]: Leaving directory
'/home/focht/projects/wine/mainline-build-arm/dlls/d2d1'
make: *** [Makefile:8385: dlls/d2d1] Error 2
--- snip ---
Re-executing failing last command with '-v -save-temps':
--- snip ---
/home/focht/projects/wine/mainline-build-x86_64/tools/winegcc/winegcc -v
-save-temps -o d2d1.dll.so  
-B/home/focht/projects/wine/mainline-build-x86_64/tools/winebuild
--sysroot=../.. -b   arm-poky-linux-gnueabi -marm -mfloat-abi=hard
-fasynchronous-unwind-tables -shared
/home/focht/projects/wine/mainline-src/dlls/d2d1/d2d1.spec   bitmap.o
bitmap_render_target.o brush.o dc_render_target.o device.o effect.o factory.o
geometry.o   hwnd_render_target.o layer.o mesh.o state_block.o stroke.o
wic_render_target.o version.res   -ldwrite -ld3d10_1
../../dlls/dxguid/libdxguid.a ../../dlls/uuid/libuuid.a -lgdi32 -luser32  
-ladvapi32 ../../libs/port/libwine_port.a -Wb,-ddwrite -Wl,-O1
-Wl,--hash-style=gnu -Wl,--as-needed
/home/focht/projects/wine/mainline-build-x86_64/tools/winebuild/winebuild -v
--save-temps --target arm-poky-linux-gnueabi -fasynchronous-unwind-tables
--cc-cmd=arm-poky-linux-gnueabi-gcc --ld-cmd=arm-poky-linux-gnueabi-ld
-D_REENTRANT -fPIC --dll -o d2d1.dll-IpMxH8.spec.o -E
/home/focht/projects/wine/mainline-src/dlls/d2d1/d2d1.spec -L../../dlls
-L../../libs/wine -marm -mfloat-abi=hard -ddwrite -rversion.res -- bitmap.o
bitmap_render_target.o brush.o dc_render_target.o device.o effect.o factory.o
geometry.o hwnd_render_target.o layer.o mesh.o state_block.o stroke.o
wic_render_target.o ../../dlls/dwrite/libdwrite.def
../../dlls/d3d10_1/libd3d10_1.def ../../dlls/dxguid/libdxguid.a
../../dlls/uuid/libuuid.a ../../dlls/gdi32/libgdi32.def
../../dlls/user32/libuser32.def ../../dlls/advapi32/libadvapi32.def
../../libs/port/libwine_port.a ../../dlls/winecrt0/libwinecrt0.a
../../dlls/kernel32/libkernel32.def ../../dlls/ntdll/libntdll.def 
arm-poky-linux-gnueabi-gcc -xassembler -c -o d2d1.nY1v31.o d2d1.CJoop5.s
arm-poky-linux-gnueabi-ld -r -o d2d1.KKS2kZ.o d2d1.nY1v31.o bitmap.o
bitmap_render_target.o brush.o dc_render_target.o device.o effect.o factory.o
geometry.o hwnd_render_target.o layer.o mesh.o state_block.o stroke.o
wic_render_target.o ../../dlls/dxguid/libdxguid.a ../../dlls/uuid/libuuid.a
../../libs/port/libwine_port.a ../../dlls/winecrt0/libwinecrt0.a
arm-poky-linux-gnueabi-gcc -xassembler -c -o d2d1.dll-IpMxH8.spec.o
d2d1.vn7wK8.s
d2d1.vn7wK8.s: Assembler messages:
d2d1.vn7wK8.s:291: Error: selected processor does not support `vpush {s0-s15}'
in ARM mode
d2d1.vn7wK8.s:310: Error: selected processor does not support `vpush {s0-s15}'
in ARM mode
winebuild: arm-poky-linux-gnueabi-gcc failed with status 1
winegcc:
/home/focht/projects/wine/mainline-build-x86_64/tools/winebuild/winebuild
failed
--- snip ---
$ cat d2d1.vn7wK8.s
/* File generated automatically from
/home/focht/projects/wine/mainline-src/dlls/d2d1/d2d1.spec; do not edit! */
/* This file can be copied, modified and distributed without restriction. */
.section ".text","ax"
    b 1f
__wine_spec_pe_header:
    .skip 69632
1:
.data
    .align 2
    .globl __wine_spec_nt_header
    .hidden __wine_spec_nt_header
__wine_spec_nt_header:
.L__wine_spec_rva_base:
...
.L__wine_spec_exports_funcs:
    .long D2D1CreateFactory
    .long D2D1MakeRotateMatrix
    .long D2D1MakeSkewMatrix
    .long D2D1IsMatrixInvertible
    .long D2D1InvertMatrix
    .long __wine_stub_D2D1ConvertColorSpace
    .long __wine_stub_D2D1CreateDevice
    .long __wine_stub_D2D1CreateDeviceContext
    .long __wine_stub_D2D1SinCos
    .long __wine_stub_D2D1Tan
    .long __wine_stub_D2D1Vec3Length
...
__wine_spec_relay_entry_points:
    nop
    .align 2
.L__wine_spec_relay_entry_point_1:
        .cfi_startproc
    push {r0-r3}
    mov r2, SP
    push {LR}
    sub SP, #4
    mov r1,#393216
    ldr r0, 2f
    add r0, PC
    ldr IP, [r0, #4]
1:    blx IP
    ldr IP, [SP, #4]
    add SP, #24
    bx IP
2:    .long .L__wine_spec_relay_descr-1b
    .cfi_endproc
    .align 2
.L__wine_spec_relay_entry_point_2:
    .cfi_startproc
    push {r0-r3}
    mov r2, SP
    vpush {s0-s15}
    push {LR}
    sub SP, #4
        mov r1,#1
    add r1,#65536
    ldr r0, 2f
    add r0, PC
    ldr IP, [r0, #4]
1:    blx IP
    ldr IP, [SP, #4]
    add SP, #88
    bx IP
2:    .long .L__wine_spec_relay_descr-1b
    .cfi_endproc
    .align 2
--- snip ---
https://source.winehq.org/git/wine.git/blob/HEAD:/tools/winebuild/spec32.c#l...
--- snip ---
...
        case CPU_ARM:
        {
            unsigned int mask, val, count = 0;
            int j, has_float = 0;
if (strcmp( float_abi_option, "soft" ))
                for (j = 0; j < odp->u.func.nb_args && !has_float; j++)
                    has_float = is_float_arg( odp, j );
val = (odp->u.func.args_str_offset << 16) | (i - spec->base);
            output( "\tpush {r0-r3}\n" );
            output( "\tmov r2, SP\n");
            if (has_float) output( "\tvpush {s0-s15}\n" );
            output( "\tpush {LR}\n" );
            output( "\tsub SP, #4\n");
            for (mask = 0xff; mask; mask <<= 8)
                if (val & mask) output( "\t%s r1,#%u\n", count++ ? "add" :
"mov", val & mask );
            if (!count) output( "\tmov r1,#0\n" );
            output( "\tldr r0, 2f\n");
            output( "\tadd r0, PC\n");
            output( "\tldr IP, [r0, #4]\n");
            output( "1:\tblx IP\n");
            output( "\tldr IP, [SP, #4]\n" );
            output( "\tadd SP, #%u\n", 24 + (has_float ? 64 : 0) );
            output( "\tbx IP\n");
            output( "2:\t.long .L__wine_spec_relay_descr-1b\n" );
            break; 
...
--- snip ---
The code was introduced by commit
https://source.winehq.org/git/wine.git/commitdiff/5e3534ee411a8c5baea5022705...
("ntdll: Fix handling of floating point arguments in relay debugging on ARM.")
The problem is that Wine makes the assumption that the assembler from hardfloat
binutils implicitly assumes '.fpu {vfp,neon}' without any command line options.
This isn't necessarily true for most hard-float cross-toolchains but for native
toolchains, running on actual targets. Usually this is solved by the
cross-toolchain environment by passing the proper '-mfloat-abi=' and/or
'-mfpu=' options when cross-compiling.
Simple test with .c file:
--- snip ---
$ cat vfptest.c
void foo(void){ __asm__ volatile("fadds s0, s0, s0"); }
--- snip ---
--- snip ---
$ arm-poky-linux-gnueabi-gcc -c vfptest.c
/tmp/cc3DpeRI.s: Assembler messages:
/tmp/cc3DpeRI.s:27: Error: selected processor does not support `fadds s0,s0,s0'
in ARM mode
$ arm-poky-linux-gnueabi-gcc -mfloat-abi=hard -c vfptest.c
--- snip ---
Test to show the available fpu target options:
--- snip ---
$ arm-poky-linux-gnueabi-gcc -mfpu=?
arm-poky-linux-gnueabi-gcc: error: unrecognized argument in option ‘-mfpu=?’
arm-poky-linux-gnueabi-gcc: note: valid arguments to ‘-mfpu=’ are:
 auto crypto-neon-fp-armv8 fp-armv8 fpv4-sp-d16 fpv5-d16 fpv5-sp-d16
 neon neon-fp-armv8 neon-fp16 neon-vfpv3 neon-vfpv4
 vfp vfp3 vfpv2 vfpv3 vfpv3-d16 vfpv3-d16-fp16 vfpv3-fp16 vfpv3xd vfpv3xd-fp16
vfpv4 vfpv4-d16
arm-poky-linux-gnueabi-gcc: fatal error: no input files
compilation terminated.
--- snip ---
Simple test with assembly file, illustrating the problem with Wine:
--- snip ---
$ cat vfptest.S
.text
test:
    fadds s0, s0, s0
--- snip ---
NOTE: 'vfp' is an alias for 'vfpv2', 'neon' is an alias for 'neon-vfpv3'
--- snip ---
$ arm-poky-linux-gnueabi-gcc -c vfptest.s
vfptest.s: Assembler messages:
vfptest.s:3: Error: selected processor does not support `fadds s0,s0,s0' in ARM
mode
$ arm-poky-linux-gnueabi-gcc -c -mfloat-abi=hard vfptest.s
vfptest.s: Assembler messages:
vfptest.s:3: Error: selected processor does not support `fadds s0,s0,s0' in ARM
mode
$ arm-poky-linux-gnueabi-gcc -c -mfpu=vfp vfptest.s
$ arm-poky-linux-gnueabi-gcc -c -mfpu=vfpv2 vfptest.s
$ arm-poky-linux-gnueabi-gcc -c -mfpu=vfpv3 vfptest.s
$ arm-poky-linux-gnueabi-gcc -c -mfpu=vfpv4 vfptest.s
$ arm-poky-linux-gnueabi-gcc -c -mfpu=neon vfptest.s
...
--- snip ---
Unlike using C with inline assembly, real assembly sources need either '.fpu'
directive, or '-mfpu=' option provided in the command line.
Using CFLAGS, EXTRACFLAGS to set '-mfpu=' doesn't work because of the 'winegcc'
wrapper limitation (like bug 46053 it builds a custom command line, not passing
these option down).
Forcing to a use hardfloat binutils with assembler that implicitly assumes
'.fpu vfp' without any command line options is IMHO not feasible. I don't want
to change/hack upstream toolchain projects.
Changing 'winebuild' to add explicit '.fpu vfp' or '.fpu neon' directives to
the assembly sources in case of 32-bit ARM hard-float seems reasonable to me.
https://source.winehq.org/git/wine.git/blob/HEAD:/tools/winebuild/spec32.c
Simple test:
--- snip ---
$ cat vfptest.s
.fpu vfp
.text
test:
    fadds s0, s0, s0
--- snip ---
--- snip ---
$ arm-poky-linux-gnueabi-gcc -c vfptest.s
--- snip ---
With that problem fixed, Wine now cross compiles successfully with a full
32-bit ARM hardfloat target sysroot (same x11/freetype/libxml2 etc. deps as for
my x86_64 host).
Next step would be running apps/tests on my x86_64 host with 'qemuarm64' and
'qemuarma9' using the generic ARM system emulation with the 'virt machine'
machine which also supports 'virtio-gpu' for full graphics stack.
$ wine --version
wine-3.19-117-g4852130c82
Regards
-- 
Do not reply to this email, post in Bugzilla using the
above URL to reply.
You are receiving this mail because:
You are watching all bug changes.

[Bug 46079] New: Support 32-bit ARM hard-float cross-toolchains with GNU binutils assembler that doesn 't implicitly assume '.fpu {vfp,neon}' (Yocto/Poky SDK)