[PATCH] Workaround ARM VFP issue

[PATCH] Workaround ARM VFP issue

[Chung-Lin Tang]

[-- Attachment #1: Type: text/plain, Size: 1255 bytes --]

This is patch to workaround the issue mentioned in this thread:
http://sourceware.org/ml/libffi-discuss/2012/msg00038.html

The issue is not really that soft-float is broken, but rather the
toolchain used is simply too old, not really because it is old-ABI, but
too old a version of binutils; the .fpu directive, nor VFP instructions
are supported in the assembler, hence even building of the VFP
call/closure shims fail.

This patch is sort of a least attempt to shield away the VFP stuff when
__ARM_EABI__ is not defined. I apologize for such a token workaround, as
it can probably be optimized a bit more, but I just feel it's not
worthwhile; even GCC (recent versions) assumes such assembler features.

Kyle, can you please see if the patch can work for you? If in case it
doesn't, I might also suggest that you try to rebuild your toolchain
with a newer binutils version (at least post circa 2005), which should
at least not fail to build the library.

Thanks,
Chung-Lin

2012-03-06  Chung-Lin Tang  <cltang@codesourcery.com>

	* src/arm/ffi.c (ffi_call): Add __ARM_EABI__ guard around call
	to ffi_call_VFP().
	(ffi_prep_closure_loc): Add __ARM_EABI__ guard around use of
	ffi_closure_VFP.
	* src/arm/sysv.S: Add __ARM_EABI__ guard around VFP code.

[-- Attachment #2: ffi.diff --]
[-- Type: text/plain, Size: 1311 bytes --]

diff --git a/src/arm/ffi.c b/src/arm/ffi.c
index 1f8597d..3ccceb9 100644
--- a/src/arm/ffi.c
+++ b/src/arm/ffi.c
@@ -251,8 +251,10 @@ void ffi_call(ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue)
       break;
 
     case FFI_VFP:
+#ifdef __ARM_EABI__
       ffi_call_VFP (fn, &ecif, cif->bytes, cif->flags, ecif.rvalue);
       break;
+#endif
 
     default:
       FFI_ASSERT(0);
@@ -609,8 +611,10 @@ ffi_prep_closure_loc (ffi_closure* closure,
 
   if (cif->abi == FFI_SYSV)
     closure_func = &ffi_closure_SYSV;
+#ifdef __ARM_EABI__
   else if (cif->abi == FFI_VFP)
     closure_func = &ffi_closure_VFP;
+#endif
   else
     return FFI_BAD_ABI;
     
diff --git a/src/arm/sysv.S b/src/arm/sysv.S
index 60e2ae3..2505305 100644
--- a/src/arm/sysv.S
+++ b/src/arm/sysv.S
@@ -334,7 +334,8 @@ ARM_FUNC_START ffi_closure_SYSV
 
 
 /* Below are VFP hard-float ABI call and closure implementations.
-   Add VFP FPU directive here. */
+   Add VFP FPU directive here. This is only compiled into the library under EABI.  */
+#ifdef __ARM_EABI__
 	.fpu	vfp
 
 	@ r0:   fn
@@ -491,6 +492,7 @@ ARM_FUNC_START ffi_closure_VFP
 .ffi_closure_VFP_end:
 	UNWIND .fnend
         .size    CNAME(ffi_closure_VFP),.ffi_closure_VFP_end-CNAME(ffi_closure_VFP)
+#endif
 
 ENTRY(ffi_arm_trampoline)
 	stmfd sp!, {r0-r3}

Re: [PATCH] Workaround ARM VFP issue

[Kyle Tinker]

Chung-Lin,

First of all, THANK you for even looking at this.  I know the toolchain 
is pretty^H^H^H^H^H^H *really* old.

I will definitely try to compile this and let you know the results. 
Unfortunately, I don't have a lot of control over the toolchain, as it's 
not mine--I'm using one provided by a vendor.  I'll see what I can do to 
update or get binutils updated.  I know that I've hit a number of 
compiler issues (in both this project and others) due to the toolchain 
being ancient. :(

Give me a few hours/days and I'll get the results back to you.

Thanks,
*Kyle Tinker* *|* *Software Developer* *|* *WorkForce Software*
*|* P: (734) 742-2616 *|* F: (734) 542-0635
*|* ktinker@workforcesoftware.com <mailto:ktinker@workforcesoftware.com>

WorkForce Software — Mitigating Complexity — Ensuring Compliance
— Enabling Strategic HR

WorkForce Software


From: Chung-Lin Tang <cltang@codesourcery.com>
Sent: Tue, Mar 06, 2012 7:37 AM
To: libffi-discuss@sourceware.org <libffi-discuss@sourceware.org>
CC: "green@moxielogic.com" <green@moxielogic.com>, Kyle Tinker 
<ktinker@workforcesoftware.com>
Subject: [PATCH] Workaround ARM VFP issue


> This is patch to workaround the issue mentioned in this thread:
> http://sourceware.org/ml/libffi-discuss/2012/msg00038.html
>
> The issue is not really that soft-float is broken, but rather the
> toolchain used is simply too old, not really because it is old-ABI, but
> too old a version of binutils; the .fpu directive, nor VFP instructions
> are supported in the assembler, hence even building of the VFP
> call/closure shims fail.
>
> This patch is sort of a least attempt to shield away the VFP stuff when
> __ARM_EABI__ is not defined. I apologize for such a token workaround, as
> it can probably be optimized a bit more, but I just feel it's not
> worthwhile; even GCC (recent versions) assumes such assembler features.
>
> Kyle, can you please see if the patch can work for you? If in case it
> doesn't, I might also suggest that you try to rebuild your toolchain
> with a newer binutils version (at least post circa 2005), which should
> at least not fail to build the library.
>
> Thanks,
> Chung-Lin
>
> 2012-03-06  Chung-Lin Tang<cltang@codesourcery.com>
>
> 	* src/arm/ffi.c (ffi_call): Add __ARM_EABI__ guard around call
> 	to ffi_call_VFP().
> 	(ffi_prep_closure_loc): Add __ARM_EABI__ guard around use of
> 	ffi_closure_VFP.
> 	* src/arm/sysv.S: Add __ARM_EABI__ guard around VFP code.

Re: [PATCH] Workaround ARM VFP issue

[Kyle Tinker]

Chung-Lin:

First, sorry for the large (3 month) delay.

I've applied the patch and gotten it to compile on our system!  THANK YOU!

We had to define the __SOFTFP__ cflag to get it to compile correctly, 
but the patch worked great once we did that.

Sorry our toolchain is so old...  performance is always a consideration, 
but we shouldn't be making THAT many libffi calls for the changes you 
made to matter.  I certainly wouldn't optimize your code for the sake of 
a toolchain that predates 2005.

Thanks,
*Kyle Tinker* *|* *Software Development Manager* *|* *WorkForce Software*

WorkForce Software — Mitigating Complexity — Ensuring Compliance
— Enabling Strategic HR


From: Chung-Lin Tang <cltang@codesourcery.com>
Sent: Tue, Mar 06, 2012 7:37 AM
To: libffi-discuss@sourceware.org <libffi-discuss@sourceware.org>
CC: "green@moxielogic.com" <green@moxielogic.com>, Kyle Tinker 
<ktinker@workforcesoftware.com>
Subject: [PATCH] Workaround ARM VFP issue


> This is patch to workaround the issue mentioned in this thread:
> http://sourceware.org/ml/libffi-discuss/2012/msg00038.html
>
> The issue is not really that soft-float is broken, but rather the
> toolchain used is simply too old, not really because it is old-ABI, but
> too old a version of binutils; the .fpu directive, nor VFP instructions
> are supported in the assembler, hence even building of the VFP
> call/closure shims fail.
>
> This patch is sort of a least attempt to shield away the VFP stuff when
> __ARM_EABI__ is not defined. I apologize for such a token workaround, as
> it can probably be optimized a bit more, but I just feel it's not
> worthwhile; even GCC (recent versions) assumes such assembler features.
>
> Kyle, can you please see if the patch can work for you? If in case it
> doesn't, I might also suggest that you try to rebuild your toolchain
> with a newer binutils version (at least post circa 2005), which should
> at least not fail to build the library.
>
> Thanks,
> Chung-Lin
>
> 2012-03-06  Chung-Lin Tang<cltang@codesourcery.com>
>
> 	* src/arm/ffi.c (ffi_call): Add __ARM_EABI__ guard around call
> 	to ffi_call_VFP().
> 	(ffi_prep_closure_loc): Add __ARM_EABI__ guard around use of
> 	ffi_closure_VFP.
> 	* src/arm/sysv.S: Add __ARM_EABI__ guard around VFP code.

[PATCH] Workaround ARM VFP issue

[Chung-Lin Tang]

[-- Attachment #1: Type: text/plain, Size: 1255 bytes --]

This is patch to workaround the issue mentioned in this thread:
http://sourceware.org/ml/libffi-discuss/2012/msg00038.html

The issue is not really that soft-float is broken, but rather the
toolchain used is simply too old, not really because it is old-ABI, but
too old a version of binutils; the .fpu directive, nor VFP instructions
are supported in the assembler, hence even building of the VFP
call/closure shims fail.

This patch is sort of a least attempt to shield away the VFP stuff when
__ARM_EABI__ is not defined. I apologize for such a token workaround, as
it can probably be optimized a bit more, but I just feel it's not
worthwhile; even GCC (recent versions) assumes such assembler features.

Kyle, can you please see if the patch can work for you? If in case it
doesn't, I might also suggest that you try to rebuild your toolchain
with a newer binutils version (at least post circa 2005), which should
at least not fail to build the library.

Thanks,
Chung-Lin

2012-03-06  Chung-Lin Tang  <cltang@codesourcery.com>

	* src/arm/ffi.c (ffi_call): Add __ARM_EABI__ guard around call
	to ffi_call_VFP().
	(ffi_prep_closure_loc): Add __ARM_EABI__ guard around use of
	ffi_closure_VFP.
	* src/arm/sysv.S: Add __ARM_EABI__ guard around VFP code.

[-- Attachment #2.1: Type: application/applefile, Size: 430 bytes --]

[-- Attachment #2.2: ffi.diff --]
[-- Type: application/octet-stream, Size: 1311 bytes --]

diff --git a/src/arm/ffi.c b/src/arm/ffi.c
index 1f8597d..3ccceb9 100644
--- a/src/arm/ffi.c
+++ b/src/arm/ffi.c
@@ -251,8 +251,10 @@ void ffi_call(ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue)
       break;
 
     case FFI_VFP:
+#ifdef __ARM_EABI__
       ffi_call_VFP (fn, &ecif, cif->bytes, cif->flags, ecif.rvalue);
       break;
+#endif
 
     default:
       FFI_ASSERT(0);
@@ -609,8 +611,10 @@ ffi_prep_closure_loc (ffi_closure* closure,
 
   if (cif->abi == FFI_SYSV)
     closure_func = &ffi_closure_SYSV;
+#ifdef __ARM_EABI__
   else if (cif->abi == FFI_VFP)
     closure_func = &ffi_closure_VFP;
+#endif
   else
     return FFI_BAD_ABI;
     
diff --git a/src/arm/sysv.S b/src/arm/sysv.S
index 60e2ae3..2505305 100644
--- a/src/arm/sysv.S
+++ b/src/arm/sysv.S
@@ -334,7 +334,8 @@ ARM_FUNC_START ffi_closure_SYSV
 
 
 /* Below are VFP hard-float ABI call and closure implementations.
-   Add VFP FPU directive here. */
+   Add VFP FPU directive here. This is only compiled into the library under EABI.  */
+#ifdef __ARM_EABI__
 	.fpu	vfp
 
 	@ r0:   fn
@@ -491,6 +492,7 @@ ARM_FUNC_START ffi_closure_VFP
 .ffi_closure_VFP_end:
 	UNWIND .fnend
         .size    CNAME(ffi_closure_VFP),.ffi_closure_VFP_end-CNAME(ffi_closure_VFP)
+#endif
 
 ENTRY(ffi_arm_trampoline)
 	stmfd sp!, {r0-r3}

[PATCH] Workaround ARM VFP issue

[Chung-Lin Tang]

[-- Attachment #1: Type: text/plain, Size: 1255 bytes --]

This is patch to workaround the issue mentioned in this thread:
http://sourceware.org/ml/libffi-discuss/2012/msg00038.html

The issue is not really that soft-float is broken, but rather the
toolchain used is simply too old, not really because it is old-ABI, but
too old a version of binutils; the .fpu directive, nor VFP instructions
are supported in the assembler, hence even building of the VFP
call/closure shims fail.

This patch is sort of a least attempt to shield away the VFP stuff when
__ARM_EABI__ is not defined. I apologize for such a token workaround, as
it can probably be optimized a bit more, but I just feel it's not
worthwhile; even GCC (recent versions) assumes such assembler features.

Kyle, can you please see if the patch can work for you? If in case it
doesn't, I might also suggest that you try to rebuild your toolchain
with a newer binutils version (at least post circa 2005), which should
at least not fail to build the library.

Thanks,
Chung-Lin

2012-03-06  Chung-Lin Tang  <cltang@codesourcery.com>

	* src/arm/ffi.c (ffi_call): Add __ARM_EABI__ guard around call
	to ffi_call_VFP().
	(ffi_prep_closure_loc): Add __ARM_EABI__ guard around use of
	ffi_closure_VFP.
	* src/arm/sysv.S: Add __ARM_EABI__ guard around VFP code.

[-- Attachment #2.1: Type: application/applefile, Size: 430 bytes --]

[-- Attachment #2.2: ffi.diff --]
[-- Type: application/octet-stream, Size: 1311 bytes --]

diff --git a/src/arm/ffi.c b/src/arm/ffi.c
index 1f8597d..3ccceb9 100644
--- a/src/arm/ffi.c
+++ b/src/arm/ffi.c
@@ -251,8 +251,10 @@ void ffi_call(ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue)
       break;
 
     case FFI_VFP:
+#ifdef __ARM_EABI__
       ffi_call_VFP (fn, &ecif, cif->bytes, cif->flags, ecif.rvalue);
       break;
+#endif
 
     default:
       FFI_ASSERT(0);
@@ -609,8 +611,10 @@ ffi_prep_closure_loc (ffi_closure* closure,
 
   if (cif->abi == FFI_SYSV)
     closure_func = &ffi_closure_SYSV;
+#ifdef __ARM_EABI__
   else if (cif->abi == FFI_VFP)
     closure_func = &ffi_closure_VFP;
+#endif
   else
     return FFI_BAD_ABI;
     
diff --git a/src/arm/sysv.S b/src/arm/sysv.S
index 60e2ae3..2505305 100644
--- a/src/arm/sysv.S
+++ b/src/arm/sysv.S
@@ -334,7 +334,8 @@ ARM_FUNC_START ffi_closure_SYSV
 
 
 /* Below are VFP hard-float ABI call and closure implementations.
-   Add VFP FPU directive here. */
+   Add VFP FPU directive here. This is only compiled into the library under EABI.  */
+#ifdef __ARM_EABI__
 	.fpu	vfp
 
 	@ r0:   fn
@@ -491,6 +492,7 @@ ARM_FUNC_START ffi_closure_VFP
 .ffi_closure_VFP_end:
 	UNWIND .fnend
         .size    CNAME(ffi_closure_VFP),.ffi_closure_VFP_end-CNAME(ffi_closure_VFP)
+#endif
 
 ENTRY(ffi_arm_trampoline)
 	stmfd sp!, {r0-r3}

[PATCH] Workaround ARM VFP issue

[Chung-Lin Tang]

[-- Attachment #1: Type: text/plain, Size: 1255 bytes --]

This is patch to workaround the issue mentioned in this thread:
http://sourceware.org/ml/libffi-discuss/2012/msg00038.html

The issue is not really that soft-float is broken, but rather the
toolchain used is simply too old, not really because it is old-ABI, but
too old a version of binutils; the .fpu directive, nor VFP instructions
are supported in the assembler, hence even building of the VFP
call/closure shims fail.

This patch is sort of a least attempt to shield away the VFP stuff when
__ARM_EABI__ is not defined. I apologize for such a token workaround, as
it can probably be optimized a bit more, but I just feel it's not
worthwhile; even GCC (recent versions) assumes such assembler features.

Kyle, can you please see if the patch can work for you? If in case it
doesn't, I might also suggest that you try to rebuild your toolchain
with a newer binutils version (at least post circa 2005), which should
at least not fail to build the library.

Thanks,
Chung-Lin

2012-03-06  Chung-Lin Tang  <cltang@codesourcery.com>

	* src/arm/ffi.c (ffi_call): Add __ARM_EABI__ guard around call
	to ffi_call_VFP().
	(ffi_prep_closure_loc): Add __ARM_EABI__ guard around use of
	ffi_closure_VFP.
	* src/arm/sysv.S: Add __ARM_EABI__ guard around VFP code.

[-- Attachment #2.1: Type: application/applefile, Size: 430 bytes --]

[-- Attachment #2.2: ffi.diff --]
[-- Type: application/octet-stream, Size: 1311 bytes --]

diff --git a/src/arm/ffi.c b/src/arm/ffi.c
index 1f8597d..3ccceb9 100644
--- a/src/arm/ffi.c
+++ b/src/arm/ffi.c
@@ -251,8 +251,10 @@ void ffi_call(ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue)
       break;
 
     case FFI_VFP:
+#ifdef __ARM_EABI__
       ffi_call_VFP (fn, &ecif, cif->bytes, cif->flags, ecif.rvalue);
       break;
+#endif
 
     default:
       FFI_ASSERT(0);
@@ -609,8 +611,10 @@ ffi_prep_closure_loc (ffi_closure* closure,
 
   if (cif->abi == FFI_SYSV)
     closure_func = &ffi_closure_SYSV;
+#ifdef __ARM_EABI__
   else if (cif->abi == FFI_VFP)
     closure_func = &ffi_closure_VFP;
+#endif
   else
     return FFI_BAD_ABI;
     
diff --git a/src/arm/sysv.S b/src/arm/sysv.S
index 60e2ae3..2505305 100644
--- a/src/arm/sysv.S
+++ b/src/arm/sysv.S
@@ -334,7 +334,8 @@ ARM_FUNC_START ffi_closure_SYSV
 
 
 /* Below are VFP hard-float ABI call and closure implementations.
-   Add VFP FPU directive here. */
+   Add VFP FPU directive here. This is only compiled into the library under EABI.  */
+#ifdef __ARM_EABI__
 	.fpu	vfp
 
 	@ r0:   fn
@@ -491,6 +492,7 @@ ARM_FUNC_START ffi_closure_VFP
 .ffi_closure_VFP_end:
 	UNWIND .fnend
         .size    CNAME(ffi_closure_VFP),.ffi_closure_VFP_end-CNAME(ffi_closure_VFP)
+#endif
 
 ENTRY(ffi_arm_trampoline)
 	stmfd sp!, {r0-r3}