[PATCH] libc: arm: Implement setjmp GCC backwards compatibility.

[PATCH] libc: arm: Implement setjmp GCC backwards compatibility.

[Victor L. Do Nascimento]

When compiling Newlib for arm targets with GCC 12.1 onward, the passing
of architecture extension information to the assembler is automatic,
making the use of `.fpu' directives instructions in assembly files
redundant.

With older versions of GCC, however, the .fpu directive must be
hard-coded into the arm/setjmp.S file to allow the assembly of
instructions concerning the storage and subsequent reloading of the
floating point registers to/from the jump buffer, respectively.

This patch conditionally adds the `.fpu vfpxd' directive based on
compile-time preprocessor macros concerning GCC version and target
architectural features, such that both the assembly and linking of
setjmp.S succeeds for older versions of Newlib.
---
 newlib/libc/machine/arm/setjmp.S | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)

diff --git a/newlib/libc/machine/arm/setjmp.S b/newlib/libc/machine/arm/setjmp.S
index c615f2428..3a9aa840d 100644
--- a/newlib/libc/machine/arm/setjmp.S
+++ b/newlib/libc/machine/arm/setjmp.S
@@ -64,6 +64,22 @@
 
 	.syntax unified
 
+/*  GCC 12.1 and later will tell the assembler exactly which floating
+    point (or MVE) unit is required and we don't want to override
+    that.  Conversely, older versions of the compiler don't pass this
+    information so we need to enable the VFP version that is most
+    appropriate.  The choice here should support all suitable VFP
+    versions that the older toolchains can handle.  */
+#if __GNUC__ && __GNUC__ < 12
+/*  While GCC > 10.1 supports MVE, the MVE instructions do not need an
+    .fpu directive, so we don't need to handle that case.  VFPxd thus
+    covers all the cases we need in this file and should be compatible
+    with all required FPUs that we need to support.  */
+# if __ARM_FP
+	.fpu vfpxd
+# endif
+#endif
+
 #if __ARM_ARCH_ISA_THUMB == 1 && !__ARM_ARCH_ISA_ARM
 /* ARMv6-M-like has to be implemented in Thumb mode.  */
 
-- 
2.36.1

Re: [PATCH] libc: arm: Implement setjmp GCC backwards compatibility.

[Richard Earnshaw]

On 01/02/2023 15:10, Victor L. Do Nascimento wrote:
> When compiling Newlib for arm targets with GCC 12.1 onward, the passing
> of architecture extension information to the assembler is automatic,
> making the use of `.fpu' directives instructions in assembly files
> redundant.
> 
> With older versions of GCC, however, the .fpu directive must be
> hard-coded into the arm/setjmp.S file to allow the assembly of
> instructions concerning the storage and subsequent reloading of the
> floating point registers to/from the jump buffer, respectively.
> 
> This patch conditionally adds the `.fpu vfpxd' directive based on
> compile-time preprocessor macros concerning GCC version and target
> architectural features, such that both the assembly and linking of
> setjmp.S succeeds for older versions of Newlib.
> ---
>   newlib/libc/machine/arm/setjmp.S | 16 ++++++++++++++++
>   1 file changed, 16 insertions(+)
> 
> diff --git a/newlib/libc/machine/arm/setjmp.S b/newlib/libc/machine/arm/setjmp.S
> index c615f2428..3a9aa840d 100644
> --- a/newlib/libc/machine/arm/setjmp.S
> +++ b/newlib/libc/machine/arm/setjmp.S
> @@ -64,6 +64,22 @@
>   
>   	.syntax unified
>   
> +/*  GCC 12.1 and later will tell the assembler exactly which floating
> +    point (or MVE) unit is required and we don't want to override
> +    that.  Conversely, older versions of the compiler don't pass this
> +    information so we need to enable the VFP version that is most
> +    appropriate.  The choice here should support all suitable VFP
> +    versions that the older toolchains can handle.  */
> +#if __GNUC__ && __GNUC__ < 12
> +/*  While GCC > 10.1 supports MVE, the MVE instructions do not need an
> +    .fpu directive, so we don't need to handle that case.  VFPxd thus
> +    covers all the cases we need in this file and should be compatible
> +    with all required FPUs that we need to support.  */
> +# if __ARM_FP
> +	.fpu vfpxd
> +# endif
> +#endif
> +
>   #if __ARM_ARCH_ISA_THUMB == 1 && !__ARM_ARCH_ISA_ARM
>   /* ARMv6-M-like has to be implemented in Thumb mode.  */
>   

Since MVE is available in GCC-10 onwards, I don't think that's quite 
enough.  Compiling with -march=armv8.1-m.main+mve still needs the 
extension register set preserving during setjmp, but this will not cause 
__ARM_FP to be defined.

So I think we need something like

# if __ARM_FP
	.fpu vfpxd
# endif
# if __ARM_FEATURE_MVE
	.arch_extension mve
# endif

This should ensure that not only the files assemble correctly but they 
have suitable build attributes as well (otherwise we might get problems 
at link time if integer-only MVE code is linked with code claiming that 
full VFP is needed).

R.