vfp: force non-conditional encoding for external Thumb2

Nick reports that the following error is produced in some cases when
using GCC+ld.bfd to build the ARM defconfig with Thumb2 enabled:

arch/arm/vfp/vfphw.o: in function `vfp_support_entry':
(.text+0xa): relocation truncated to fit: R_ARM_THM_JUMP19 against
symbol `vfp_kmode_exception' defined in .text.unlikely section in
arch/arm/vfp/vfpmodule.o

$ arm-linux-gnueabihf-ld --version
GNU ld (GNU Binutils for Debian) 2.34

Generally, the linker should be able to fix up out of range branches by
emitting veneers, but apparently, it fails to do so in this particular
case, i.e., a conditional 'tail call' to vfp_kmode_exception(), which
is not defined in the same object file.

So let's force the use of a non-conditional encoding of the B instruction,
which has more space for an immediate offset. To compensate for the
additional 2 byte IT opcode, switch the preceding TEQ to CMP, which can
be emitted in 2 bytes instead of 4 bytes as well.

Fixes: eff8728fe698 ("vmlinux.lds.h: Add PGO and AutoFDO input sections")
Reported-by: Nick Desaulniers <(address hidden)>
Tested-by: Nick Desaulniers <(address hidden)>
Signed-off-by: Ard Biesheuvel <(address hidden)>
---

diff --git a/arch/arm/vfp/vfphw.S b/arch/arm/vfp/vfphw.S
index 4fcff9f59947..06ff091c0932 100644
--- a/arch/arm/vfp/vfphw.S
+++ b/arch/arm/vfp/vfphw.S
@@ -81,7 +81,8 @@ ENTRY(vfp_support_entry)
 	.fpu	vfpv2
 	ldr	r3, [sp, #S_PSR]	@ Neither lazy restore nor FP exceptions
 	and	r3, r3, #MODE_MASK	@ are supported in kernel mode
-	teq	r3, #USR_MODE
+	cmp	r3, #USR_MODE
+THUMB(	it	ne			)
 	bne	vfp_kmode_exception	@ Returns through lr
 
 	VFPFMRX	r1, FPEXC		@ Is the VFP enabled?