| Message ID | 20220810090446.1469742-1-saproj@gmail.com |
|---|---|
| State | Changes Requested |
| Delegated to: | Tom Rini |
| Headers | show |
| Series | arm: ARMv4 assembly compatibility | expand |
On Wed, 10 Aug 2022 at 12:05, Sergei Antonov <saproj@gmail.com> wrote: > > There is currently a problem that U-Boot can not work on ARMv4 > because assembly imlementations of memcpy() and some other functions > use "bx lr" instruction that is not available on ARMv4 ("mov pc, lr" > should be used instead). Hello! There has been no replies to this message. I hope it did not make it into spam folders. The proposed patch allows U-Boot to run on ARMv4 systems without breaking (or even changing) other systems. Please, consider it.
On Wed, 10 Aug 2022 12:04:46 +0300 Sergei Antonov <saproj@gmail.com> wrote: Hi, > There is currently a problem that U-Boot can not work on ARMv4 > because assembly imlementations of memcpy() and some other functions > use "bx lr" instruction that is not available on ARMv4 ("mov pc, lr" > should be used instead). So there is a comment in the current code, right above what's shown here in the first hunk, which says: /* * We only support cores that support at least Thumb-1 and thus we use * 'bx lr' */ This was added by Tom in commit 431afb4ef9fe, when he removed some code very close to what you are adding back now. So what is the story here? This commit seems to suggest U-Boot doesn't support even ARMv5 without "T", has this changed? There are probably other code places which would need adjustment to run on ARMv4? Tom, can you say why support for Thumb-less code / older architectures was dropped? No users, I guess? Sergei, can you say what is this for? Are you adding support for a SoC with an ARM810 core from 1996? Cheers, Andre > > A working preprocessor-based solution to this problem is found in > arch/arm/lib/relocate.S. Move it to the "ret" macro in > arch/arm/include/asm/assembler.h and change all "bx lr" code > to "ret lr" in functions that may run on ARMv4. Linux source code > deals with this problem in the same manner. > > Signed-off-by: Sergei Antonov <saproj@gmail.com> > CC: Samuel Holland <samuel@sholland.org> > CC: Ye Li <ye.li@nxp.com> > CC: Simon Glass <sjg@chromium.org> > CC: Andre Przywara <andre.przywara@arm.com> > CC: Marek Vasut <marex@denx.de> > CC: Sean Anderson <sean.anderson@seco.com> > CC: Tom Rini <trini@konsulko.com> > CC: Wolfgang Denk <wd@denx.de> > --- > arch/arm/include/asm/assembler.h | 6 ++++++ > arch/arm/lib/lib1funcs.S | 8 ++++---- > arch/arm/lib/memcpy.S | 6 +++--- > arch/arm/lib/relocate.S | 10 ++-------- > arch/arm/lib/setjmp.S | 4 ++-- > 5 files changed, 17 insertions(+), 17 deletions(-) > > diff --git a/arch/arm/include/asm/assembler.h b/arch/arm/include/asm/assembler.h > index b146918586..911f7a1b49 100644 > --- a/arch/arm/include/asm/assembler.h > +++ b/arch/arm/include/asm/assembler.h > @@ -63,11 +63,17 @@ > */ > .irp c,,eq,ne,cs,cc,mi,pl,vs,vc,hi,ls,ge,lt,gt,le,hs,lo > .macro ret\c, reg > + > + /* ARMv4- don't know bx lr but the assembler fails to see that */ > +#ifdef __ARM_ARCH_4__ > + mov\c pc, \reg > +#else > .ifeqs "\reg", "lr" > bx\c \reg > .else > mov\c pc, \reg > .endif > +#endif > .endm > .endr > > diff --git a/arch/arm/lib/lib1funcs.S b/arch/arm/lib/lib1funcs.S > index 700eee5fbb..7ff4446dd6 100644 > --- a/arch/arm/lib/lib1funcs.S > +++ b/arch/arm/lib/lib1funcs.S > @@ -377,7 +377,7 @@ ENTRY(__gnu_thumb1_case_sqi) > lsls r1, r1, #1 > add lr, lr, r1 > pop {r1} > - bx lr > + ret lr > ENDPROC(__gnu_thumb1_case_sqi) > .popsection > > @@ -391,7 +391,7 @@ ENTRY(__gnu_thumb1_case_uqi) > lsls r1, r1, #1 > add lr, lr, r1 > pop {r1} > - bx lr > + ret lr > ENDPROC(__gnu_thumb1_case_uqi) > .popsection > > @@ -406,7 +406,7 @@ ENTRY(__gnu_thumb1_case_shi) > lsls r1, r1, #1 > add lr, lr, r1 > pop {r0, r1} > - bx lr > + ret lr > ENDPROC(__gnu_thumb1_case_shi) > .popsection > > @@ -421,7 +421,7 @@ ENTRY(__gnu_thumb1_case_uhi) > lsls r1, r1, #1 > add lr, lr, r1 > pop {r0, r1} > - bx lr > + ret lr > ENDPROC(__gnu_thumb1_case_uhi) > .popsection > #endif > diff --git a/arch/arm/lib/memcpy.S b/arch/arm/lib/memcpy.S > index eee7a219ce..a1c996f94e 100644 > --- a/arch/arm/lib/memcpy.S > +++ b/arch/arm/lib/memcpy.S > @@ -59,7 +59,7 @@ > #endif > ENTRY(memcpy) > cmp r0, r1 > - bxeq lr > + reteq lr > > enter r4, lr > > @@ -148,7 +148,7 @@ ENTRY(memcpy) > str1b r0, ip, cs, abort=21f > > exit r4, lr > - bx lr > + ret lr > > 9: rsb ip, ip, #4 > cmp ip, #2 > @@ -258,7 +258,7 @@ ENTRY(memcpy) > > .macro copy_abort_end > ldmfd sp!, {r4, lr} > - bx lr > + ret lr > .endm > > ENDPROC(memcpy) > diff --git a/arch/arm/lib/relocate.S b/arch/arm/lib/relocate.S > index 5102bfabde..dd6f2e3bd5 100644 > --- a/arch/arm/lib/relocate.S > +++ b/arch/arm/lib/relocate.S > @@ -61,7 +61,7 @@ ENTRY(relocate_vectors) > stmia r1!, {r2-r8,r10} > #endif > #endif > - bx lr > + ret lr > > ENDPROC(relocate_vectors) > > @@ -127,13 +127,7 @@ relocate_done: > mcr p15, 0, r0, c7, c10, 4 /* drain write buffer */ > #endif > > - /* ARMv4- don't know bx lr but the assembler fails to see that */ > - > -#ifdef __ARM_ARCH_4__ > - mov pc, lr > -#else > - bx lr > -#endif > + ret lr > > ENDPROC(relocate_code) > > diff --git a/arch/arm/lib/setjmp.S b/arch/arm/lib/setjmp.S > index 176a1d5315..2f041aeef0 100644 > --- a/arch/arm/lib/setjmp.S > +++ b/arch/arm/lib/setjmp.S > @@ -17,7 +17,7 @@ ENTRY(setjmp) > mov ip, sp > stm a1, {v1-v8, ip, lr} > mov a1, #0 > - bx lr > + ret lr > ENDPROC(setjmp) > .popsection > > @@ -31,6 +31,6 @@ ENTRY(longjmp) > bne 1f > mov a1, #1 > 1: > - bx lr > + ret lr > ENDPROC(longjmp) > .popsection
On Tue, Aug 16, 2022 at 12:17 PM Andre Przywara <andre.przywara@arm.com> wrote: > > So what is the story here? This commit seems to suggest U-Boot doesn't support > even ARMv5 without "T", has this changed? There are probably other code > places which would need adjustment to run on ARMv4? Note that gcc 6.0 and later considers armv4 (non-Thumb) support to be deprecated. [1] [2] [1] https://gcc.gnu.org/onlinedocs/gcc/ARM-Options.html [2] https://lkml.iu.edu/hypermail/linux/kernel/1712.2/05158.html On Wed, 10 Aug 2022 12:04:46 +0300 Sergei Antonov <saproj@gmail.com> wrote: > > A working preprocessor-based solution to this problem is found in > arch/arm/lib/relocate.S. Move it to the "ret" macro in > arch/arm/include/asm/assembler.h and change all "bx lr" code > to "ret lr" in functions that may run on ARMv4. Linux source code > deals with this problem in the same manner. Just wanted to point out another option: the linker has a --fix-v4bx flat, which performs the same conversion of "bx lr" to "mov pc lr". Although it is kind of an unusual place for such a transformation, it has the advantage that one does not need to fixup the source code of every piece of code you might like to compile. I used this some years ago [3], and it seems it still works in 2020 [4] [3] http://lists.infradead.org/pipermail/netwinder/2017-August/000267.html [4] https://lists.gnu.org/archive/html/bug-binutils/2020-02/msg00174.html Ralph
On Tue, 16 Aug 2022 at 19:17, Andre Przywara <andre.przywara@arm.com> wrote: > > On Wed, 10 Aug 2022 12:04:46 +0300 > Sergei Antonov <saproj@gmail.com> wrote: > > Hi, > > > There is currently a problem that U-Boot can not work on ARMv4 > > because assembly imlementations of memcpy() and some other functions > > use "bx lr" instruction that is not available on ARMv4 ("mov pc, lr" > > should be used instead). > > So there is a comment in the current code, right above what's shown here in > the first hunk, which says: > /* > * We only support cores that support at least Thumb-1 and thus we use > * 'bx lr' > */ I did not notice it. Thanks. > This was added by Tom in commit 431afb4ef9fe, when he removed some code > very close to what you are adding back now. > > So what is the story here? This commit seems to suggest U-Boot doesn't support > even ARMv5 without "T", has this changed? There are probably other code > places which would need adjustment to run on ARMv4? I can not explain why there was __ARM_ARCH_5E__ in the removed code. I am dealing with the issue that ARMv4 ("four" without letter "t") does not understand "bx lr" while ARMv4T does. The device on which I run modern U-Boot has FA526 - a CPU similar to ARM810 which is ARMv4 without Thumb. Its MIDR register is 0x66015261 which means that Architecture=1. > Tom, can you say why support for Thumb-less code / older architectures was > dropped? No users, I guess? > > Sergei, can you say what is this for? Are you adding support for a SoC > with an ARM810 core from 1996? Yes. And U-Boot for this device is already working: able to boot Linux and do TFTP stuff. My company is going to use it on hundreds of devices. It would be great for us to not have to patch U-Boot locally and have ARMv4-compatible assembly in the mainline.
On Wed, Aug 17, 2022 at 11:27:39AM +0300, Sergei Antonov wrote: > On Tue, 16 Aug 2022 at 19:17, Andre Przywara <andre.przywara@arm.com> wrote: > > > > On Wed, 10 Aug 2022 12:04:46 +0300 > > Sergei Antonov <saproj@gmail.com> wrote: > > > > Hi, > > > > > There is currently a problem that U-Boot can not work on ARMv4 > > > because assembly imlementations of memcpy() and some other functions > > > use "bx lr" instruction that is not available on ARMv4 ("mov pc, lr" > > > should be used instead). > > > > So there is a comment in the current code, right above what's shown here in > > the first hunk, which says: > > /* > > * We only support cores that support at least Thumb-1 and thus we use > > * 'bx lr' > > */ > > I did not notice it. Thanks. > > > This was added by Tom in commit 431afb4ef9fe, when he removed some code > > very close to what you are adding back now. > > > > So what is the story here? This commit seems to suggest U-Boot doesn't support > > even ARMv5 without "T", has this changed? There are probably other code > > places which would need adjustment to run on ARMv4? > > I can not explain why there was __ARM_ARCH_5E__ in the removed code. I > am dealing with the issue that ARMv4 ("four" without letter "t") does > not understand "bx lr" while ARMv4T does. The device on which I run > modern U-Boot has FA526 - a CPU similar to ARM810 which is ARMv4 > without Thumb. Its MIDR register is 0x66015261 which means that > Architecture=1. > > > Tom, can you say why support for Thumb-less code / older architectures was > > dropped? No users, I guess? > > > > Sergei, can you say what is this for? Are you adding support for a SoC > > with an ARM810 core from 1996? > > Yes. > And U-Boot for this device is already working: able to boot Linux and > do TFTP stuff. My company is going to use it on hundreds of devices. > It would be great for us to not have to patch U-Boot locally and have > ARMv4-compatible assembly in the mainline. So, my recollection at the time was that we were / had removed all upstream platforms that were using a core design that didn't accept "bx lr". If you're going to upstream the rest of your platform, please do a v2 that also updates the comment Andre pointed out. But by itself, the general policy is that we support platforms that are upstream, and anything else is functionally dead code.
On Sat, 20 Aug 2022 at 22:13, Tom Rini <trini@konsulko.com> wrote: > So, my recollection at the time was that we were / had removed all > upstream platforms that were using a core design that didn't accept "bx > lr". If you're going to upstream the rest of your platform, please do a > v2 that also updates the comment Andre pointed out. But by itself, the > general policy is that we support platforms that are upstream, and > anything else is functionally dead code. Yes, I am going to upstream the rest of my platform. It will involve resurrecting some removed non-DM drivers and porting them to DM. Submitting now the v2 of the patch with an updated commend.
diff --git a/arch/arm/include/asm/assembler.h b/arch/arm/include/asm/assembler.h index b146918586..911f7a1b49 100644 --- a/arch/arm/include/asm/assembler.h +++ b/arch/arm/include/asm/assembler.h @@ -63,11 +63,17 @@ */ .irp c,,eq,ne,cs,cc,mi,pl,vs,vc,hi,ls,ge,lt,gt,le,hs,lo .macro ret\c, reg + + /* ARMv4- don't know bx lr but the assembler fails to see that */ +#ifdef __ARM_ARCH_4__ + mov\c pc, \reg +#else .ifeqs "\reg", "lr" bx\c \reg .else mov\c pc, \reg .endif +#endif .endm .endr diff --git a/arch/arm/lib/lib1funcs.S b/arch/arm/lib/lib1funcs.S index 700eee5fbb..7ff4446dd6 100644 --- a/arch/arm/lib/lib1funcs.S +++ b/arch/arm/lib/lib1funcs.S @@ -377,7 +377,7 @@ ENTRY(__gnu_thumb1_case_sqi) lsls r1, r1, #1 add lr, lr, r1 pop {r1} - bx lr + ret lr ENDPROC(__gnu_thumb1_case_sqi) .popsection @@ -391,7 +391,7 @@ ENTRY(__gnu_thumb1_case_uqi) lsls r1, r1, #1 add lr, lr, r1 pop {r1} - bx lr + ret lr ENDPROC(__gnu_thumb1_case_uqi) .popsection @@ -406,7 +406,7 @@ ENTRY(__gnu_thumb1_case_shi) lsls r1, r1, #1 add lr, lr, r1 pop {r0, r1} - bx lr + ret lr ENDPROC(__gnu_thumb1_case_shi) .popsection @@ -421,7 +421,7 @@ ENTRY(__gnu_thumb1_case_uhi) lsls r1, r1, #1 add lr, lr, r1 pop {r0, r1} - bx lr + ret lr ENDPROC(__gnu_thumb1_case_uhi) .popsection #endif diff --git a/arch/arm/lib/memcpy.S b/arch/arm/lib/memcpy.S index eee7a219ce..a1c996f94e 100644 --- a/arch/arm/lib/memcpy.S +++ b/arch/arm/lib/memcpy.S @@ -59,7 +59,7 @@ #endif ENTRY(memcpy) cmp r0, r1 - bxeq lr + reteq lr enter r4, lr @@ -148,7 +148,7 @@ ENTRY(memcpy) str1b r0, ip, cs, abort=21f exit r4, lr - bx lr + ret lr 9: rsb ip, ip, #4 cmp ip, #2 @@ -258,7 +258,7 @@ ENTRY(memcpy) .macro copy_abort_end ldmfd sp!, {r4, lr} - bx lr + ret lr .endm ENDPROC(memcpy) diff --git a/arch/arm/lib/relocate.S b/arch/arm/lib/relocate.S index 5102bfabde..dd6f2e3bd5 100644 --- a/arch/arm/lib/relocate.S +++ b/arch/arm/lib/relocate.S @@ -61,7 +61,7 @@ ENTRY(relocate_vectors) stmia r1!, {r2-r8,r10} #endif #endif - bx lr + ret lr ENDPROC(relocate_vectors) @@ -127,13 +127,7 @@ relocate_done: mcr p15, 0, r0, c7, c10, 4 /* drain write buffer */ #endif - /* ARMv4- don't know bx lr but the assembler fails to see that */ - -#ifdef __ARM_ARCH_4__ - mov pc, lr -#else - bx lr -#endif + ret lr ENDPROC(relocate_code) diff --git a/arch/arm/lib/setjmp.S b/arch/arm/lib/setjmp.S index 176a1d5315..2f041aeef0 100644 --- a/arch/arm/lib/setjmp.S +++ b/arch/arm/lib/setjmp.S @@ -17,7 +17,7 @@ ENTRY(setjmp) mov ip, sp stm a1, {v1-v8, ip, lr} mov a1, #0 - bx lr + ret lr ENDPROC(setjmp) .popsection @@ -31,6 +31,6 @@ ENTRY(longjmp) bne 1f mov a1, #1 1: - bx lr + ret lr ENDPROC(longjmp) .popsection
There is currently a problem that U-Boot can not work on ARMv4 because assembly imlementations of memcpy() and some other functions use "bx lr" instruction that is not available on ARMv4 ("mov pc, lr" should be used instead). A working preprocessor-based solution to this problem is found in arch/arm/lib/relocate.S. Move it to the "ret" macro in arch/arm/include/asm/assembler.h and change all "bx lr" code to "ret lr" in functions that may run on ARMv4. Linux source code deals with this problem in the same manner. Signed-off-by: Sergei Antonov <saproj@gmail.com> CC: Samuel Holland <samuel@sholland.org> CC: Ye Li <ye.li@nxp.com> CC: Simon Glass <sjg@chromium.org> CC: Andre Przywara <andre.przywara@arm.com> CC: Marek Vasut <marex@denx.de> CC: Sean Anderson <sean.anderson@seco.com> CC: Tom Rini <trini@konsulko.com> CC: Wolfgang Denk <wd@denx.de> --- arch/arm/include/asm/assembler.h | 6 ++++++ arch/arm/lib/lib1funcs.S | 8 ++++---- arch/arm/lib/memcpy.S | 6 +++--- arch/arm/lib/relocate.S | 10 ++-------- arch/arm/lib/setjmp.S | 4 ++-- 5 files changed, 17 insertions(+), 17 deletions(-)