On Sat, Aug 4, 2018 at 6:44 AM Tom Rini trini@konsulko.com wrote:
On Sat, Aug 04, 2018 at 05:32:18AM -0500, Adam Ford wrote:
On Fri, Aug 3, 2018 at 10:20 AM Tom Rini trini@konsulko.com wrote:
On Fri, Aug 03, 2018 at 10:02:17AM -0500, Adam Ford wrote:
On Fri, Aug 3, 2018 at 10:00 AM Tom Rini trini@konsulko.com wrote:
On Fri, Aug 03, 2018 at 05:43:46AM -0500, Adam Ford wrote:
On Tue, Jul 25, 2017 at 11:08 AM Tom Rini trini@konsulko.com wrote: > > On Tue, Jul 25, 2017 at 08:14:57AM -0500, Adam Ford wrote: > > > The readme file for OMAP indicates that we compile using armv5 to "to > > allow more compilers to work" > > > > We have our arch/arm/mach-omap2/omap3/lowlevel_init.S file also noting > > some special assembly insturctions becuse we use armv5. The barriers > > defined also indicate we're using CP15 instead of the separate barrier > > instructions for armv7 because we're using armv5 instead. > > > > I just wonder in this day and age when we're noting a GCC version and > > generating warnings based on the GCC warning, do we still need to > > compile as armv5 any more? It seems like "to allow more compilers to > > work" would not really apply any more we're trying to push newer > > versions of GCC. > > So, these are historical notes that really should be corrected. > Initially, when ARMv7 support was added, most people did not have > compilers new enough to recognize -march=armv7-a. We still even support > them, see the logic in arch/arm/Makefile around CONFIG_CPU_V7 (the > options are any sort of modern gcc, llvm, ancient gcc). When we move to > gcc-6 being the oldest gcc supported for ARM we can fixup those comments > and logic as well.
My understanding is that we've made the requirement for GCC 6 now. I just pushed a patch which enabled mtune=armv7-a-generic when CONFIG_CPU_V7A is enabled and that seems to shrink the code a bit on omap3_logic. Does it make sense to remove the , -march=armv5 from arch/arm/Makefile and or the plain -march=armv7 since CONFIG_CPU_V7A implies armv-a?
Yes, we can probably at least drop the v5 portion of that logic. As noted in the other patch, LLVM and GCC disagree on "armv7" vs "armv7-a" as being the appropriate name, or at least did in the past.
Can you point me to an example of how to tune armv7 for both GCC and LLVM? I was looking around and I'm not seeing an obvious syntax. I'll do a 2-part patch. One to remove the armv5, and one to enable the armv7 optimization.
In arch/arm/Makefile: arch-$(CONFIG_CPU_V7A) =$(call cc-option, -march=armv7-a, \ $(call cc-option, -march=armv7, -march=armv5))
this works in both cases. And doc/README.clang is correct on how to build with LLVM on ARM.
I reset my head to origin/master, but using the following, I am not able to build omap3_logic with clang on Ubuntu 18.04:
make HOSTCC=clang omap3_logic make HOSTCC=clang CROSS_COMPILE=arm-linux-gnueabi- \ CC="clang -target arm-linux-gnueabi" -j8
arch/arm/lib/memset.S:43:2: error: predicated instructions must be in IT block stmiage ip!, {r1, r3, r8, lr} @ 64 bytes at a time. ^ arch/arm/lib/memset.S:44:2: error: predicated instructions must be in IT block stmiage ip!, {r1, r3, r8, lr} ^ arch/arm/lib/memset.S:45:2: error: predicated instructions must be in IT block stmiage ip!, {r1, r3, r8, lr} ^ arch/arm/lib/memset.S:46:2: error: predicated instructions must be in IT block stmiage ip!, {r1, r3, r8, lr} ^ arch/arm/lib/memset.S:48:2: error: predicated instructions must be in IT block ldmfdeq sp!, {r8, pc} @ Now <64 bytes to go. ^ arch/arm/lib/memset.S:53:2: error: predicated instructions must be in IT block stmiane ip!, {r1, r3, r8, lr} ^ arch/arm/lib/memset.S:54:2: error: predicated instructions must be in IT block stmiane ip!, {r1, r3, r8, lr} ^ arch/arm/lib/memset.S:56:2: error: predicated instructions must be in IT block stmiane ip!, {r1, r3, r8, lr} ^ arch/arm/lib/memset.S:103:2: error: predicated instructions must be in IT block stmiane ip!, {r1, r3} ^ arch/arm/lib/memset.S:105:2: error: predicated instructions must be in IT block strne r1, [ip], #4 ^ arch/arm/lib/memset.S:111:2: error: predicated instructions must be in IT block strbne r1, [ip], #1 ^ arch/arm/lib/memset.S:112:2: error: predicated instructions must be in IT block strbne r1, [ip], #1 ^ arch/arm/lib/memset.S:114:2: error: predicated instructions must be in IT block strbne r1, [ip], #1 ^ arch/arm/lib/memset.S:120:2: error: predicated instructions must be in IT block strblt r1, [ip], #1 @ 1 ^ arch/arm/lib/memset.S:121:2: error: predicated instructions must be in IT block strble r1, [ip], #1 @ 1 ^ scripts/Makefile.build:314: recipe for target 'arch/arm/lib/memset.o' failed make[1]: *** [arch/arm/lib/memset.o] Error 1 Makefile:1373: recipe for target 'arch/arm/lib' failed make: *** [arch/arm/lib] Error 2 aford@aford-IdeaCentre-A730:~/src/u-boot$
If you have any suggestions, I'm open to try them. I have never used clang before, but I would really like to get these armv7a optimizations in to make SPL smaller.
I guess I forgot that you will need to turn off ARCH_MEMCPY/ARCH_MEMSET and EFI support for clang. That said, since you're just making sure the right options are being passed in, in both cases, you can also just use V=1 and just build common/main.o or something like that.
I am not getting compiler errors due to the-mtune=generic-armv7-a on clang, or at least the clang version that comes with Ubuntu 18.04
clang -target arm-none-eabi -Wp,-MD,common/.main.o.d -nostdinc -isystem /usr/lib/llvm-6.0/lib/clang/6.0.0/include -Iinclude -I./arch/arm/include -include ./include/linux/kconfig.h -D__KERNEL__ -D__UBOOT__ -Qunused-arguments -Wall -Wstrict-prototypes -Wno-format-security -fno-builtin -ffreestanding -std=gnu11 -fshort-wchar -Os -fno-stack-protector -g -Wno-format-nonliteral -Wno-format-invalid-specifier -Wno-gnu -Wno-address-of-packed-member -Werror=date-time -D__ARM__ -marm -mabi=aapcs-linux -fno-pic -mno-unaligned-access -ffunction-sections -fdata-sections -fno-common -ffixed-r9 -msoft-float -mno-movt -pipe -march=armv7-a -D__LINUX_ARM_ARCH__=7 -mtune=generic-armv7-a -I./arch/arm/mach-omap2/include -D"KBUILD_STR(s)=#s" -D"KBUILD_BASENAME=KBUILD_STR(main)" -D"KBUILD_MODNAME=KBUILD_STR(main)" -c -o common/main.o common/main.c
aford@aford-IdeaCentre-A730:~/src/u-boot$ arm-linux-gnueabi-size common/main.o text data bss dec hex filename 181 0 0 181 b5 common/main.o
If this is good enough for you, I'd like to push a patch that removes the armv5 reference in the Makefile, adds mtune to the Makefile and removes the armv5 reference in doc/README.omap3.
Stock text data bss dec hex filename 50910 429 67580 118919 1d087 spl/u-boot-spl 540713 22700 327072 890485 d9675 u-boot
Without Armv5 text data bss dec hex filename 50916 429 67580 118925 1d08d spl/u-boot-spl 540719 22700 327064 890483 d9673 u-boot
-mtune=generic-armv7-a text data bss dec hex filename 50932 429 67580 118941 1d09d spl/u-boot-spl 540519 22700 327080 890299 d95bb u-boot
SPL grew a little, but U-Boot seems to shrink (with GCC)
I don't know much about the omap3 assembly code found in Lowlevel_init.S, but I was hoping someone might know a better way to optimize (or eliminate) it.
adam
-- Tom