Switch default "-mcpu" for 32-bit ARCv3 processors from "hs5x" to "hs58"

[abrodkin]

GCC for 32-bit ARCv3 processors supports 2 mcpu values: hs5x & hs58. With the only difference of hs58 supporting dual loads/stores (AKA ldd, std, llockd & scondd instructions) while hs5x doesn't. See https://foss-for-synopsys-dwc-arc-processors.github.io/toolchain/baremetal/gcc-mcpu.html#arcv3-toolchain. I.e. -mcpu=hs58 is exactly the same as -mcpu=hs5x -mll64, or in reverse -mcpu=hs5x is exactky the same as -mcpu=hs58 -mno-ll64.

Now in our current Linxu port for ARCv3 we have by default -mcpu=hs5x and...

ifndef CONFIG_ARC_HAS_LL64
cflags-y                += -mno-ll64
endif

I.e. if in the kernel configuration we unset CONFIG_ARC_HAS_LL64 then -mno-ll64 is used, while n case of enabled CONFIG_ARC_HAS_LL64 nothing really happens. Which means effectively dual loads/stores are not used now at all.

But given reverse logic of CONFIG_ARC_HAS_LL64 processing is still used for ARCv2, we don't want to change that. Thus leaves us just one option - switching -mcpu to the one which enables dual loads/stores by default. So, let's finally do it.

[pavelvkozlov]

For some reason combination of -mcpu=hs58 and -mno-ll64 doesn't work as expected. Adding -mno-ll64 has no effect and compiler continue to put ldd/std instructions. Does compiler handle option -mno-ll64? At the same time combination of -mcpu=hs5x and -mll64works as expected.

[abrodkin]

For some reason combination of -mcpu=hs58 and -mno-ll64 doesn't work as expected. Adding -mno-ll64 has no effect and compiler continue to put ldd/std instructions. Does compiler handle option -mno-ll64? At the same time combination of -mcpu=hs5x and -mll64works as expected.

How do you tell if -mno-ll64 works or not? Just grepping the vmlinux disassembly? I guess some ldd/std come from assembly code, thus the compiler option doesn't affect that part. But if you count a number of ldd/std instructions with -mcpu=hs58 vs -mcpu=hs58 -mno-ll64 I expect that you'll see the difference.

[pavelvkozlov]

Yes, I see ldd/std instructions in the vmlinux disassembly (C functions like start_kernel, mount_root and others), even without option CONFIG_ARC_HAS_LL64. I talk about situation after my change in Makefile from current -mcpu=hs5x to -mcpu=hs58. I've also checked combination of -mcpu=hs58 -mno-ll64 on simple C program:

#include <stdio.h>
#include <inttypes.h>

int main(int argc, char **argv)
{
    uint64_t value = 0x100000001ull;

    printf("64-bit value %llx", value);
    return 0;
}

arc32-linux-gcc -mcpu=hs58 -mno-ll64 ldd_test.c -o ldd_test

And I see than -mno-ll64 option has no effect. With option -mcpu=hs58 compiler puts ldd/std instructions with and without -mno-ll64. Compiler for archs with -mcpu=hs38 doesn't have this issue and option -mno-ll64 works as expected. I've been using latest ARC 2022.09 toolchain for arc32. @claziss, can you please also look?

So, if this issue be confirmed, I want offer to use combination of -mcpu=hs5x and -mll64 in Makefile for ARCv3 32 Linux kernel to workaround it.

[claziss]

For hs58 cpu the ll64 is enabled by default and cannot be turned off, as the compiler will use the ll64 optimized libraries. If you don't want ll64 option on, please use the default hs5x.

[pavelvkozlov]

@claziss, thank you for your comment. Is there another difference between use of -mcpu=hs5x and -mcpu=hs58 or just ll64? What is preferable? I'm going to use -mcpu=hs5x + -mll64 to enable ll64.

[claziss]

@pavelvkozlov not yet, but we may choose to grow this difference. More info here: https://github.com/foss-for-synopsys-dwc-arc-processors/toolchain/blob/arc-dev/doc/baremetal/gcc-mcpu.rst And I think your solution is the right one.

[pavelvkozlov]

Thanks to all for comments. I've created PR #106. To provides correct logic for HS5x, I had to add extra conditions. Now ldd/std instructions appear as expected. @abrodkin please take a look. Is it ok?

[pavelvkozlov]

Merged. Commit: c8444e0