RE: [MUSL] microMIPS32R2 O32 port

[Jaydeep Patil <Jaydeep.Patil@imgtec.com>]

>-----Original Message-----
>From: Rich Felker [mailto:dalias@aerifal.cx] On Behalf Of Rich Felker
>Sent: 24 April 2017 PM 07:18
>To: Jaydeep Patil
>Cc: Szabolcs Nagy; musl@lists.openwall.com; Andre McCurdy
>Subject: Re: [musl] [MUSL] microMIPS32R2 O32 port
>
>On Mon, Apr 24, 2017 at 05:30:34AM +0000, Jaydeep Patil wrote:
>> >-----Original Message-----
>> >From: Rich Felker [mailto:dalias@aerifal.cx] On Behalf Of Rich Felker
>> >Sent: 21 April 2017 PM 07:03
>> >To: Jaydeep Patil
>> >Cc: Szabolcs Nagy; musl@lists.openwall.com; Andre McCurdy
>> >Subject: Re: [musl] [MUSL] microMIPS32R2 O32 port
>> >
>> >On Thu, Apr 13, 2017 at 10:37:05AM +0000, Jaydeep Patil wrote:
>> >> Hi Szabolcs,
>> >>
>> >> Please find the attached patch.
>> >
>> >I still don't follow the motivation of all the changes in this patch.
>> >See comments below:
>> >
>> >> [...]
>> >> diff --git a/arch/mips/crt_arch.h b/arch/mips/crt_arch.h index
>> >> 9fc50d7..78832b0 100644
>> >> --- a/arch/mips/crt_arch.h
>> >> +++ b/arch/mips/crt_arch.h
>> >> @@ -1,6 +1,7 @@
>> >>  __asm__(
>> >>  ".set push\n"
>> >>  ".set noreorder\n"
>> >> +".set nomicromips\n"
>> >>  ".text \n"
>> >>  ".global _" START "\n"
>> >>  ".global " START "\n"
>> >
>> >Is there a need for crt1.o (or other users of this file like
>> >dlstart/rcrt1) to be built as plain mips rather than micromips? If
>> >so, please explain. Arbitrary changes like this without an
>> >explanation of why they're being made (what was broken before, and
>> >why this is the correct fix) are not acceptable.
>>
>> The crt1.o and __cp_begin are built as pure MIPS functions rather than
>> microMIPS as they are reading data using $ra register (with ISA bit
>> set). We can clear the ISA bit (using INS) and choose to compile these
>> functions as microMIPS.
>
>I see. So all the .s files are assembled as plain (32-bit) MIPS, and these files are
>just affected because the asm is included in .c files that might be compiled in
>microMIPS mode. That makes sense. I think it's probably just better to make
>the code microMIPS-compatible though by clearing the low bits. 

Ok, I will remove the  "nomicromips" restriction and clear the ISA bit instead.

> But syscall_cp.s needs some care because saved instruction pointer values are
>compared against these labels. In micromips mode, do the labels evaluate
>with the +1 low bit offset?

Yes, in microMIPS mode, ISA bit (0th bit) is set for labels. However I don't see any issue with following comparison

pc >= (uintptr_t)__cp_begin && pc < (uintptr_t)__cp_end

The ISA bit will be set even for PC in the saved context.

>> >> diff --git a/arch/mips/reloc.h b/arch/mips/reloc.h index
>> >> b3d59a4..772b3aa 100644
>> >> --- a/arch/mips/reloc.h
>> >> +++ b/arch/mips/reloc.h
>> >> @@ -36,15 +36,23 @@
>> >>  #define CRTJMP(pc,sp) __asm__ __volatile__( \
>> >>  	"move $sp,%1 ; jr %0" : : "r"(pc), "r"(sp) : "memory" )
>> >>
>> >> +/*
>> >> + * When compiled for microMIPS, .align makes sure that .gpword
>> >> + * is placed at word boundary. $ra must point to first .gpword.
>> >> + * ISA bit of $ra must be cleared for microMIPS before using it
>> >> + * as a base address. For MIPS, ISA bit is always zero.
>> >> +*/
>> >>  #define GETFUNCSYM(fp, sym, got) __asm__ ( \
>> >>  	".hidden " #sym "\n" \
>> >>  	".set push \n" \
>> >>  	".set noreorder \n" \
>> >> +	"	.align 2 \n" \
>> >>  	"	bal 1f \n" \
>> >>  	"	 nop \n" \
>> >>  	"	.gpword . \n" \
>> >>  	"	.gpword " #sym " \n" \
>> >> -	"1:	lw %0, ($ra) \n" \
>> >> +	"1:	ins $ra, $0, 0, 1 \n" \
>> >> +	"	lw %0, ($ra) \n" \
>> >>  	"	subu %0, $ra, %0 \n" \
>> >>  	"	lw $ra, 4($ra) \n" \
>> >>  	"	addu %0, %0, $ra \n" \
>> >
>> >By this, do you mean that .gpword is producing a value that's
>> >relative to the actual byte position of the directive rather than the
>> >logical (ISA bit set) value of "." at the point of .gpword?
>>
>> ISA bit is set for all microMIPS symbols (including the dot symbol).
>> The GETFUNCSYM macro cannot be compiled as pure MIPS code as it might
>> be expanded in a microMIPS function. We need to clear the ISA bit of
>> $ra before using it as a base address in load/store instructions.
>
>OK. And the same appproach would work for crt_arch.h and other places,
>right?

Yes.

>> >> diff --git a/src/thread/mips/syscall_cp.s
>> >> b/src/thread/mips/syscall_cp.s index d284626..9c5f55e 100644
>> >> --- a/src/thread/mips/syscall_cp.s
>> >> +++ b/src/thread/mips/syscall_cp.s
>> >> @@ -1,5 +1,5 @@
>> >>  .set    noreorder
>> >> -
>> >> +.set    nomicromips
>> >>  .global __cp_begin
>> >>  .hidden __cp_begin
>> >>  .type   __cp_begin,@function
>> >
>> >I'm also unclear on the motivation of this one. Before (v1) you had a
>> >lot of changes to replace .s files with something
>> >micromips-compatible (removing branch delay slots); now (v2) those
>> >changes are not included. So are .s files even being built as
>> >micromips at all? If not, why is the above needed? If so, how do the files
>with delay slots work?
>>
>> Branch delay slots are removed (called as compact instructions) in the
>> newer MIPS/microMIPS cores (in development).
>> The MIPS/microMIPS R2-R6 still support instructions with delay slot.
>> Assembler takes care of converting a BRANCH + NOP to its appropriate
>> compact instruction (BEQ + NOP to BEQC).
>> With the v1 branch I was trying to create generic hand-written
>> assembly which can be used for newer cores with the compact
>> instructions.
>> However I realized that it would appropriate to create a new arch
>> instead of creating generic assembly.
>> Thus in v2 branch I modified only those functions which would create
>> issues when compiled with interlinking on.
>
>Based on the discussions so far, I don't think pure-micromips qualifies as a
>new arch. If it would be possible to take a program compiled as micromips-
>only, and run it with the libc.so/ldso built for plain mips on a machine that
>supports both forms of code, then it's not a separate arch, and as I
>understand it this should be possible.

Yes, in the context of miroMIPSR2-R5, we don't need to create a new arch.

>Rich

I will create v3 if you are OK with this approach.

Thanks,
Jaydeep