From mboxrd@z Thu Jan 1 00:00:00 1970 X-Msuck: nntp://news.gmane.org/gmane.linux.lib.musl.general/11281 Path: news.gmane.org!.POSTED!not-for-mail From: Rich Felker Newsgroups: gmane.linux.lib.musl.general Subject: Re: [MUSL] microMIPS32R2 O32 port Date: Mon, 24 Apr 2017 09:48:08 -0400 Message-ID: <20170424134808.GQ17319@brightrain.aerifal.cx> References: <20170407141941.GQ17319@brightrain.aerifal.cx> <20170412192535.GG2082@port70.net> <20170412202721.GY17319@brightrain.aerifal.cx> <20170413090036.GH2082@port70.net> <20170421133300.GE17319@brightrain.aerifal.cx> Reply-To: musl@lists.openwall.com NNTP-Posting-Host: blaine.gmane.org Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii X-Trace: blaine.gmane.org 1493041708 18757 195.159.176.226 (24 Apr 2017 13:48:28 GMT) X-Complaints-To: usenet@blaine.gmane.org NNTP-Posting-Date: Mon, 24 Apr 2017 13:48:28 +0000 (UTC) User-Agent: Mutt/1.5.21 (2010-09-15) Cc: Szabolcs Nagy , "musl@lists.openwall.com" , Andre McCurdy To: Jaydeep Patil Original-X-From: musl-return-11296-gllmg-musl=m.gmane.org@lists.openwall.com Mon Apr 24 15:48:24 2017 Return-path: Envelope-to: gllmg-musl@m.gmane.org Original-Received: from mother.openwall.net ([195.42.179.200]) by blaine.gmane.org with smtp (Exim 4.84_2) (envelope-from ) id 1d2eLn-0004kA-5F for gllmg-musl@m.gmane.org; Mon, 24 Apr 2017 15:48:23 +0200 Original-Received: (qmail 16163 invoked by uid 550); 24 Apr 2017 13:48:27 -0000 Mailing-List: contact musl-help@lists.openwall.com; run by ezmlm Precedence: bulk List-Post: List-Help: List-Unsubscribe: List-Subscribe: List-ID: Original-Received: (qmail 16142 invoked from network); 24 Apr 2017 13:48:26 -0000 Content-Disposition: inline In-Reply-To: Original-Sender: Rich Felker Xref: news.gmane.org gmane.linux.lib.musl.general:11281 Archived-At: 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. 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? > >> 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? > >> 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. Rich