Re: port-mips/57680: printf("%.1f") shows wrong resultsonR3000mipseb

[Izumi Tsutsui <tsutsui%ceres.dti.ne.jp@localhost>]

The following reply was made to PR port-mips/57680; it has been noted by GNATS.

From: Izumi Tsutsui <tsutsui%ceres.dti.ne.jp@localhost>
To: rvp%SDF.ORG@localhost
Cc: gnats-bugs%netbsd.org@localhost, tsutsui%ceres.dti.ne.jp@localhost
Subject: Re: port-mips/57680: printf("%.1f") shows wrong resultsonR3000mipseb
Date: Sat, 4 Nov 2023 12:12:53 +0900

 > On Fri, 3 Nov 2023, Izumi Tsutsui wrote:
 > 
 > > So at least gcc -O2 (including dtoa.c in libc.so) causes
 > > something wrong on MIPS1.
 > >
 > 
 > Very intriguing results! Does adding an explicit arch flag like `-march=mips1'
 > along with `-O2' help?
 
 Looks no change:
 ---
 news3470-% uname -a
 NetBSD galant 9.3 NetBSD 9.3 (GENERIC) #0: Sat Nov  4 02:48:54 JST 2023  tsutsui@mirage:/s/tsutsui/netbsd-src/sys/arch/newsmips/compile/obj.newsmips/GENERIC newsmips
 news3470-% sysctl hw.model
 hw.model = news3400
 news3470-% grep ^cpu /var/run/dmesg.boot
 cpu0 at mainbus0: MIPS R3000 CPU (0x220) Rev. 2.0 with MIPS R3010 FPC Rev. 2.0
 cpu0: 64KB/4B direct-mapped Instruction cache, 64 TLB entries
 cpu0: 64KB/4B direct-mapped write-through Data cache
 news3470-% cc -O2 -march=mips1 -I/s/netbsd-9/src/lib/libc/gdtoa/ -I/s/netbsd-9/src/lib/libc/arch/mips/gdtoa/ -DDEBUG -DHonor_FLT_ROUNDS -DNO_LONG_LONG -o m2-O2-mips1 m.c /s/netbsd-9/src/lib/libc/gdtoa/dtoa.c /s/netbsd-9/src/lib/libc/gdtoa/dmisc.c /s/netbsd-9/src/lib/libc/gdtoa/misc.c
 news3470-% ./m2-O2-mips1 
 .O
 40490FD0
 3*5.8
 439D145A
 ?*8.7
 44C45970
 news3470-% 
 ---
 ('*'s are also non US-ASCII 8bit chars)
 
 > 
 > > (I still wonder if it's compiler bug or improper implementation like
 > >  undefined behaviour etc.)
 > >
 > 
 > I don't know anything about the MIPS arch. so I can't definitively say, but,
 > IMO I think it's the code GCC is generating for MIPS1:
 > 
 > a) One of the first things I tried on a recent 9.3_STABLE running on amd64
 >     was to try out different FE_* rounding modes--just like what you added.
 >     Compile with: cc -m32 -O2 -ffast-math (this last sometimes causes problems).
 > 
 >     No issues.
 > 
 > b) Next I forced the FPU to do the rounding by calling fesetround(FE_*) first
 >     thing. This made some of the final bits in the precision differ, but, this
 >     is expected.
 
 Yes, using FE_TOWARDZERO as default works around on NWS-3470 (R3000+R3010).
 
 > c) The same binaries work on the newer MIPS arches.
 
 Yes, at least on NWS-5000 (R4400 mipseb).
 
 > d) You say it works fine in an emulator.
 
 I've also confirmed on NetBSD/pmax 9.3 on GXemul emulating
 DECstation5000/200 (3MAX) R3000:
  https://gist.github.com/tsutsui/043243066bb77f3f0315c70b53269192
  https://user-images.githubusercontent.com/3054480/280441532-81e27faf-1e5b-4bee-a9a1-170e8ded09ea.png
 
 ---
 
 Unless some MIPS guru has proper idea how this issue should be fixed,
 I'd like to commit workaround fix to disable -DHonor_FLT_ROUNDS in
 src/lib/libc/gdtoa/Makefile.inc in ${MACHINE_CPU}=="mips" case.
 
 Thanks,
 ---
 Izumi Tsutsui