AArch64 Compiling

packetnathan · 2016-12-22, 20:01

I'm attempting to compile mprime on an aarch64-based system and am running into some issues.

- Can you provide the source code for factor64.c so that it can be compiled for aarch64?
- Can you provide guidance on where the fldcw and fnstcw instructions are being referenced when compiling gwnum? Presumably cctraipy.s is being generated in one of the compilation steps but I'm not seeing it.

It appears as though mprime or some supporting libraries rely specifically on instructions that aren't available in the ARMv8 CPU architecture, specifically fnstcw and fldcw.

Code:

root@aarch64-test:~/p95v2810/gwnum# make clean && make -f makeaarch64
gcc -I.. -DAARCH64 -O2 -c cpuid.c
gcc -I.. -DAARCH64 -O2 -c gwnum.c
gcc -I.. -DAARCH64 -O2 -c gwtables.c
gcc -I.. -DAARCH64 -O2 -c gwthread.c
gcc -I.. -DAARCH64 -O2 -c gwini.c
gwini.c: In function ‘writeIniFile’:
gwini.c:148:3: warning: ignoring return value of ‘write’, declared with attribute warn_unused_result [-Wunused-result]
   (void) _write (fd, buf, (unsigned int) strlen (buf));
   ^
gcc -I.. -DAARCH64 -O2 -c gwutil.c
g++ -I.. -I../qd -DAARCH64 -O2 -c gwdbldbl.cpp
/tmp/cctraipy.s: Assembler messages:
/tmp/cctraipy.s:4374: Error: unknown mnemonic `fnstcw' -- `fnstcw [x29,116]'
/tmp/cctraipy.s:4383: Error: unknown mnemonic `fldcw' -- `fldcw [x29,118]'
/tmp/cctraipy.s:4716: Error: unknown mnemonic `fldcw' -- `fldcw [x29,116]'
/tmp/cctraipy.s:4928: Error: unknown mnemonic `fnstcw' -- `fnstcw [x29,52]'
/tmp/cctraipy.s:4937: Error: unknown mnemonic `fldcw' -- `fldcw [x29,54]'
/tmp/cctraipy.s:4998: Error: unknown mnemonic `fldcw' -- `fldcw [x29,52]'
<snip>
makeaarch64:32: recipe for target 'gwdbldbl.o' failed
make: *** [gwdbldbl.o] Error 1

It also seems as though factor64.o is provided without the source code and as such ld can't use it since it needs a different ELF loader obviously.

Code:

/usr/bin/ld: factor64.o: Relocations in generic ELF (EM: 62)
factor64.o: error adding symbols: File in wrong format

fivemack · 2016-12-22, 22:50

The fnstcw and fldcw instructions manipulate the control word for the x87 FPU, to determine whether instructions will work in single, double or extended precision, as well as the rounding mode and the infinity control; the rounding mode is fairly critical for getting double-double-precision to work.

The instructions are generated in qd/dd.h around line 260

There is nothing analogous to the precision control in AArch64, though fiddling with the FPCR will let you set flush-to-zero, NaN propagation, and the rounding mode:

Code:

MRS <Xt>, FPCR; Read FPCR into Xt
MSR FPCR, <Xt>; Write Xt to FPCR

Prime95 · 2016-12-23, 01:36

The rather harmless gwini.c warning should be fixed in 29.1 although source is not available yet.

Factor64.o is created from MASM source. I can give you the source code, but MASM only runs under windows. I then use Agner Fog's objconv program to convert from Windows COFF object file format to ELF file format. The easiest solution would be to see if there are any options in Agner Fog's program to create an object file you can use.

In version 29.1, I've started using HJWasm which is an open source MASM replacement. Thus, in version 29.1 and later I can provide the MASM sources for you to assemble in a variety of environments. See if HJWasm works for you. If it does that provides us with a path forward.

Version 29.1 also uses the hwloc library -- I don't know if a version of that library is available for your architecture.

2016-12-22, 22:50	#2
fivemack (loop (#_fork)) Feb 2006 Cambridge, England 2³·11·73 Posts	The fnstcw and fldcw instructions manipulate the control word for the x87 FPU, to determine whether instructions will work in single, double or extended precision, as well as the rounding mode and the infinity control; the rounding mode is fairly critical for getting double-double-precision to work. The instructions are generated in qd/dd.h around line 260 There is nothing analogous to the precision control in AArch64, though fiddling with the FPCR will let you set flush-to-zero, NaN propagation, and the rounding mode: Code: MRS <Xt>, FPCR; Read FPCR into Xt MSR FPCR, <Xt>; Write Xt to FPCR

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Compiling mfaktc on Mac OS X	dozba	GPU Computing	23	2017-10-03 19:50
Compiling GNFS sievers on AArch64 platform	wombatman	Programming	11	2017-03-11 03:12
compiling GMP-ECM	ATH	GMP-ECM	69	2017-01-04 12:03
compiling mprime	ATH	Software	9	2015-02-15 12:59
Compiling 24.14	CBoland	Software	6	2007-08-01 00:11

2016-12-23, 01:36	#3
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 1110101110111₂ Posts	The rather harmless gwini.c warning should be fixed in 29.1 although source is not available yet. Factor64.o is created from MASM source. I can give you the source code, but MASM only runs under windows. I then use Agner Fog's objconv program to convert from Windows COFF object file format to ELF file format. The easiest solution would be to see if there are any options in Agner Fog's program to create an object file you can use. In version 29.1, I've started using HJWasm which is an open source MASM replacement. Thus, in version 29.1 and later I can provide the MASM sources for you to assemble in a variety of environments. See if HJWasm works for you. If it does that provides us with a path forward. Version 29.1 also uses the hwloc library -- I don't know if a version of that library is available for your architecture.