Date: Sun, 16 Feb 2003 01:21:44 +0800 From: Huang wen hui <hwh@gddsn.org.cn> To: shudo@computer.org Cc: java@FreeBSD.ORG, stable@freebsd.org Subject: Optimization problem (Re: Math.pow bug for jdk1.3.1-p8 ?) Message-ID: <3E4E7728.8030803@gddsn.org.cn> References: <3E4C9DDD.4040204@gddsn.org.cn> <20030215.175421.1015281127.shudo@localhost> <20030216014802.A10817@misty.eyesbeyond.com> <20030216.005020.894433699.shudo@localhost>
next in thread | previous in thread | raw e-mail | index | archive | help
shudo@computer.org ??: >From: Greg Lewis <glewis@eyesbeyond.com> > > > >>On Sat, Feb 15, 2003 at 05:54:21PM +0900, shudo@computer.org wrote: >> >> >>>I could confirm __j__ieee754_pow() in libjava.so behaves wrong and one >>>in libjava_g.so works correctly. >>> >>>I also tried writing a test program which calls __j__ieee754_pow() and >>>linked the program with e_pow.o, w_sqrt.o, e_sqrt.o, s_fabs.o, >>>s_scalbn.o and s_copysign.o generated during compilation process of >>>JDK 1.3.1. The program produces the incorrect value (0, not 512). >>> >>>It is certain that there is a problem around fdlibm, FreeBSD's gcc >>>2.95.4 or how the fdlibm compiled (compiler flags?). >>> >>> >>Sounds like its quite possibly due to problems with optimisation. >>Obviously, java_g uses no optimisation when compiling, >> >> > >I agree with you. > >Next thing we can do to track the cause down is to compile e_pow.c, >w_sqrt.c, e_sqrt.c, s_fabs.c, s_scalbn.c and s_copysign.c with >different compiler options (-O0, -O2 and so on) and see the result of >the ieee754_pow() function. It may be good to supply different >options to each source code and identify which source code is badly >affected by optimization. > >We should report the bug of FreeBSD's gcc 2.95.4 to appropriate >persons if we could recognize the problem is due to the gcc. > > Kazuyuki Shudo shudo@computer.org http://www.shudo.net/ > >To Unsubscribe: send mail to majordomo@FreeBSD.org >with "unsubscribe freebsd-java" in the body of the message > > > > This code is from j2sdk1.3.1/src/share/native/java/lang/fdlibm/src/e_pow.c in jdk13 ports, gcc's optimization may has problem. #include <stdio.h> #define __HI(x) *(1+(int*)&x) #define __LO(x) *(int*)&x double tst(double x, double y) { double z,ax,z_h,z_l,p_h,p_l; double y1,t1,t2,r,s,t,u,v,w; int i,j,k,yisint,n; int hx,hy,ix,iy; unsigned lx,ly; u=1; v=1; t1=u+v; __LO(t1) = 0; t2=10; y1 = y; __LO(y1) = 0; p_l = (y-y1)*t1+y*t2; p_h = y1*t1; z = p_l+p_h; // return t1; //ok j = __HI(z); // return t1; //bad t1 return t1; } main() { double d = tst(2.0,9.0); printf("%f\n", d); } ----output---- gcc -O3 -o tst tst.c # ./tst 0.000000 # gcc -O2 -o tst tst.c # ./tst 0.000000 # gcc -O1 -o tst tst.c # ./tst 0.000000 # gcc -O0 -o tst tst.c # ./tst 2.000000 # gcc -v Using builtin specs. gcc version 2.95.4 20020320 [FreeBSD] # uname -a FreeBSD wfdb.gddsn.org.cn 4.7-STABLE FreeBSD 4.7-STABLE #0: Fri Feb 14 12:04:07 CST 2003 root@wfdb.gddsn.org.cn:/usr/obj/usr/src/sys/WFDB i386 To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-stable" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?3E4E7728.8030803>