gcc 2.95.2-2 (and gcc 2.95.3-2) bug with long long math

["Robert Blenis" <blenis at mindspring dot com>]

I'm using a ppc-elf cross compiler on Window 2000, and think there is a bug
in the recent gcc compilers (2.95.2 and 2.95.3), that was not present in gcc
2.90.29 (egcs-1.0.3).  Below is a section of the code and resulting
assembler output for gcc 2.90.29 and gcc 2.95.2.  In either case, if I
replace the line "dSdx = 2*A*x;" with a function call to do the equivalent,
the function call assembly does the right thing for both versions, but when
in a large function, gcc 2.95.2 generates bad code. I tried looking at the
source for gcc but am new to it and did not know where to look. Let me know
what other information is needed.

And here are the flags passed to gcc:
-Wall -ansi -B
c:/dvt/bin/ppc-elf/2.95.2/ -DUSE_ECONOCR -mcpu=860 -mmultiple -DDVT -DDVT_RT
OS -DDVT_600 -DENABLE_DECIMAL_MODE -DMARK_IMAGE -O2 -fno-exceptions -I
c:/dvt/include -I c:/dvt/DVTLib/include -I ../600 -I ../ppc -I ../net -I
../Libraries -I ../AB_Enet -I ../AB_Enet/Ported -I ../AB_Enet/Msg -I .. -I
../sensors -I ../dvtrtos -I ../ImageProc -I ../Terminals -I ../Script -I
../Libraries -DDEBUG -DDEBUG_MALLOC -DTOSHIBA_ROIBOT_DRIVER -Wa,-ahls=.lst -
pipe -DNEW_SERIAL_DRIVER

The options used to build/configure gcc were (going from memory):
../gcc-2.90.29/configure --target=ppc-elf --prefix=/tools-ppc -with-headers=
/usr/include --enable-languages=c,c++

and same for gcc-2.95.2 (and 2.95.3)

I'm reporting this to cygwin since I built the compiler with the cygwin
sources, but I also reported the bug to gcc-bugs@gcc.gnu.org.


///////// Source code (prim.c)
#define INT64 long long
#define INT32 int

	INT64 x, y, dSdx, dSdy;


	/* ... lots of other code ... */


  	DbgPrint(TRACE_OF,"  ------- maxnumpnts = %ld\n\r", (INT32)maxnumpnts);

	/* get starting octant */
    x = (INT64)start->X;
    y = (INT64)start->Y;
#if 1
    dSdx = 2*A*x;		/* the code for this line is being generated incorrectly!
*/
#else
    dSdx = FuzzyMath(A,x);
#endif

    dSdy =   2;
    octant = _GetOctant( dSdx, dSdy);


///////// output from gcc 2.90.29

 1434 1228 48000001 		bl DbgPrint
 1435 122c A81B0000 		lha 0,0(27)
 1436 1230 9001003C 		stw 0,60(1)
 1437 1234 7C00FE70 		srawi 0,0,31
 1438 1238 90010038 		stw 0,56(1)
 1439 123c 81C10038 		lwz 14,56(1)
 1440 1240 81E1003C 		lwz 15,60(1)
 1441 1244 38A00000 		li 5,0
 1442 1248 38C00002 		li 6,2
 1443 124c 7DD27378 		mr 18,14			/* copy A (r14,r15) into (r18,r19) */
 1444 1250 7DF37B78 		mr 19,15			/* ... */
 1445 1254 7D4F9814 		addc 10,15,19		/* multiply A by 2 (using original A
and copy of A) */
 1446 1258 7D2E9114 		adde 9,14,18		/* and store result in r9,r10 */
 1447 125c 7E7F5016 		mulhwu 19,31,10
 1448 1260 7E9F51D6 		mullw 20,31,10
 1449 1264 7C1F49D6 		mullw 0,31,9
 1450 1268 7D2AF1D6 		mullw 9,10,30
 1451 126c 92610090 		stw 19,144(1)
 1452 1270 92810094 		stw 20,148(1)
 1453 1274 82810090 		lwz 20,144(1)
 1454 1278 7E940214 		add 20,20,0
 1455 127c 92810090 		stw 20,144(1)
 1456 1280 7E95A378 		mr 21,20
 1457 1284 7EB54A14 		add 21,21,9
 1458 1288 92A10090 		stw 21,144(1)
 1459 128c 80610090 		lwz 3,144(1)
 1460 1290 80810094 		lwz 4,148(1)
 1461 1294 48004029 		bl _GetOctant



///////// output from gcc 2.95.2-2

 1435 127c 48000001 	bl DbgPrint
 1436 1280 A8150000 	lha 0,0(21)
 1437 1284 8101002C 	lwz 8,44(1)
 1438 1288 7C030378 	mr 3,0
 1439 128c 7C00FE70 	srawi 0,0,31
 1440 1290 9061003C 	stw 3,60(1)
 1441 1294 90010038 	stw 0,56(1)
 1442 1298 80810038 	lwz 4,56(1)
 1443 129c 80A1003C 	lwz 5,60(1)
 1444 12a0 7D0B4378 	mr 11,8
 1445 12a4 7CA62B78 	mr 6,5				/* copy A (r4,r5) into (r5,r6) */
 1446 12a8 7C852378 	mr 5,4				/* ... */
 1447 12ac 7D453014 	addc 10,5,6				/* multiply A by 2 (using original A and
copy of A) */
 1448 12b0 7D242914 	adde 9,4,5				/* and store result in r9,r10 */
									/* the problem is the addc 10,5,6 adds r5 and r6, but */
									/* r5 has just been replaced by the contents of r4! */
 1449 12b4 7CC85016 	mulhwu 6,8,10
 1450 12b8 7CE851D6 	mullw 7,8,10
 1451 12bc 81810028 	lwz 12,40(1)
 1452 12c0 7C0B49D6 	mullw 0,11,9
 1453 12c4 90C10080 	stw 6,128(1)
 1454 12c8 90E10084 	stw 7,132(1)
 1455 12cc 7D6A61D6 	mullw 11,10,12
 1456 12d0 7CC60214 	add 6,6,0
 1457 12d4 7CCE3378 	mr 14,6
 1458 12d8 7DCE5A14 	add 14,14,11
 1459 12dc 90C10080 	stw 6,128(1)
 1460 12e0 91C10080 	stw 14,128(1)
 1461 12e4 80610080 	lwz 3,128(1)
 1462 12e8 80810084 	lwz 4,132(1)
 1463 12ec 38A00000 	li 5,0
 1464 12f0 38C00002 	li 6,2
 1465 12f4 48003DD1 	bl _GetOctant


Robert Blenis
DVT Corporation
R&D Engineering Manager
770-449-4960
robert.blenis@dvtsensors.com


--
Want to unsubscribe from this list?
Check out: http://cygwin.com/ml/#unsubscribe-simple