------- Additional Comments From daney at gcc dot gnu dot org 2005-02-22 22:02
-------
The new 'C' test case will reproduce the bug on:
$ mipsel-linux-gcc -v
Using built-in specs.
Target: mipsel-linux
Configured with: ../gcccvsmain/gcc/configure --target=mipsel-linux
--with-sysroot=/usr/local/mipsel-linux-test --with-arch=mips32 --with-float=soft
--prefix=/usr/local/mipsel-linux-test --with-system-zlib
--enable-languages=c,c++,java
Thread model: posix
gcc version 4.0.0 20050222 (experimental)
I have not been able to generate a 'C' test case that fails on 3.4.x. A similar
java testcase failed on 3.3.1, but I have not tried the attached java case on
3.3.1.
Here is the setup for failure:
$ mipsel-linux-gcc -O3 -c tm.c
$ mipsel-linux-gcc -O3 -c tm_driver.c
$ mipsel-linux-gcc -o tm tm_driver.o tm.o
Then on the host:
#./tm
The answer is: 3989327253662920704
Here is the setup for success:
$ mipsel-linux-gcc -O0 -c tm.c
$ mipsel-linux-gcc -O3 -c tm_driver.c
$ mipsel-linux-gcc -o tm tm_driver.o tm.o
Then on the host:
# ./tm
The answer is: 1109030400000
If we look at the relevent part of the disassembly we see:
li $6,86376448 # 0x5260000
ori $6,$6,0x5c00
multu $2,$6
lw $28,16($sp)
li $5,92340224 # 0x5810000
ori $5,$5,0x1392
lw $25,%call16(f4)($28)
mflo $4
sw $4,24($sp)
lw $4,%got($LC0)($28)
madd $6,$3
The problem is that before the madd, we need to copy the contents of hi -> low
and then clear hi to zero. This is not happening.
--
What |Removed |Added
----------------------------------------------------------------------------
Known to fail| |4.0.0 3.4.3 3.3.1
Summary|MIPS wrong-code for 64-bit |MIPS wrong-code for 64-bit
|multiply from java program. |multiply.
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19683
