Re: nontrapping math in Java

[Bryce McKinlay <bryce at waitaki dot otago dot ac dot nz>]

Jan Hubicka wrote:

>Did someone run benchmark with and without -fnon-call-exceptions?
>

I did.

My testing with the small "Sieve.java" on PowerPC @ -O2 showed no 
difference to generated code as a result of -fnon-call-exceptions, 
except for more unwind info being generated.

With a larger benchmark (jBYTEmark) on x86, there are some differences:

$ gcj -O2 *.java -o jbyte --main=jBYTEmark
$ ./jbyte
jBYTEmark ver. 0.9 +
OS: Linux 2.4.9-13
VM: GNU libgcj 3.2 20020324 (experimental)

Numeric Sort:            161.58
String Sort:              42.61
Bitfield Operations:     244.57
FP Emulation:            127.11
Fourier:                   8.22
Assignment:              102.92
IDEA Encryption:          27.62
Huffman Compression:     108.35
Neural Net:              111.54
LU Decomposition:         73.33
                        -------
Integer Index:            94.22
FP Index:                 40.67

$ gcj -O2 *.java -o jbyte --main=jBYTEmark -fnon-call-exceptions
$ ./jbyte
jBYTEmark ver. 0.9 +
OS: Linux 2.4.9-13
VM: GNU libgcj 3.2 20020324 (experimental)

Numeric Sort:            163.63
String Sort:              42.84
Bitfield Operations:     212.46
FP Emulation:            143.70
Fourier:                   8.19
Assignment:              102.53
IDEA Encryption:          27.67
Huffman Compression:     108.54
Neural Net:              111.54
LU Decomposition:         80.44
                        -------
Integer Index:            94.22
FP Index:                 41.88

I get similar results with "scimark2". So on my P3, some tests get 
faster, some get slower, and the overall result is about the same.

regards

Bryce.