RFR: 8231895: Avoid String allocations in JVM_FindLoadedClass

[Ioi Lam]

Looks good.

We have a few places that convert from . to /. I am wondering if we 
should consolidate them to a single place:

open/src/hotspot/share$ find . -name \*.cpp -print -o -name \*.hpp 
-print | xargs grep "'/'" | grep '[^=!]=[^=]'
./prims/unsafe.cpp:      if (utfName[i] == '.')   utfName[i] = '/';
./compiler/compilerOracle.cpp:      if (name[i] == '.')  name[i] = '/';  
// package prefix uses '/'
./compiler/methodMatcher.cpp:      if (*lp == '.')  *lp = '/'; // dots 
build the package prefix
./classfile/javaAssertions.cpp:    if (name_copy[i] == '.') name_copy[i] 
= '/';

Thanks
- Ioi


On 10/5/19 3:50 PM, Claes Redestad wrote:
> Hi,
>
> JVM_FindLoadedClass allocates a java String from JVM code to transform
> class names from external to internal representation (java.lang.Class ->
> java/lang/Class). This is both inefficient and redundant since we'll
> directly transform it to a utf-8 char* and then ignore the newly created
> String.
>
> Instead we should mutate the C string. There's a small improvement in
> a targeted microbenchmark (order of 100ns/op, see bug for details),
> although effect on most applications is likely to be in the noise.
>
> Webrev: http://cr.openjdk.java.net/~redestad/8231895/open.00/
> Bug:    https://bugs.openjdk.java.net/browse/JDK-8231895
>
> Testing: tier1-3
>
> Thanks!
>
> /Claes