testgc2: Make sure GC allocations are not optimized out. #1191
Conversation
|
Now just using a global variable. This is enough for LDC without link-time optimizations turned on to not optimize the allocations away, but of course not an ideal solution. If anybody has a better idea, please let me know. |
|
Seems OK to me. |
| assert(0); | ||
| dummy = new byte[size_t.max / 3]; | ||
| version (Windows) | ||
| assert(0); |
There was a problem hiding this comment.
This shouldn't be Windows only, right?
There was a problem hiding this comment.
To be honest, I have no idea why it is in there, so I just didn't touch it. Speculation mode on: Allocating size_t.max / 3 bytes might be possible on some 32 bit systems if they provide an uncluttered address space to applications, but maybe this is guaranteed to not happen on Windows?
There was a problem hiding this comment.
Ah, now you shed some light on this. Windows has only 3GB address space for 32 bit applications.
If it's not an actual problem then it's good for now. If LTO is really breaking this you could try to make it a shared variable and/or use atomicStore/asm. |
|
You know, on second thought... Why does this test even exist? Shouldn't it be in druntime, if anything? And even then, don't we have plenty of code virtually everywhere that tests this stuff? And besides, maintaining it seems to turn into a battle against the compiler's optimizer(s) which may behave differently over time. Or am I being silly? |
|
No, you're not being silly. Some of the tests in the test suite are very, very old (years older than D1.00), and date from long before the runtime even existed. Even now, there are very few tests in druntime. All these kind of tests definitely belong there. On the other hand, we currently have no mechanism for stand-alone tests for phobos and druntime. |
|
Using the allocated data in a call to a runtime function (getCapacity?) would be a more robust way to prevent it from being optimized out. |
|
Rebased and changed the code to just |
|
Ping? |
Tabs -> spaces, removed empty version(none) block.
|
Rebased again. Should be entirely uncontroversial now (well, in theory a clever compiler could just set capacity to 0 and remove the allocations anyway, but I doubt this will ever be worth it). |
testgc2: Make sure GC allocations are not optimized out.
This makes sure the GC allocations are not optimized out, causing the assert to be hit, in addition to some cosmetic changes in a separate commit.