I think this is one of the issues... saying that things are optimizations without citing even a single perf number. I still think that the biggest perf regression from -fno-bullshit would be the loss of strict aliasing.
Conversation
But folks describing things like llvm's undef or signed overflow as an "optimization" is disingenuous since *lots* of code doesn't benefit from those "optimizations".
1
There's nothing disingenuous about saying that an optimization is an optimization. Many of the things that have been mentioned in this thread have a substantial impact on optimization including the basic memory safety guarantees, which optimizers lean on very heavily.
2
1
There have just been a whole bunch of threads about these issues. It's you being disingenuous by picking out 2 specific things that are not a huge part of how optimizations are implemented. No one has claimed that those 2 things enable major optimizations. It's a strawman.
1
1
I have a large list of things that don't enable major optimizations. Not sure which 2 things you think I'm picking out. So I'll repeat my point again: the UB feature that most helps perf is strict aliasing. Everything else is crumbs.
1
The basic memory safety guarantees provide a lot more broadly used information about memory dependencies / aliasing than type-based aliasing metadata. Even the guarantees for pointer arithmetic have a huge impact. Just try compiling code using iterators without those guarantees.
2
Simply taking away the basic memory safety and pointer arithmetic guarantees forbids most code motion (like hoisting out of loops), makes the compiler unable to see that tons of loops terminate or only access within the bounds of an object, etc. It kills most loop optimizations.
1
Even the simple loop optimizations, not just more aggressive things like unrolling and vectorization. The optimization stack is based on basic guarantees from the assumption that memory safety isn't broken. No data races, no out-of-bounds accesses, no use-after-free, etc.
2
If you force the compiler to assume that memory can change at any point, you forbid tons of optimization / code movement. Every single memory access has to act as a strong, sequential memory barrier. Even optimization in code without explicit pointers ends up disallowed.
1
Type-based alias analysis is a tiny part of it. Most of the basic information needed to perform optimizations is provided by the guarantees of memory safety - no out-of-bounds accesses, no data races, etc. so memory doesn't just change without a write through a derived pointer.