jmorse wrote: G'day -- we've got some tests for -O0 that, with this patch applied, generate different code with-and-without the `-g` flag, if `-fdebug-info-for-profiling` is enabled. Example godbolt: https://godbolt.org/z/qooo5Eah1 .
It seems the intention of this patch is generating additional loads and stores with debug-info enabled, which is usually highly undesirable. It's guarded by `-fdebug-info-for-profiling`, however we (Sony) use this flag in a bunch of places, and would prefer it to continue generating stable + consistent code. A couple of possibilities: * Put this behaviour behind a cc1 flag, I think that was the plan in the original PR? * Instead of creating an alloca, store+load, and dbg.declare, connect a dbg.value to the relevant pointer Value. The 2nd point is more work, but it means you're only adding more debug-info, and nothing else. A dbg.value connected to the pointer Value is the inevitable consequence of the alloca+store+load idiom after mem2reg/SROA anyway. https://github.com/llvm/llvm-project/pull/94100 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
