Destroyerrrocket wrote: The reason for puting the classes with DeclContext closer together is to achieve a better code generation. Here's an example of the asssembly with just the removal of the DECL_CONTEXT_BASE macro: ``` clang::Decl::castToDeclContext(clang::Decl const*): # @clang::Decl::castToDeclContext(clang::Decl const*) .L_ZN5clang4Decl17castToDeclContextEPKS0_$local: movl 28(%rdi), %edx leaq .LJTI66_0(%rip), %rsi movq %rdi, %rax movl $40, %ecx andl $127, %edx decl %edx movslq (%rsi,%rdx,4), %rdx addq %rsi, %rdx jmpq *%rdx .LBB66_5: movl $48, %ecx addq %rcx, %rax retq .LBB66_1: addq %rcx, %rax retq .LBB66_2: movl $72, %ecx addq %rcx, %rax retq .LBB66_6: movl $64, %ecx addq %rcx, %rax retq .LBB66_7: movl $56, %ecx addq %rcx, %rax retq .LBB66_8: .LJTI66_0: <jump table> ``` While this is better, it is still not performing ideally compared to just the load and an add of the final implementation. I'm of course open to suggestions if there is a better way of getting llvm to emit the right thing.
I agree with @cor3ntin that a comment explaining the need for this order priorization is needed. https://github.com/llvm/llvm-project/pull/76825 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
