https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66790
--- Comment #40 from Bernd Schmidt <bernds at gcc dot gnu.org> --- (In reply to Pierre-Marie de Rodat from comment #38) > Besides, Eric told me live that instead of artificially clearing the > entry basic block’s IN set, we could use the confluence_0 for that. I > did this and took this opportunity to minimize code duplication: the > df_mir_init function now just calls df_mir_reset. Hmm, is that actually the right thing for the entry block? I notice that we simulate artificial defs, but it seems like that set is not conservatively correct. For a simple function taking a single integer argument: ;; entry block defs 0 [ax] 1 [dx] 2 [cx] 4 [si] 5 [di] 7 [sp] 21 [xmm0] 22 [xmm1] 23 [xmm2] 24 [xmm3] 25 [xmm4] 26 [xmm5] 27 [xmm6] 28 [xmm7] 37 [r8] 38 [r9] where it would appear that only regs 5 (DI) and 7 (SP) are actually reliably defined, but df_get_entry_block_def_set adds everything that's a FUNCTION_ARG_REGNO. But the entire set makes it through the mir in/out sets in the entire function. (gdb) p df_dump(stdout) [...] ; mir in 0 [ax] 1 [dx] 2 [cx] 4 [si] 5 [di] 7 [sp] 21 [xmm0] 22 [xmm1] 23 [xmm2] 24 [xmm3] 25 [xmm4] 26 [xmm5] 27 [xmm6] 28 [xmm7] 37 [r8] 38 [r9] [...] ;; mir out 0 [ax] 1 [dx] 2 [cx] 4 [si] 5 [di] 7 [sp] 17 [flags] 21 [xmm0] 22 [xmm1] 23 [xmm2] 24 [xmm3] 25 [xmm4] 26 [xmm5] 27 [xmm6] 28 [xmm7] 37 [r8] 38 [r9] Maybe it's best _not_ to include artificial defs, except for a subset of things that are known good - such as the stack pointer. Apart from that, starting to look good, thanks for the effort. Future versions should probably go to gcc-patches.
