On Wed, Mar 14, 2018 at 8:03 AM, Alexandre Oliva <aol...@redhat.com> wrote: > On Mar 12, 2018, "Bin.Cheng" <amker.ch...@gmail.com> wrote: > >> internal compiler error: in aarch64_classify_address, at >> config/aarch64/aarch64.c:5678 >> 0xfe3c29 aarch64_classify_address >> /.../build/src/gcc/gcc/config/aarch64/aarch64.c:5677 >> 0xfe8be8 aarch64_legitimate_address_hook_p >> /.../build/src/gcc/gcc/config/aarch64/aarch64.c:5958 >> 0xc0149e default_addr_space_legitimate_address_p(machine_mode, >> rtx_def*, bool, unsigned char) >> /.../build/src/gcc/gcc/targhooks.c:1476 >> 0xb5b9f1 memory_address_addr_space_p(machine_mode, rtx_def*, unsigned char) >> /.../build/src/gcc/gcc/recog.c:1334 >> 0xb5d278 address_operand(rtx_def*, machine_mode) >> /.../build/src/gcc/gcc/recog.c:1073 >> 0xb5e186 asm_operand_ok(rtx_def*, char const*, char const**) >> /.../build/src/gcc/gcc/recog.c:1816 >> 0x73f440 expand_asm_stmt >> /.../build/src/gcc/gcc/cfgexpand.c:3138 >> 0x742d3c expand_gimple_stmt_1 >> /.../build/src/gcc/gcc/cfgexpand.c:3621 > >> Not sure if it reveals latent bug or just inconsistent issue with >> backend though. > > Possibly both, in a way. This triggers at expand, which is much earlier > than the patch affects compilation; that call to address_operand was > already there before. So the ICE was latent, but only in the sense that > we didn't have tests that would have triggered it. Any asm stmt with a > p constraint and a non-address operand (with another constraint to > accept it) would have exercised this aarch64-specific ICE, and it would > do so before hitting the machine-independent ICE that the patch fixed. > > aarch64_classify_address should probably return false instead of failing > that assertion. When it comes to asm operands, there's no guarantee > that you'll get something that even looks like an address, when you're > part of the very code that's supposed to tell the rest of the compiler > what a legitimate address should look like. Yes, the assert imposes stronger condition on input, which it could simply return false indicating an invalid address. I will test a change to see if there is any fallout.
Thanks, bin > > -- > Alexandre Oliva, freedom fighter http://FSFLA.org/~lxoliva/ > You must be the change you wish to see in the world. -- Gandhi > Be Free! -- http://FSFLA.org/ FSF Latin America board member > Free Software Evangelist|Red Hat Brasil GNU Toolchain Engineer
