On 10/22/2015 10:24 AM, Kyrill Tkachov wrote:
Hi all,The ICE in this PR (reduced from LLVM) occurs because the atomic_load builtin expansion ends up calling expand_atomic_compare_and_swap without caring about the result. The convention in the optabs/builtin machinery is to use const0_rtx for the target destination when that is the case. However, expand_atomic_compare_and_swap didn't check for that, only for the case where the pointer to the target is NULL. This patch fixes that by handling the case where the ptarget_oval and ptarget_bool pointers are not-NULL but contain const0_rtx. N.B. This is one of the fallback paths for expanding atomic_load when the target doesn't provide expanders for it, like the arm target with -march=armv4t. Bootstrapped and tested on arm-none-linux-gnueabihf, aarch64-linux-gnu, x86_64-unknown-linux-gnu. This ICE appears on the GCC 5 and GCC 4.9 branch. I'll be bootstrapping and testing this patch on those branches. Meanwhile, is this ok for trunk? Thanks, Kyrill 2015-10-22 Kyrylo Tkachov <kyrylo.tkac...@arm.com> PR middle-end/67989 * optabs.c (expand_atomic_compare_and_swap): Handle case when ptarget_oval or ptarget_bool are const0_rtx. 2015-10-22 Kyrylo Tkachov <kyrylo.tkac...@arm.com> PR middle-end/67989 * g++.dg/pr67989.C: New test.
OK. jeff
