https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61225
--- Comment #2 from zhenqiang.chen at linaro dot org --- Patch @r210457 enhances move_insn_for_shrink_wrap to sink more instructions out of entry block. For ICE, it tries to sink a SP adjustment instruction. But some useful information is lost during sinking. A patch is in testing. shrink-wrap-loop.c is a new test. I will update it to skip it for "-m32". As pr49095.c, peephole2 generates inefficient code with the changes in shrink-wrap. Need more investigation to identify the root cause.