https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116275
--- Comment #4 from Roger Sayle <roger at nextmovesoftware dot com> --- Created attachment 58868 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=58868&action=edit proposed patch Here's my proposed fix (the first of two patches) that resolves the ICE with the testcase. The problem is that i386.md has a *extenddi2_doubleword_highpart pattern on !TARGET_64BIT for noticing that sign-extending a doubleword can be optimized to just manipulating the highpart. The STV pass recognizes this as a candidate, where previously ASHIFTRT of V2DI required AVX512VL, which runs into problems as the post-reload splitter *extendv2di2_highpart_stv is conditional on AVX512VL. The first patch resolves this problem by adding a pre-reload splitter for this case, called imaginatively called *extendv2di2_highpart_stv_noavx512vl. This relies on split1's recursive splitting, so that the generated code for (x<<8)>>8 in the testcase becomes: vpsllq $8, %xmm0, %xmm0 vpsrad $8, %xmm0, %xmm1 vpsrlq $8, %xmm0, %xmm0 vpblendd $5, %xmm0, %xmm1, %xmm0 I'll post a follow-up patch (part 2) that provides a better implementation for this V2DI highpart sign extension, which can be done in 3 insns (taking advantage of SSE's ability to arithmetic shift right the "highpart"): vpsllq $8, %xmm0, %xmm1 vpsrad $8, %xmm1, %xmm1 vpblendd $5, %xmm0, %xmm1, %xmm0 Sorry for the inconvenience. In the meantime, -mno-stv can be used as a workaround. The bootstrap and regression testing should finish in a little while.
