https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112891
Bug ID: 112891 Summary: [10/11/12/13/14 Regression] Missing vzeroupper insert. Product: gcc Version: 14.0 Status: UNCONFIRMED Severity: normal Priority: P3 Component: target Assignee: unassigned at gcc dot gnu.org Reporter: liuhongt at gcc dot gnu.org Target Milestone: --- #include<math.h> void __attribute__((noinline)) bar (double* a) { a[0] = 1.0; a[1] = 2.0; } void __attribute__((noinline)) foo (double* __restrict a, double* b) { a[0] += b[0]; a[1] += b[1]; a[2] += b[2]; a[3] += b[3]; bar (b); } double foo1 (double* __restrict a, double* b) { foo (a, b); return exp (b[1]); } gcc -O3 -mavx2 Got bar(double*): vmovapd xmm0, XMMWORD PTR .LC0[rip] vmovupd XMMWORD PTR [rdi], xmm0 ret foo(double*, double*): mov rax, rdi vmovupd ymm0, YMMWORD PTR [rsi] mov rdi, rsi vaddpd ymm0, ymm0, YMMWORD PTR [rax] vmovupd YMMWORD PTR [rax], ymm0 jmp bar(double*) foo1(double*, double*): sub rsp, 8 call foo(double*, double*) vmovsd xmm0, QWORD PTR [rsi+8] add rsp, 8 jmp exp .LC0: .long 0 .long 1072693248 .long 0 .long 1073741824 In foo, 256-bit ymm are used, and the upper bits are dirty, but there's no vzeroupper inserted by exp which cause big avx->sse transition penalty.