@@ -1325,8 +1325,10 @@ static int kvm_put_xsave(X86CPU *cpu)
xsave->region[XSAVE_FTW_FOP] = (uint32_t)(env->fpop << 16) + twd;
memcpy(&xsave->region[XSAVE_CWD_RIP], &env->fpip, sizeof(env->fpip));
memcpy(&xsave->region[XSAVE_CWD_RDP], &env->fpdp, sizeof(env->fpdp));
- memcpy(&xsave->region[XSAVE_ST_SPACE], env->fpregs,
- sizeof env->fpregs);
+ for (i = 0; i < 8; i++) {
+ memcpy(&xsave_region[XSAVE_ST_SPACE + i * 4],
+ &env->fpregs[(env->fpstt + i) & 7], 16);
+ }
xsave->region[XSAVE_MXCSR] = env->mxcsr;
*(uint64_t *)&xsave->region[XSAVE_XSTATE_BV] = env->xstate_bv;
memcpy(&xsave->region[XSAVE_BNDREGS], env->bnd_regs,
@@ -1745,8 +1747,10 @@ static int kvm_get_xsave(X86CPU *cpu)
memcpy(&env->fpip, &xsave->region[XSAVE_CWD_RIP], sizeof(env->fpip));
memcpy(&env->fpdp, &xsave->region[XSAVE_CWD_RDP], sizeof(env->fpdp));
env->mxcsr = xsave->region[XSAVE_MXCSR];
- memcpy(env->fpregs, &xsave->region[XSAVE_ST_SPACE],
- sizeof env->fpregs);
+ for (i = 0; i < 8; i++) {
+ memcpy(&env->fpregs[(env->fpstt + i) & 7],
+ &xsave_region[XSAVE_ST_SPACE + i * 4], 16);
+ }
env->xstate_bv = *(uint64_t *)&xsave->region[XSAVE_XSTATE_BV];
memcpy(env->bnd_regs, &xsave->region[XSAVE_BNDREGS],
sizeof env->bnd_regs);
Many thanks for the reply. I'm sorry for the typo in the previous patch. Below comes a new one. Best regards, asia Stick to the right order of the FPU registers in xsave (as specified in the Intel manual.) Otherwise, instructions loading the FPU state don't set it up correctly. To set up FPU, software needs to provide a buffer of 80 bytes storing 8 FPU registers. They are organized in a stack. FPU assumes that the first field of the buffer is ST0, then ST1, and so on. QEMU maintains a circular buffer. When preparing these 80 bytes for KVM, QEMU just uses memcpy instead of copying the elements in a proper order. Signed-off-by: Asia Slowinska <asia@lastline.com> --- target-i386/kvm.c | 12 ++++++++---- 1 file changed, 8 insertions(+), 4 deletions(-)