Remove npxgetregs(), npxsetregs(), fpugetregs() and fpusetregs()

functions, they are unused. Remove 'user' from npxgetuserregs()
etc. names.

For {npx,fpu}{get,set}regs(), always use pcb->pcb_user_save for FPU
context storage. This eliminates the need for ugly copying with
overwrite of the newly added and reserved fields in ucontext on i386
to satisfy alignment requirements for fpusave() and fpurstor().

pc98 version was copied from i386.

Suggested and reviewed by:	bde
Tested by:    pho (i386 and amd64)
MFC after:    1 week

9 files changed, 115 insertions, 291 deletions

diff --git a/sys/amd64/amd64/fpu.c b/sys/amd64/amd64/fpu.c
index 4b1583a..482b5da 100644
--- a/sys/amd64/amd64/fpu.c
+++ b/sys/amd64/amd64/fpu.c
@@ -426,9 +426,7 @@ fpudna(void)
 		fxrstor(&fpu_initialstate);
 		if (pcb->pcb_initial_fpucw != __INITIAL_FPUCW__)
 			fldcw(pcb->pcb_initial_fpucw);
-		pcb->pcb_flags |= PCB_FPUINITDONE;
-		if (PCB_USER_FPU(pcb))
-			pcb->pcb_flags |= PCB_USERFPUINITDONE;
+		fpuuserinited(curthread);
 	} else
 		fxrstor(pcb->pcb_save);
 	critical_exit();
@@ -448,60 +446,50 @@ fpudrop()
 }
 
 /*
- * Get the state of the FPU without dropping ownership (if possible).
- * It returns the FPU ownership status.
+ * Get the user state of the FPU into pcb->pcb_user_save without
+ * dropping ownership (if possible).  It returns the FPU ownership
+ * status.
  */
 int
-fpugetuserregs(struct thread *td, struct savefpu *addr)
+fpugetregs(struct thread *td)
 {
 	struct pcb *pcb;
 
 	pcb = td->td_pcb;
 	if ((pcb->pcb_flags & PCB_USERFPUINITDONE) == 0) {
-		bcopy(&fpu_initialstate, addr, sizeof(fpu_initialstate));
-		addr->sv_env.en_cw = pcb->pcb_initial_fpucw;
-		return (_MC_FPOWNED_NONE);
+		bcopy(&fpu_initialstate, &pcb->pcb_user_save,
+		    sizeof(fpu_initialstate));
+		pcb->pcb_user_save.sv_env.en_cw = pcb->pcb_initial_fpucw;
+		fpuuserinited(td);
+		return (_MC_FPOWNED_PCB);
 	}
 	critical_enter();
 	if (td == PCPU_GET(fpcurthread) && PCB_USER_FPU(pcb)) {
-		fxsave(addr);
+		fxsave(&pcb->pcb_user_save);
 		critical_exit();
 		return (_MC_FPOWNED_FPU);
 	} else {
 		critical_exit();
-		bcopy(&pcb->pcb_user_save, addr, sizeof(*addr));
 		return (_MC_FPOWNED_PCB);
 	}
 }
 
-int
-fpugetregs(struct thread *td, struct savefpu *addr)
+void
+fpuuserinited(struct thread *td)
 {
 	struct pcb *pcb;
 
 	pcb = td->td_pcb;
-	if ((pcb->pcb_flags & PCB_FPUINITDONE) == 0) {
-		bcopy(&fpu_initialstate, addr, sizeof(fpu_initialstate));
-		addr->sv_env.en_cw = pcb->pcb_initial_fpucw;
-		return (_MC_FPOWNED_NONE);
-	}
-	critical_enter();
-	if (td == PCPU_GET(fpcurthread)) {
-		fxsave(addr);
-		critical_exit();
-		return (_MC_FPOWNED_FPU);
-	} else {
-		critical_exit();
-		bcopy(pcb->pcb_save, addr, sizeof(*addr));
-		return (_MC_FPOWNED_PCB);
-	}
+	if (PCB_USER_FPU(pcb))
+		pcb->pcb_flags |= PCB_FPUINITDONE;
+	pcb->pcb_flags |= PCB_USERFPUINITDONE;
 }
 
 /*
  * Set the state of the FPU.
  */
 void
-fpusetuserregs(struct thread *td, struct savefpu *addr)
+fpusetregs(struct thread *td, struct savefpu *addr)
 {
 	struct pcb *pcb;
 
@@ -514,29 +502,8 @@ fpusetuserregs(struct thread *td, struct savefpu *addr)
 	} else {
 		critical_exit();
 		bcopy(addr, &td->td_pcb->pcb_user_save, sizeof(*addr));
-		if (PCB_USER_FPU(pcb))
-			pcb->pcb_flags |= PCB_FPUINITDONE;
-		pcb->pcb_flags |= PCB_USERFPUINITDONE;
-	}
-}
-
-void
-fpusetregs(struct thread *td, struct savefpu *addr)
-{
-	struct pcb *pcb;
-
-	pcb = td->td_pcb;
-	critical_enter();
-	if (td == PCPU_GET(fpcurthread)) {
-		fxrstor(addr);
-		critical_exit();
-	} else {
-		critical_exit();
-		bcopy(addr, td->td_pcb->pcb_save, sizeof(*addr));
+		fpuuserinited(td);
 	}
-	if (PCB_USER_FPU(pcb))
-		pcb->pcb_flags |= PCB_USERFPUINITDONE;
-	pcb->pcb_flags |= PCB_FPUINITDONE;
 }
 
 /*
diff --git a/sys/amd64/amd64/machdep.c b/sys/amd64/amd64/machdep.c
index 738427f..194cf71 100644
--- a/sys/amd64/amd64/machdep.c
+++ b/sys/amd64/amd64/machdep.c
@@ -1975,6 +1975,8 @@ int
 fill_fpregs(struct thread *td, struct fpreg *fpregs)
 {
 
+	KASSERT(TD_IS_SUSPENDED(td), ("not suspended thread %p", td));
+	fpugetregs(td);
 	fill_fpregs_xmm(&td->td_pcb->pcb_user_save, fpregs);
 	return (0);
 }
@@ -1985,6 +1987,7 @@ set_fpregs(struct thread *td, struct fpreg *fpregs)
 {
 
 	set_fpregs_xmm(fpregs, &td->td_pcb->pcb_user_save);
+	fpuuserinited(td);
 	return (0);
 }
 
@@ -2099,8 +2102,9 @@ static void
 get_fpcontext(struct thread *td, mcontext_t *mcp)
 {
 
-	mcp->mc_ownedfp = fpugetuserregs(td,
-	    (struct savefpu *)&mcp->mc_fpstate);
+	mcp->mc_ownedfp = fpugetregs(td);
+	bcopy(&td->td_pcb->pcb_user_save, &mcp->mc_fpstate,
+	    sizeof(mcp->mc_fpstate));
 	mcp->mc_fpformat = fpuformat();
 }
 
@@ -2120,7 +2124,7 @@ set_fpcontext(struct thread *td, const mcontext_t *mcp)
 	    mcp->mc_ownedfp == _MC_FPOWNED_PCB) {
 		fpstate = (struct savefpu *)&mcp->mc_fpstate;
 		fpstate->sv_env.en_mxcsr &= cpu_mxcsr_mask;
-		fpusetuserregs(td, fpstate);
+		fpusetregs(td, fpstate);
 	} else
 		return (EINVAL);
 	return (0);
diff --git a/sys/amd64/ia32/ia32_reg.c b/sys/amd64/ia32/ia32_reg.c
index 30fcffb..da6eb0f 100644
--- a/sys/amd64/ia32/ia32_reg.c
+++ b/sys/amd64/ia32/ia32_reg.c
@@ -145,13 +145,18 @@ set_regs32(struct thread *td, struct reg32 *regs)
 int
 fill_fpregs32(struct thread *td, struct fpreg32 *regs)
 {
-	struct save87 *sv_87 = (struct save87 *)regs;
-	struct env87 *penv_87 = &sv_87->sv_env;
-	struct savefpu *sv_fpu = &td->td_pcb->pcb_user_save;
-	struct envxmm *penv_xmm = &sv_fpu->sv_env;
+	struct savefpu *sv_fpu;
+	struct save87 *sv_87;
+	struct env87 *penv_87;
+	struct envxmm *penv_xmm;
 	int i;
 
 	bzero(regs, sizeof(*regs));
+	sv_87 = (struct save87 *)regs;
+	penv_87 = &sv_87->sv_env;
+	fpugetregs(td);
+	sv_fpu = &td->td_pcb->pcb_user_save;
+	penv_xmm = &sv_fpu->sv_env;
 	
 	/* FPU control/status */
 	penv_87->en_cw = penv_xmm->en_cw;
@@ -200,6 +205,7 @@ set_fpregs32(struct thread *td, struct fpreg32 *regs)
 		sv_fpu->sv_fp[i].fp_acc = sv_87->sv_ac[i];
 	for (i = 8; i < 16; ++i)
 		bzero(&sv_fpu->sv_fp[i].fp_acc, sizeof(sv_fpu->sv_fp[i].fp_acc));
+	fpuuserinited(td);
 
 	return (0);
 }
diff --git a/sys/amd64/ia32/ia32_signal.c b/sys/amd64/ia32/ia32_signal.c
index d85a3bb..aefe9f0 100644
--- a/sys/amd64/ia32/ia32_signal.c
+++ b/sys/amd64/ia32/ia32_signal.c
@@ -99,8 +99,9 @@ ia32_get_fpcontext(struct thread *td, struct ia32_mcontext *mcp)
 	 * 64bit instruction and data pointers. Ignore the difference
 	 * for now, it should be irrelevant for most applications.
 	 */
-	mcp->mc_ownedfp = fpugetuserregs(td,
-	    (struct savefpu *)&mcp->mc_fpstate);
+	mcp->mc_ownedfp = fpugetregs(td);
+	bcopy(&td->td_pcb->pcb_user_save, &mcp->mc_fpstate,
+	    sizeof(mcp->mc_fpstate));
 	mcp->mc_fpformat = fpuformat();
 }
 
@@ -117,7 +118,7 @@ ia32_set_fpcontext(struct thread *td, const struct ia32_mcontext *mcp)
 		fpstate_drop(td);
 	else if (mcp->mc_ownedfp == _MC_FPOWNED_FPU ||
 	    mcp->mc_ownedfp == _MC_FPOWNED_PCB) {
-		fpusetuserregs(td, (struct savefpu *)&mcp->mc_fpstate);
+		fpusetregs(td, (struct savefpu *)&mcp->mc_fpstate);
 	} else
 		return (EINVAL);
 	return (0);
diff --git a/sys/amd64/include/fpu.h b/sys/amd64/include/fpu.h
index ca0ac8f..50b3819 100644
--- a/sys/amd64/include/fpu.h
+++ b/sys/amd64/include/fpu.h
@@ -112,12 +112,11 @@ void	fpudna(void);
 void	fpudrop(void);
 void	fpuexit(struct thread *td);
 int	fpuformat(void);
-int	fpugetregs(struct thread *td, struct savefpu *addr);
-int	fpugetuserregs(struct thread *td, struct savefpu *addr);
+int	fpugetregs(struct thread *td);
 void	fpuinit(void);
 void	fpusetregs(struct thread *td, struct savefpu *addr);
-void	fpusetuserregs(struct thread *td, struct savefpu *addr);
 int	fputrap(void);
+void	fpuuserinited(struct thread *td);
 int	fpu_kern_enter(struct thread *td, struct fpu_kern_ctx *ctx,
 	    u_int flags);
 int	fpu_kern_leave(struct thread *td, struct fpu_kern_ctx *ctx);
diff --git a/sys/i386/i386/machdep.c b/sys/i386/i386/machdep.c
index f89de32..935a8d20 100644
--- a/sys/i386/i386/machdep.c
+++ b/sys/i386/i386/machdep.c
@@ -633,13 +633,6 @@ sendsig(sig_t catcher, ksiginfo_t *ksi, sigset_t *mask)
 	sf.sf_uc.uc_mcontext.mc_gs = rgs();
 	bcopy(regs, &sf.sf_uc.uc_mcontext.mc_fs, sizeof(*regs));
 	sf.sf_uc.uc_mcontext.mc_len = sizeof(sf.sf_uc.uc_mcontext); /* magic */
-
-	/*
-	 * The get_fpcontext() call must be placed before assignments
-	 * to mc_fsbase and mc_gsbase due to the alignment-override
-	 * code in get_fpcontext() that possibly clobbers 12 bytes of
-	 * mcontext after mc_fpstate.
-	 */
 	get_fpcontext(td, &sf.sf_uc.uc_mcontext);
 	fpstate_drop(td);
 	/*
@@ -3209,28 +3202,33 @@ set_fpregs_xmm(sv_87, sv_xmm)
 int
 fill_fpregs(struct thread *td, struct fpreg *fpregs)
 {
+
+	KASSERT(TD_IS_SUSPENDED(td), ("not suspended thread %p", td));
+	npxgetregs(td);
 #ifdef CPU_ENABLE_SSE
-	if (cpu_fxsr) {
+	if (cpu_fxsr)
 		fill_fpregs_xmm(&td->td_pcb->pcb_user_save.sv_xmm,
-						(struct save87 *)fpregs);
-		return (0);
-	}
+		    (struct save87 *)fpregs);
+	else
 #endif /* CPU_ENABLE_SSE */
-	bcopy(&td->td_pcb->pcb_user_save.sv_87, fpregs, sizeof *fpregs);
+		bcopy(&td->td_pcb->pcb_user_save.sv_87, fpregs,
+		    sizeof(*fpregs));
 	return (0);
 }
 
 int
 set_fpregs(struct thread *td, struct fpreg *fpregs)
 {
+
 #ifdef CPU_ENABLE_SSE
-	if (cpu_fxsr) {
+	if (cpu_fxsr)
 		set_fpregs_xmm((struct save87 *)fpregs,
 		    &td->td_pcb->pcb_user_save.sv_xmm);
-		return (0);
-	}
+	else
 #endif /* CPU_ENABLE_SSE */
-	bcopy(fpregs, &td->td_pcb->pcb_user_save.sv_87, sizeof *fpregs);
+		bcopy(fpregs, &td->td_pcb->pcb_user_save.sv_87,
+		    sizeof(*fpregs));
+	npxuserinited(td);
 	return (0);
 }
 
@@ -3272,13 +3270,6 @@ get_mcontext(struct thread *td, mcontext_t *mcp, int flags)
 	mcp->mc_esp = tp->tf_esp;
 	mcp->mc_ss = tp->tf_ss;
 	mcp->mc_len = sizeof(*mcp);
-
-	/*
-	 * The get_fpcontext() call must be placed before assignments
-	 * to mc_fsbase and mc_gsbase due to the alignment-override
-	 * code in get_fpcontext() that possibly clobbers 12 bytes of
-	 * mcontext after mc_fpstate.
-	 */
 	get_fpcontext(td, mcp);
 	sdp = &td->td_pcb->pcb_fsd;
 	mcp->mc_fsbase = sdp->sd_hibase << 24 | sdp->sd_lobase;
@@ -3329,39 +3320,14 @@ set_mcontext(struct thread *td, const mcontext_t *mcp)
 static void
 get_fpcontext(struct thread *td, mcontext_t *mcp)
 {
+
 #ifndef DEV_NPX
 	mcp->mc_fpformat = _MC_FPFMT_NODEV;
 	mcp->mc_ownedfp = _MC_FPOWNED_NONE;
 #else
-	union savefpu *addr;
-
-	/*
-	 * XXX mc_fpstate might be misaligned, since its declaration is not
-	 * unportabilized using __attribute__((aligned(16))) like the
-	 * declaration of struct savemm, and anyway, alignment doesn't work
-	 * for auto variables since we don't use gcc's pessimal stack
-	 * alignment.  Work around this by abusing the spare fields after
-	 * mcp->mc_fpstate.
-	 *
-	 * XXX unpessimize most cases by only aligning when fxsave might be
-	 * called, although this requires knowing too much about
-	 * npxgetuserregs()'s internals.
-	 */
-	addr = (union savefpu *)&mcp->mc_fpstate;
-	if (td == PCPU_GET(fpcurthread) &&
-#ifdef CPU_ENABLE_SSE
-	    cpu_fxsr &&
-#endif
-	    ((uintptr_t)(void *)addr & 0xF)) {
-		do
-			addr = (void *)((char *)addr + 4);
-		while ((uintptr_t)(void *)addr & 0xF);
-	}
-	mcp->mc_ownedfp = npxgetuserregs(td, addr);
-	if (addr != (union savefpu *)&mcp->mc_fpstate) {
-		bcopy(addr, &mcp->mc_fpstate, sizeof(mcp->mc_fpstate));
-		bzero(&mcp->mc_spare2, sizeof(mcp->mc_spare2));
-	}
+	mcp->mc_ownedfp = npxgetregs(td);
+	bcopy(&td->td_pcb->pcb_user_save, &mcp->mc_fpstate,
+	    sizeof(mcp->mc_fpstate));
 	mcp->mc_fpformat = npxformat();
 #endif
 }
@@ -3369,7 +3335,6 @@ get_fpcontext(struct thread *td, mcontext_t *mcp)
 static int
 set_fpcontext(struct thread *td, const mcontext_t *mcp)
 {
-	union savefpu *addr;
 
 	if (mcp->mc_fpformat == _MC_FPFMT_NODEV)
 		return (0);
@@ -3381,30 +3346,14 @@ set_fpcontext(struct thread *td, const mcontext_t *mcp)
 		fpstate_drop(td);
 	else if (mcp->mc_ownedfp == _MC_FPOWNED_FPU ||
 	    mcp->mc_ownedfp == _MC_FPOWNED_PCB) {
-		/* XXX align as above. */
-		addr = (union savefpu *)&mcp->mc_fpstate;
-		if (td == PCPU_GET(fpcurthread) &&
-#ifdef CPU_ENABLE_SSE
-		    cpu_fxsr &&
-#endif
-		    ((uintptr_t)(void *)addr & 0xF)) {
-			do
-				addr = (void *)((char *)addr + 4);
-			while ((uintptr_t)(void *)addr & 0xF);
-			bcopy(&mcp->mc_fpstate, addr, sizeof(mcp->mc_fpstate));
-		}
 #ifdef DEV_NPX
 #ifdef CPU_ENABLE_SSE
 		if (cpu_fxsr)
-			addr->sv_xmm.sv_env.en_mxcsr &= cpu_mxcsr_mask;
+			((union savefpu *)&mcp->mc_fpstate)->sv_xmm.sv_env.
+			    en_mxcsr &= cpu_mxcsr_mask;
 #endif
-		npxsetuserregs(td, addr);
+		npxsetregs(td, (union savefpu *)&mcp->mc_fpstate);
 #endif
-		/*
-		 * Don't bother putting things back where they were in the
-		 * misaligned case, since we know that the caller won't use
-		 * them again.
-		 */
 	} else
 		return (EINVAL);
 	return (0);
@@ -3422,12 +3371,12 @@ fpstate_drop(struct thread *td)
 #endif
 	/*
 	 * XXX force a full drop of the npx.  The above only drops it if we
-	 * owned it.  npxgetuserregs() has the same bug in the !cpu_fxsr case.
+	 * owned it.  npxgetregs() has the same bug in the !cpu_fxsr case.
 	 *
-	 * XXX I don't much like npxgetuserregs()'s semantics of doing a full
+	 * XXX I don't much like npxgetregs()'s semantics of doing a full
 	 * drop.  Dropping only to the pcb matches fnsave's behaviour.
 	 * We only need to drop to !PCB_INITDONE in sendsig().  But
-	 * sendsig() is the only caller of npxgetuserregs()... perhaps we just
+	 * sendsig() is the only caller of npxgetregs()... perhaps we just
 	 * have too many layers.
 	 */
 	curthread->td_pcb->pcb_flags &= ~(PCB_NPXINITDONE |
diff --git a/sys/i386/include/npx.h b/sys/i386/include/npx.h
index 71073a6..6b56bb4 100644
--- a/sys/i386/include/npx.h
+++ b/sys/i386/include/npx.h
@@ -151,13 +151,12 @@ int	npxdna(void);
 void	npxdrop(void);
 void	npxexit(struct thread *td);
 int	npxformat(void);
-int	npxgetregs(struct thread *td, union savefpu *addr);
-int	npxgetuserregs(struct thread *td, union savefpu *addr);
+int	npxgetregs(struct thread *td);
 void	npxinit(void);
 void	npxsave(union savefpu *addr);
 void	npxsetregs(struct thread *td, union savefpu *addr);
-void	npxsetuserregs(struct thread *td, union savefpu *addr);
 int	npxtrap(void);
+void	npxuserinited(struct thread *);
 int	fpu_kern_enter(struct thread *td, struct fpu_kern_ctx *ctx,
 	    u_int flags);
 int	fpu_kern_leave(struct thread *td, struct fpu_kern_ctx *ctx);
diff --git a/sys/i386/isa/npx.c b/sys/i386/isa/npx.c
index 8d37562..9ec5d25 100644
--- a/sys/i386/isa/npx.c
+++ b/sys/i386/isa/npx.c
@@ -684,9 +684,7 @@ npxdna(void)
 		fpurstor(&npx_initialstate);
 		if (pcb->pcb_initial_npxcw != __INITIAL_NPXCW__)
 			fldcw(pcb->pcb_initial_npxcw);
-		pcb->pcb_flags |= PCB_NPXINITDONE;
-		if (PCB_USER_FPU(pcb))
-			pcb->pcb_flags |= PCB_NPXUSERINITDONE;
+		npxuserinited(curthread);
 	} else {
 		/*
 		 * The following fpurstor() may cause an IRQ13 when the
@@ -767,11 +765,12 @@ npxdrop()
 }
 
 /*
- * Get the state of the FPU without dropping ownership (if possible).
- * It returns the FPU ownership status.
+ * Get the user state of the FPU into pcb->pcb_user_save without
+ * dropping ownership (if possible).  It returns the FPU ownership
+ * status.
  */
 int
-npxgetregs(struct thread *td, union savefpu *addr)
+npxgetregs(struct thread *td)
 {
 	struct pcb *pcb;
 
@@ -780,48 +779,15 @@ npxgetregs(struct thread *td, union savefpu *addr)
 
 	pcb = td->td_pcb;
 	if ((pcb->pcb_flags & PCB_NPXINITDONE) == 0) {
-		bcopy(&npx_initialstate, addr, sizeof(npx_initialstate));
-		SET_FPU_CW(addr, pcb->pcb_initial_npxcw);
-		return (_MC_FPOWNED_NONE);
-	}
-	critical_enter();
-	if (td == PCPU_GET(fpcurthread)) {
-		fpusave(addr);
-#ifdef CPU_ENABLE_SSE
-		if (!cpu_fxsr)
-#endif
-			/*
-			 * fnsave initializes the FPU and destroys whatever
-			 * context it contains.  Make sure the FPU owner
-			 * starts with a clean state next time.
-			 */
-			npxdrop();
-		critical_exit();
-		return (_MC_FPOWNED_FPU);
-	} else {
-		critical_exit();
-		bcopy(pcb->pcb_save, addr, sizeof(*addr));
+		bcopy(&npx_initialstate, &pcb->pcb_user_save,
+		    sizeof(npx_initialstate));
+		SET_FPU_CW(&pcb->pcb_user_save, pcb->pcb_initial_npxcw);
+		npxuserinited(td);
 		return (_MC_FPOWNED_PCB);
 	}
-}
-
-int
-npxgetuserregs(struct thread *td, union savefpu *addr)
-{
-	struct pcb *pcb;
-
-	if (!hw_float)
-		return (_MC_FPOWNED_NONE);
-
-	pcb = td->td_pcb;
-	if ((pcb->pcb_flags & PCB_NPXUSERINITDONE) == 0) {
-		bcopy(&npx_initialstate, addr, sizeof(npx_initialstate));
-		SET_FPU_CW(addr, pcb->pcb_initial_npxcw);
-		return (_MC_FPOWNED_NONE);
-	}
 	critical_enter();
-	if (td == PCPU_GET(fpcurthread) && PCB_USER_FPU(pcb)) {
-		fpusave(addr);
+	if (td == PCPU_GET(fpcurthread)) {
+		fpusave(&pcb->pcb_user_save);
 #ifdef CPU_ENABLE_SSE
 		if (!cpu_fxsr)
 #endif
@@ -835,42 +801,24 @@ npxgetuserregs(struct thread *td, union savefpu *addr)
 		return (_MC_FPOWNED_FPU);
 	} else {
 		critical_exit();
-		bcopy(&pcb->pcb_user_save, addr, sizeof(*addr));
 		return (_MC_FPOWNED_PCB);
 	}
 }
 
-/*
- * Set the state of the FPU.
- */
 void
-npxsetregs(struct thread *td, union savefpu *addr)
+npxuserinited(struct thread *td)
 {
 	struct pcb *pcb;
 
-	if (!hw_float)
-		return;
-
 	pcb = td->td_pcb;
-	critical_enter();
-	if (td == PCPU_GET(fpcurthread)) {
-#ifdef CPU_ENABLE_SSE
-		if (!cpu_fxsr)
-#endif
-			fnclex();	/* As in npxdrop(). */
-		fpurstor(addr);
-		critical_exit();
-	} else {
-		critical_exit();
-		bcopy(addr, pcb->pcb_save, sizeof(*addr));
-	}
 	if (PCB_USER_FPU(pcb))
-		pcb->pcb_flags |= PCB_NPXUSERINITDONE;
-	pcb->pcb_flags |= PCB_NPXINITDONE;
+		pcb->pcb_flags |= PCB_NPXINITDONE;
+	pcb->pcb_flags |= PCB_NPXUSERINITDONE;
 }
 
+
 void
-npxsetuserregs(struct thread *td, union savefpu *addr)
+npxsetregs(struct thread *td, union savefpu *addr)
 {
 	struct pcb *pcb;
 
@@ -884,15 +832,17 @@ npxsetuserregs(struct thread *td, union savefpu *addr)
 		if (!cpu_fxsr)
 #endif
 			fnclex();	/* As in npxdrop(). */
-		fpurstor(addr);
+		if (((uintptr_t)addr & 0xf) != 0) {
+			bcopy(addr, &pcb->pcb_user_save, sizeof(*addr));
+			fpurstor(&pcb->pcb_user_save);
+		} else
+			fpurstor(addr);
 		critical_exit();
 		pcb->pcb_flags |= PCB_NPXUSERINITDONE | PCB_NPXINITDONE;
 	} else {
 		critical_exit();
 		bcopy(addr, &pcb->pcb_user_save, sizeof(*addr));
-		if (PCB_USER_FPU(pcb))
-			pcb->pcb_flags |= PCB_NPXINITDONE;
-		pcb->pcb_flags |= PCB_NPXUSERINITDONE;
+		npxuserinited(td);
 	}
 }
 
diff --git a/sys/pc98/pc98/machdep.c b/sys/pc98/pc98/machdep.c
index 942aa45..03933a6 100644
--- a/sys/pc98/pc98/machdep.c
+++ b/sys/pc98/pc98/machdep.c
@@ -568,13 +568,6 @@ sendsig(sig_t catcher, ksiginfo_t *ksi, sigset_t *mask)
 	sf.sf_uc.uc_mcontext.mc_gs = rgs();
 	bcopy(regs, &sf.sf_uc.uc_mcontext.mc_fs, sizeof(*regs));
 	sf.sf_uc.uc_mcontext.mc_len = sizeof(sf.sf_uc.uc_mcontext); /* magic */
-
-	/*
-	 * The get_fpcontext() call must be placed before assignments
-	 * to mc_fsbase and mc_gsbase due to the alignment-override
-	 * code in get_fpcontext() that possibly clobbers 12 bytes of
-	 * mcontext after mc_fpstate.
-	 */
 	get_fpcontext(td, &sf.sf_uc.uc_mcontext);
 	fpstate_drop(td);
 	/*
@@ -2553,28 +2546,33 @@ set_fpregs_xmm(sv_87, sv_xmm)
 int
 fill_fpregs(struct thread *td, struct fpreg *fpregs)
 {
+
+	KASSERT(TD_IS_SUSPENDED(td), ("not suspended thread %p", td));
+	npxgetregs(td);
 #ifdef CPU_ENABLE_SSE
-	if (cpu_fxsr) {
-		fill_fpregs_xmm(&td->td_pcb->pcb_save->sv_xmm,
-						(struct save87 *)fpregs);
-		return (0);
-	}
+	if (cpu_fxsr)
+		fill_fpregs_xmm(&td->td_pcb->pcb_user_save.sv_xmm,
+		    (struct save87 *)fpregs);
+	else
 #endif /* CPU_ENABLE_SSE */
-	bcopy(&td->td_pcb->pcb_save->sv_87, fpregs, sizeof *fpregs);
+		bcopy(&td->td_pcb->pcb_user_save.sv_87, fpregs,
+		    sizeof(*fpregs));
 	return (0);
 }
 
 int
 set_fpregs(struct thread *td, struct fpreg *fpregs)
 {
+
 #ifdef CPU_ENABLE_SSE
-	if (cpu_fxsr) {
+	if (cpu_fxsr)
 		set_fpregs_xmm((struct save87 *)fpregs,
-					   &td->td_pcb->pcb_save->sv_xmm);
-		return (0);
-	}
+		    &td->td_pcb->pcb_user_save.sv_xmm);
+	else
 #endif /* CPU_ENABLE_SSE */
-	bcopy(fpregs, &td->td_pcb->pcb_save->sv_87, sizeof *fpregs);
+		bcopy(fpregs, &td->td_pcb->pcb_user_save.sv_87,
+		    sizeof(*fpregs));
+	npxuserinited(td);
 	return (0);
 }
 
@@ -2616,13 +2614,6 @@ get_mcontext(struct thread *td, mcontext_t *mcp, int flags)
 	mcp->mc_esp = tp->tf_esp;
 	mcp->mc_ss = tp->tf_ss;
 	mcp->mc_len = sizeof(*mcp);
-
-	/*
-	 * The get_fpcontext() call must be placed before assignments
-	 * to mc_fsbase and mc_gsbase due to the alignment-override
-	 * code in get_fpcontext() that possibly clobbers 12 bytes of
-	 * mcontext after mc_fpstate.
-	 */
 	get_fpcontext(td, mcp);
 	sdp = &td->td_pcb->pcb_fsd;
 	mcp->mc_fsbase = sdp->sd_hibase << 24 | sdp->sd_lobase;
@@ -2673,39 +2664,14 @@ set_mcontext(struct thread *td, const mcontext_t *mcp)
 static void
 get_fpcontext(struct thread *td, mcontext_t *mcp)
 {
+
 #ifndef DEV_NPX
 	mcp->mc_fpformat = _MC_FPFMT_NODEV;
 	mcp->mc_ownedfp = _MC_FPOWNED_NONE;
 #else
-	union savefpu *addr;
-
-	/*
-	 * XXX mc_fpstate might be misaligned, since its declaration is not
-	 * unportabilized using __attribute__((aligned(16))) like the
-	 * declaration of struct savemm, and anyway, alignment doesn't work
-	 * for auto variables since we don't use gcc's pessimal stack
-	 * alignment.  Work around this by abusing the spare fields after
-	 * mcp->mc_fpstate.
-	 *
-	 * XXX unpessimize most cases by only aligning when fxsave might be
-	 * called, although this requires knowing too much about
-	 * npxgetuserregs()'s internals.
-	 */
-	addr = (union savefpu *)&mcp->mc_fpstate;
-	if (td == PCPU_GET(fpcurthread) &&
-#ifdef CPU_ENABLE_SSE
-	    cpu_fxsr &&
-#endif
-	    ((uintptr_t)(void *)addr & 0xF)) {
-		do
-			addr = (void *)((char *)addr + 4);
-		while ((uintptr_t)(void *)addr & 0xF);
-	}
-	mcp->mc_ownedfp = npxgetuserregs(td, addr);
-	if (addr != (union savefpu *)&mcp->mc_fpstate) {
-		bcopy(addr, &mcp->mc_fpstate, sizeof(mcp->mc_fpstate));
-		bzero(&mcp->mc_spare2, sizeof(mcp->mc_spare2));
-	}
+	mcp->mc_ownedfp = npxgetregs(td);
+	bcopy(&td->td_pcb->pcb_user_save, &mcp->mc_fpstate,
+	    sizeof(mcp->mc_fpstate));
 	mcp->mc_fpformat = npxformat();
 #endif
 }
@@ -2713,7 +2679,6 @@ get_fpcontext(struct thread *td, mcontext_t *mcp)
 static int
 set_fpcontext(struct thread *td, const mcontext_t *mcp)
 {
-	union savefpu *addr;
 
 	if (mcp->mc_fpformat == _MC_FPFMT_NODEV)
 		return (0);
@@ -2725,30 +2690,14 @@ set_fpcontext(struct thread *td, const mcontext_t *mcp)
 		fpstate_drop(td);
 	else if (mcp->mc_ownedfp == _MC_FPOWNED_FPU ||
 	    mcp->mc_ownedfp == _MC_FPOWNED_PCB) {
-		/* XXX align as above. */
-		addr = (union savefpu *)&mcp->mc_fpstate;
-		if (td == PCPU_GET(fpcurthread) &&
-#ifdef CPU_ENABLE_SSE
-		    cpu_fxsr &&
-#endif
-		    ((uintptr_t)(void *)addr & 0xF)) {
-			do
-				addr = (void *)((char *)addr + 4);
-			while ((uintptr_t)(void *)addr & 0xF);
-			bcopy(&mcp->mc_fpstate, addr, sizeof(mcp->mc_fpstate));
-		}
 #ifdef DEV_NPX
 #ifdef CPU_ENABLE_SSE
 		if (cpu_fxsr)
-			addr->sv_xmm.sv_env.en_mxcsr &= cpu_mxcsr_mask;
+			((union savefpu *)&mcp->mc_fpstate)->sv_xmm.sv_env.
+			    en_mxcsr &= cpu_mxcsr_mask;
 #endif
-		npxsetuserregs(td, addr);
+		npxsetregs(td, (union savefpu *)&mcp->mc_fpstate);
 #endif
-		/*
-		 * Don't bother putting things back where they were in the
-		 * misaligned case, since we know that the caller won't use
-		 * them again.
-		 */
 	} else
 		return (EINVAL);
 	return (0);
@@ -2765,12 +2714,12 @@ fpstate_drop(struct thread *td)
 #endif
 	/*
 	 * XXX force a full drop of the npx.  The above only drops it if we
-	 * owned it.  npxusergetregs() has the same bug in the !cpu_fxsr case.
+	 * owned it.  npxgetregs() has the same bug in the !cpu_fxsr case.
 	 *
-	 * XXX I don't much like npxgetuserregs()'s semantics of doing a full
+	 * XXX I don't much like npxgetregs()'s semantics of doing a full
 	 * drop.  Dropping only to the pcb matches fnsave's behaviour.
 	 * We only need to drop to !PCB_INITDONE in sendsig().  But
-	 * sendsig() is the only caller of npxgetuserregs()... perhaps we just
+	 * sendsig() is the only caller of npxgetregs()... perhaps we just
 	 * have too many layers.
 	 */
 	curthread->td_pcb->pcb_flags &= ~(PCB_NPXINITDONE |