From 0001229827afcc7d5e168b2711dab3a26ab5df12 Mon Sep 17 00:00:00 2001
From: Stefan Lankes <slankes@eonerc.rwth-aachen.de>
Date: Sat, 10 Sep 2016 16:55:33 +0200
Subject: [PATCH] save /restore only MXCSR & the x87 status word

- reduce overhead for a "user-level" context switch
---
 hermit/arch/x86/include/asm/stddef.h      |  2 ++
 hermit/arch/x86/include/asm/tasks_types.h | 16 ++++++++++++++++
 hermit/arch/x86/kernel/entry.asm          | 17 ++++++++---------
 hermit/include/hermit/syscall.h           |  4 +++-
 hermit/usr/newlib                         |  2 +-
 5 files changed, 30 insertions(+), 11 deletions(-)

diff --git a/hermit/arch/x86/include/asm/stddef.h b/hermit/arch/x86/include/asm/stddef.h
index c40070bde..9f0eb045d 100644
--- a/hermit/arch/x86/include/asm/stddef.h
+++ b/hermit/arch/x86/include/asm/stddef.h
@@ -144,6 +144,8 @@ typedef struct mregs {
 	uint64_t rsp;
 	/// RIP
 	uint64_t rip;
+	/// MSXCSR
+	uint32_t mxcsr;
 } mregs_t;
 
 /// This defines what the stack looks like after the task context is saved
diff --git a/hermit/arch/x86/include/asm/tasks_types.h b/hermit/arch/x86/include/asm/tasks_types.h
index 45f44882d..172f08a4a 100644
--- a/hermit/arch/x86/include/asm/tasks_types.h
+++ b/hermit/arch/x86/include/asm/tasks_types.h
@@ -108,6 +108,22 @@ union fpu_state {
 	xsave_t xsave;
 };
 
+typedef struct {
+	uint16_t control_word;
+	uint16_t unused1;
+	uint16_t status_word;
+	uint16_t unused2;
+	uint16_t tags;
+	uint16_t unused3;
+	uint32_t eip;
+	uint16_t cs_selector;
+	uint32_t opcode:11;
+	uint32_t unused4:5;
+	uint32_t data_offset;
+	uint16_t data_selector;
+	uint16_t unused5;
+} fenv_t;
+
 typedef void (*handle_fpu_state)(union fpu_state* state);
 
 extern handle_fpu_state save_fpu_state;
diff --git a/hermit/arch/x86/kernel/entry.asm b/hermit/arch/x86/kernel/entry.asm
index 150f3f81c..9bbffa6b4 100644
--- a/hermit/arch/x86/kernel/entry.asm
+++ b/hermit/arch/x86/kernel/entry.asm
@@ -493,7 +493,6 @@ isrsyscall:
     o64 sysret
 %endif
 
-extern save_fpu_state
 global getcontext
 align 64
 getcontext:
@@ -515,23 +514,23 @@ getcontext:
     mov QWORD [rdi + 0x60], rax
     mov rax, QWORD [rsp]
     mov QWORD [rdi + 0x68], rax
-    ; restore FPU state
-    add rdi, 0x80
-    call [save_fpu_state]
-    sub rdi, 0x80
+    ; save FPU state
+    fnstenv [rdi + 0x74]
+    fldenv [rdi + 0x74]
+    lea rax, [rdi + 0x70]
+    stmxcsr [rax]
     xor rax, rax
     sti
     ret
 
-extern restore_fpu_state
 global setcontext
 align 64
 setcontext:
     cli
     ; restore FPU state
-    add rdi, 0x80
-    call [restore_fpu_state]
-    sub rdi, 0x80
+    fldenv [rdi + 0x74]
+    lea rax, [rdi + 0x70]
+    ldmxcsr [rax]
     ; restore general purpose registers
     mov r15, QWORD [rdi + 0x00]
     mov r14, QWORD [rdi + 0x08]
diff --git a/hermit/include/hermit/syscall.h b/hermit/include/hermit/syscall.h
index a7da385fb..5d7da4c50 100644
--- a/hermit/include/hermit/syscall.h
+++ b/hermit/include/hermit/syscall.h
@@ -95,13 +95,15 @@ int sys_signal(signal_handler_t handler);
 
 typedef struct ucontext {
 	mregs_t		uc_mregs;
-	union fpu_state	uc_fpu;
+	fenv_t		uc_fenv;
 	struct ucontext	*uc_link;
 	stack_t		uc_stack;
 } ucontext_t;
 
 void makecontext(ucontext_t *ucp, void (*func)(), int argc, ...);
 int swapcontext(ucontext_t *oucp, const ucontext_t *ucp);
+int getcontext(ucontext_t *ucp);
+int setcontext(ucontext_t *ucp);
 
 #define __NR_exit 		0
 #define __NR_write		1
diff --git a/hermit/usr/newlib b/hermit/usr/newlib
index ba4dae88b..30a3e67a8 160000
--- a/hermit/usr/newlib
+++ b/hermit/usr/newlib
@@ -1 +1 @@
-Subproject commit ba4dae88bdc2fda079fc96cf70e7b238282229ca
+Subproject commit 30a3e67a8641c3a6884ad3d18e1fc8752652e89c