Merge pull request #4352 from evincarofautumn/fix-xmm-scanning-mac-4.2mono-4.2.0-branch

Backport Mac 64 SIMD fixes to 4.2

11 files changed, 338 insertions, 78 deletions

diff --git a/mono/metadata/sgen-os-mach.c b/mono/metadata/sgen-os-mach.c
index eee4d837d3f..c57b06179c4 100644
--- a/mono/metadata/sgen-os-mach.c
+++ b/mono/metadata/sgen-os-mach.c
@@ -49,8 +49,8 @@ sgen_resume_thread (SgenThreadInfo *info)
 gboolean
 sgen_suspend_thread (SgenThreadInfo *info)
 {
-	mach_msg_type_number_t num_state;
-	thread_state_t state;
+	mach_msg_type_number_t num_state, num_fpstate;
+	thread_state_t state, fpstate;
 	kern_return_t ret;
 	ucontext_t ctx;
 	mcontext_t mctx;
@@ -58,17 +58,20 @@ sgen_suspend_thread (SgenThreadInfo *info)
 	gpointer stack_start;
 
 	state = (thread_state_t) alloca (mono_mach_arch_get_thread_state_size ());
+	fpstate = (thread_state_t) alloca (mono_mach_arch_get_thread_fpstate_size ());
 	mctx = (mcontext_t) alloca (mono_mach_arch_get_mcontext_size ());
 
 	ret = thread_suspend (info->client_info.info.native_handle);
 	if (ret != KERN_SUCCESS)
 		return FALSE;
 
-	ret = mono_mach_arch_get_thread_state (info->client_info.info.native_handle, state, &num_state);
+	do {
+		ret = mono_mach_arch_get_thread_states (info->client_info.info.native_handle, state, &num_state, fpstate, &num_fpstate);
+	} while (ret == KERN_ABORTED);
 	if (ret != KERN_SUCCESS)
 		return FALSE;
 
-	mono_mach_arch_thread_state_to_mcontext (state, mctx);
+	mono_mach_arch_thread_states_to_mcontext (state, fpstate, mctx);
 	ctx.uc_mcontext = mctx;
 
 	info->client_info.stopped_domain = mono_thread_info_tls_get (info, TLS_KEY_DOMAIN);
diff --git a/mono/mini/mini-darwin.c b/mono/mini/mini-darwin.c
index 9c1e46ed829..f07fc5ea01c 100644
--- a/mono/mini/mini-darwin.c
+++ b/mono/mini/mini-darwin.c
@@ -298,8 +298,8 @@ gboolean
 mono_thread_state_init_from_handle (MonoThreadUnwindState *tctx, MonoThreadInfo *info)
 {
 	kern_return_t ret;
-	mach_msg_type_number_t num_state;
-	thread_state_t state;
+	mach_msg_type_number_t num_state, num_fpstate;
+	thread_state_t state, fpstate;
 	ucontext_t ctx;
 	mcontext_t mctx;
 	MonoJitTlsData *jit_tls;
@@ -315,13 +315,16 @@ mono_thread_state_init_from_handle (MonoThreadUnwindState *tctx, MonoThreadInfo
 	tctx->unwind_data [MONO_UNWIND_DATA_JIT_TLS] = NULL;
 
 	state = (thread_state_t) alloca (mono_mach_arch_get_thread_state_size ());
+	fpstate = (thread_state_t) alloca (mono_mach_arch_get_thread_fpstate_size ());
 	mctx = (mcontext_t) alloca (mono_mach_arch_get_mcontext_size ());
 
-	ret = mono_mach_arch_get_thread_state (info->native_handle, state, &num_state);
+	do {
+		ret = mono_mach_arch_get_thread_states (info->native_handle, state, &num_state, fpstate, &num_fpstate);
+	} while (ret == KERN_ABORTED);
 	if (ret != KERN_SUCCESS)
 		return FALSE;
 
-	mono_mach_arch_thread_state_to_mcontext (state, mctx);
+	mono_mach_arch_thread_states_to_mcontext (state, fpstate, mctx);
 	ctx.uc_mcontext = mctx;
 
 	mono_sigctx_to_monoctx (&ctx, &tctx->ctx);
diff --git a/mono/utils/mach-support-amd64.c b/mono/utils/mach-support-amd64.c
index 7944e2c986b..42a66d541ce 100644
--- a/mono/utils/mach-support-amd64.c
+++ b/mono/utils/mach-support-amd64.c
@@ -56,21 +56,23 @@ mono_mach_arch_get_mcontext_size ()
 }
 
 void
-mono_mach_arch_thread_state_to_mcontext (thread_state_t state, void *context)
+mono_mach_arch_thread_states_to_mcontext (thread_state_t state, thread_state_t fpstate, void *context)
 {
 	x86_thread_state64_t *arch_state = (x86_thread_state64_t *) state;
+	x86_float_state64_t *arch_fpstate = (x86_float_state64_t *) fpstate;
 	struct __darwin_mcontext64 *ctx = (struct __darwin_mcontext64 *) context;
-
 	ctx->__ss = *arch_state;
+	ctx->__fs = *arch_fpstate;
 }
 
 void
-mono_mach_arch_mcontext_to_thread_state (void *context, thread_state_t state)
+mono_mach_arch_mcontext_to_thread_states (void *context, thread_state_t state, thread_state_t fpstate)
 {
 	x86_thread_state64_t *arch_state = (x86_thread_state64_t *) state;
+	x86_float_state64_t *arch_fpstate = (x86_float_state64_t *) fpstate;
 	struct __darwin_mcontext64 *ctx = (struct __darwin_mcontext64 *) context;
-
 	*arch_state = ctx->__ss;
+	*arch_fpstate = ctx->__fs;
 }
 
 int
@@ -79,23 +81,39 @@ mono_mach_arch_get_thread_state_size ()
 	return sizeof (x86_thread_state64_t);
 }
 
+int
+mono_mach_arch_get_thread_fpstate_size ()
+{
+	return sizeof (x86_float_state64_t);
+}
+
 kern_return_t
-mono_mach_arch_get_thread_state (thread_port_t thread, thread_state_t state, mach_msg_type_number_t *count)
+mono_mach_arch_get_thread_states (thread_port_t thread, thread_state_t state, mach_msg_type_number_t *count, thread_state_t fpstate, mach_msg_type_number_t *fpcount)
 {
-	x86_thread_state64_t *arch_state = (x86_thread_state64_t *) state;
+	x86_thread_state64_t *arch_state = (x86_thread_state64_t *)state;
+	x86_float_state64_t *arch_fpstate = (x86_float_state64_t *)fpstate;
 	kern_return_t ret;
 
 	*count = x86_THREAD_STATE64_COUNT;
+	*fpcount = x86_FLOAT_STATE64_COUNT;
 
-	ret = thread_get_state (thread, x86_THREAD_STATE64, (thread_state_t) arch_state, count);
+	ret = thread_get_state (thread, x86_THREAD_STATE64, (thread_state_t)arch_state, count);
+	if (ret != KERN_SUCCESS)
+		return ret;
 
+	ret = thread_get_state (thread, x86_FLOAT_STATE64, (thread_state_t)arch_fpstate, fpcount);
 	return ret;
 }
 
 kern_return_t
-mono_mach_arch_set_thread_state (thread_port_t thread, thread_state_t state, mach_msg_type_number_t count)
+mono_mach_arch_set_thread_states (thread_port_t thread, thread_state_t state, mach_msg_type_number_t count, thread_state_t fpstate, mach_msg_type_number_t fpcount)
 {
-	return thread_set_state (thread, x86_THREAD_STATE64, state, count);
+	kern_return_t ret;
+	ret = thread_set_state (thread, x86_THREAD_STATE64, state, count);
+	if (ret != KERN_SUCCESS)
+		return ret;
+	ret = thread_set_state (thread, x86_FLOAT_STATE64, fpstate, fpcount);
+	return ret;
 }
 
 void *
diff --git a/mono/utils/mach-support-arm.c b/mono/utils/mach-support-arm.c
index cdf701cd8f4..9e04ddc508c 100644
--- a/mono/utils/mach-support-arm.c
+++ b/mono/utils/mach-support-arm.c
@@ -82,7 +82,7 @@ mono_mach_arch_get_mcontext_size ()
 }
 
 void
-mono_mach_arch_thread_state_to_mcontext (thread_state_t state, void *context)
+mono_mach_arch_thread_states_to_mcontext (thread_state_t state, thread_state_t fpstate, void *context)
 {
 #ifdef TARGET_ARM64
 	arm_unified_thread_state_t *arch_state = (arm_unified_thread_state_t *) state;
@@ -98,7 +98,7 @@ mono_mach_arch_thread_state_to_mcontext (thread_state_t state, void *context)
 }
 
 void
-mono_mach_arch_mcontext_to_thread_state (void *context, thread_state_t state)
+mono_mach_arch_mcontext_to_thread_states (void *context, thread_state_t state, thread_state_t fpstate)
 {
 #ifdef TARGET_ARM64
 	arm_unified_thread_state_t *arch_state = (arm_unified_thread_state_t *) state;
@@ -113,6 +113,20 @@ mono_mach_arch_mcontext_to_thread_state (void *context, thread_state_t state)
 #endif
 }
 
+void
+mono_mach_arch_thread_states_to_mono_context (thread_state_t state, thread_state_t fpstate, MonoContext *context)
+{
+	int i;
+	arm_thread_state_t *arch_state = (arm_thread_state_t *) state;
+	for (i = 0; i < 13; ++i)
+		context->regs [i] = arch_state->__r [i];
+	context->regs [ARMREG_R13] = arch_state->__sp;
+	context->regs [ARMREG_R14] = arch_state->__lr;
+	context->regs [ARMREG_R15] = arch_state->__pc;
+	context->pc = arch_state->__pc;
+	context->cpsr = arch_state->__cpsr;
+}
+
 int
 mono_mach_arch_get_thread_state_size ()
 {
@@ -123,8 +137,14 @@ mono_mach_arch_get_thread_state_size ()
 #endif
 }
 
+int
+mono_mach_arch_get_thread_fpstate_size ()
+{
+	return sizeof (arm_neon_state_t);
+}
+
 kern_return_t
-mono_mach_arch_get_thread_state (thread_port_t thread, thread_state_t state, mach_msg_type_number_t *count)
+mono_mach_arch_get_thread_states (thread_port_t thread, thread_state_t state, mach_msg_type_number_t *count, thread_state_t fpstate, mach_msg_type_number_t *fpcount)
 {
 #ifdef TARGET_ARM64
 	arm_unified_thread_state_t *arch_state = (arm_unified_thread_state_t *) state;
@@ -145,7 +165,7 @@ mono_mach_arch_get_thread_state (thread_port_t thread, thread_state_t state, mac
 }
 
 kern_return_t
-mono_mach_arch_set_thread_state (thread_port_t thread, thread_state_t state, mach_msg_type_number_t count)
+mono_mach_arch_set_thread_states (thread_port_t thread, thread_state_t state, mach_msg_type_number_t count, thread_state_t fpstate, mach_msg_type_number_t fpcount)
 {
 #ifdef TARGET_ARM64
 	return thread_set_state (thread, ARM_UNIFIED_THREAD_STATE, state, count);
diff --git a/mono/utils/mach-support-unknown.c b/mono/utils/mach-support-unknown.c
index 9fece227216..88eb187da44 100644
--- a/mono/utils/mach-support-unknown.c
+++ b/mono/utils/mach-support-unknown.c
@@ -35,17 +35,22 @@ mono_mach_arch_get_mcontext_size ()
 }
 
 void
-mono_mach_arch_thread_state_to_mcontext (thread_state_t state, void *context)
+mono_mach_arch_thread_states_to_mcontext (thread_state_t state, thread_state_t fpstate, void *context)
 {
 	g_assert_not_reached ();
 }
 
 void
-mono_mach_arch_mcontext_to_thread_state (void *context, thread_state_t state)
+mono_mach_arch_mcontext_to_thread_states (void *context, thread_state_t state, thread_state_t fpstate)
 {
 	g_assert_not_reached ();
 }
 
+void
+mono_mach_arch_thread_states_to_mono_context (thread_state_t state, thread_state_t fpstate, MonoContext *context)
+{
+	g_assert_not_reached ();
+}
 
 int
 mono_mach_arch_get_thread_state_size ()
@@ -53,14 +58,20 @@ mono_mach_arch_get_thread_state_size ()
 	g_assert_not_reached ();
 }
 
+int
+mono_mach_arch_get_thread_fpstate_size ()
+{
+	g_assert_not_reached ();
+}
+
 kern_return_t
-mono_mach_arch_get_thread_state (thread_port_t thread, thread_state_t state, mach_msg_type_number_t *count)
+mono_mach_arch_get_thread_states (thread_port_t thread, thread_state_t state, mach_msg_type_number_t *count, thread_state_t fpstate, mach_msg_type_number_t *fpcount)
 {
 	g_assert_not_reached ();
 }
 
 kern_return_t
-mono_mach_arch_set_thread_state (thread_port_t thread, thread_state_t state, mach_msg_type_number_t count)
+mono_mach_arch_set_thread_states (thread_port_t thread, thread_state_t state, mach_msg_type_number_t count, thread_state_t fpstate, mach_msg_type_number_t fpcount)
 {
 	g_assert_not_reached ();	
 }
diff --git a/mono/utils/mach-support-x86.c b/mono/utils/mach-support-x86.c
index 967aeb7abcf..7535b253a4b 100644
--- a/mono/utils/mach-support-x86.c
+++ b/mono/utils/mach-support-x86.c
@@ -16,6 +16,9 @@
 #include "utils/mono-sigcontext.h"
 #include "mach-support.h"
 
+// For reg numbers
+#include <mono/arch/amd64/amd64-codegen.h>
+
 /* Known offsets used for TLS storage*/
 
 /* All OSX versions up to 10.8 */
@@ -55,21 +58,47 @@ mono_mach_arch_get_mcontext_size ()
 }
 
 void
-mono_mach_arch_thread_state_to_mcontext (thread_state_t state, void *context)
+mono_mach_arch_thread_states_to_mcontext (thread_state_t state, thread_state_t fpstate, void *context)
 {
 	x86_thread_state32_t *arch_state = (x86_thread_state32_t *) state;
+	x86_float_state32_t *arch_fpstate = (x86_float_state32_t *) fpstate;
 	struct __darwin_mcontext32 *ctx = (struct __darwin_mcontext32 *) context;
-
 	ctx->__ss = *arch_state;
+	ctx->__fs = *arch_fpstate;
 }
 
 void
-mono_mach_arch_mcontext_to_thread_state (void *context, thread_state_t state)
+mono_mach_arch_mcontext_to_thread_states (void *context, thread_state_t state, thread_state_t fpstate)
 {
 	x86_thread_state32_t *arch_state = (x86_thread_state32_t *) state;
+	x86_float_state32_t *arch_fpstate = (x86_float_state32_t *) fpstate;
 	struct __darwin_mcontext32 *ctx = (struct __darwin_mcontext32 *) context;
-
 	*arch_state = ctx->__ss;
+	*arch_fpstate = ctx->__fs;
+}
+
+void
+mono_mach_arch_thread_states_to_mono_context (thread_state_t state, thread_state_t fpstate, MonoContext *context)
+{
+	x86_thread_state32_t *arch_state = (x86_thread_state32_t *) state;
+	x86_float_state32_t *arch_fpstate = (x86_float_state32_t *) state;
+	context->eax = arch_state->__eax;
+	context->ebx = arch_state->__ebx;
+	context->ecx = arch_state->__ecx;
+	context->edx = arch_state->__edx;
+	context->ebp = arch_state->__ebp;
+	context->esp = arch_state->__esp;
+	context->esi = arch_state->__edi;
+	context->edi = arch_state->__esi;
+	context->eip = arch_state->__eip;
+	context->fregs [X86_XMM0] = arch_fpstate->__fpu_xmm0;
+	context->fregs [X86_XMM1] = arch_fpstate->__fpu_xmm1;
+	context->fregs [X86_XMM2] = arch_fpstate->__fpu_xmm2;
+	context->fregs [X86_XMM3] = arch_fpstate->__fpu_xmm3;
+	context->fregs [X86_XMM4] = arch_fpstate->__fpu_xmm4;
+	context->fregs [X86_XMM5] = arch_fpstate->__fpu_xmm5;
+	context->fregs [X86_XMM6] = arch_fpstate->__fpu_xmm6;
+	context->fregs [X86_XMM7] = arch_fpstate->__fpu_xmm7;
 }
 
 int
@@ -78,23 +107,39 @@ mono_mach_arch_get_thread_state_size ()
 	return sizeof (x86_thread_state32_t);
 }
 
+int
+mono_mach_arch_get_thread_fpstate_size ()
+{
+	return sizeof (x86_float_state32_t);
+}
+
 kern_return_t
-mono_mach_arch_get_thread_state (thread_port_t thread, thread_state_t state, mach_msg_type_number_t *count)
+mono_mach_arch_get_thread_states (thread_port_t thread, thread_state_t state, mach_msg_type_number_t *count, thread_state_t fpstate, mach_msg_type_number_t *fpcount)
 {
 	x86_thread_state32_t *arch_state = (x86_thread_state32_t *) state;
+	x86_float_state32_t *arch_fpstate = (x86_float_state32_t *) fpstate;
 	kern_return_t ret;
 
 	*count = x86_THREAD_STATE32_COUNT;
+	*fpcount = x86_FLOAT_STATE32_COUNT;
 
-	ret = thread_get_state (thread, x86_THREAD_STATE32, (thread_state_t) arch_state, count);
+	ret = thread_get_state (thread, x86_THREAD_STATE32, (thread_state_t)arch_state, count);
+	if (ret != KERN_SUCCESS)
+		return ret;
 
+	ret = thread_get_state (thread, x86_FLOAT_STATE32, (thread_state_t)arch_fpstate, fpcount);
 	return ret;
 }
 
 kern_return_t
-mono_mach_arch_set_thread_state (thread_port_t thread, thread_state_t state, mach_msg_type_number_t count)
+mono_mach_arch_set_thread_states (thread_port_t thread, thread_state_t state, mach_msg_type_number_t count, thread_state_t fpstate, mach_msg_type_number_t fpcount)
 {
-	return thread_set_state (thread, x86_THREAD_STATE32, state, count);
+	kern_return_t ret;
+	ret = thread_set_state (thread, x86_THREAD_STATE32, state, count);
+	if (ret != KERN_SUCCESS)
+		return ret;
+	ret = thread_set_state (thread, x86_FLOAT_STATE32, fpstate, fpcount);
+	return ret;
 }
 
 void *
diff --git a/mono/utils/mach-support.h b/mono/utils/mach-support.h
index 940cf0356a0..4ce00952c0a 100644
--- a/mono/utils/mach-support.h
+++ b/mono/utils/mach-support.h
@@ -28,14 +28,16 @@ void *mono_mach_arch_get_sp (thread_state_t state);
 void mono_mach_init (pthread_key_t key);
 
 int mono_mach_arch_get_mcontext_size (void);
-void mono_mach_arch_thread_state_to_mcontext (thread_state_t state, void *context);
-void mono_mach_arch_mcontext_to_thread_state (void *context, thread_state_t state);
+void mono_mach_arch_thread_states_to_mcontext (thread_state_t state, thread_state_t fpstate, void *context);
+void mono_mach_arch_mcontext_to_thread_states (void *context, thread_state_t state, thread_state_t fpstate);
 
+/* FIXME: Should return size_t, not int. */
 int mono_mach_arch_get_thread_state_size (void);
+int mono_mach_arch_get_thread_fpstate_size (void);
 kern_return_t mono_mach_get_threads (thread_act_array_t *threads, guint32 *count);
 kern_return_t mono_mach_free_threads (thread_act_array_t threads, guint32 count);
-kern_return_t mono_mach_arch_get_thread_state (thread_port_t thread, thread_state_t state, mach_msg_type_number_t *count);
-kern_return_t mono_mach_arch_set_thread_state (thread_port_t thread, thread_state_t state, mach_msg_type_number_t count);
+kern_return_t mono_mach_arch_get_thread_states (thread_port_t thread, thread_state_t state, mach_msg_type_number_t *count, thread_state_t fpstate, mach_msg_type_number_t *fpcount);
+kern_return_t mono_mach_arch_set_thread_states (thread_port_t thread, thread_state_t state, mach_msg_type_number_t count, thread_state_t fpstate, mach_msg_type_number_t fpcount);
 void *mono_mach_arch_get_tls_value_from_thread (pthread_t thread, guint32 key);
 void *mono_mach_get_tls_address_from_thread (pthread_t thread, pthread_key_t key);
 
diff --git a/mono/utils/mono-context.c b/mono/utils/mono-context.c
index a05c7bab76d..c3c173518a3 100644
--- a/mono/utils/mono-context.c
+++ b/mono/utils/mono-context.c
@@ -163,6 +163,26 @@ mono_sigctx_to_monoctx (void *sigctx, MonoContext *mctx)
 	mctx->gregs [AMD64_R14] = UCONTEXT_REG_R14 (ctx);
 	mctx->gregs [AMD64_R15] = UCONTEXT_REG_R15 (ctx);
 	mctx->gregs [AMD64_RIP] = UCONTEXT_REG_RIP (ctx);
+
+#ifdef UCONTEXT_REG_XMM
+	mctx->fregs [0] = UCONTEXT_REG_XMM0 (ctx);
+	mctx->fregs [1] = UCONTEXT_REG_XMM1 (ctx);
+	mctx->fregs [2] = UCONTEXT_REG_XMM2 (ctx);
+	mctx->fregs [3] = UCONTEXT_REG_XMM3 (ctx);
+	mctx->fregs [4] = UCONTEXT_REG_XMM4 (ctx);
+	mctx->fregs [5] = UCONTEXT_REG_XMM5 (ctx);
+	mctx->fregs [6] = UCONTEXT_REG_XMM6 (ctx);
+	mctx->fregs [7] = UCONTEXT_REG_XMM7 (ctx);
+	mctx->fregs [8] = UCONTEXT_REG_XMM8 (ctx);
+	mctx->fregs [9] = UCONTEXT_REG_XMM9 (ctx);
+	mctx->fregs [10] = UCONTEXT_REG_XMM10 (ctx);
+	mctx->fregs [11] = UCONTEXT_REG_XMM11 (ctx);
+	mctx->fregs [12] = UCONTEXT_REG_XMM12 (ctx);
+	mctx->fregs [13] = UCONTEXT_REG_XMM13 (ctx);
+	mctx->fregs [14] = UCONTEXT_REG_XMM14 (ctx);
+	mctx->fregs [15] = UCONTEXT_REG_XMM15 (ctx);
+#endif
+
 #elif defined(HOST_WIN32)
 	CONTEXT *context = (CONTEXT*)sigctx;
 
@@ -233,6 +253,26 @@ mono_monoctx_to_sigctx (MonoContext *mctx, void *sigctx)
 	UCONTEXT_REG_R14 (ctx) = mctx->gregs [AMD64_R14];
 	UCONTEXT_REG_R15 (ctx) = mctx->gregs [AMD64_R15];
 	UCONTEXT_REG_RIP (ctx) = mctx->gregs [AMD64_RIP];
+
+#ifdef UCONTEXT_REG_XMM
+	UCONTEXT_REG_XMM0 (ctx) = mctx->fregs [0];
+	UCONTEXT_REG_XMM1 (ctx) = mctx->fregs [1];
+	UCONTEXT_REG_XMM2 (ctx) = mctx->fregs [2];
+	UCONTEXT_REG_XMM3 (ctx) = mctx->fregs [3];
+	UCONTEXT_REG_XMM4 (ctx) = mctx->fregs [4];
+	UCONTEXT_REG_XMM5 (ctx) = mctx->fregs [5];
+	UCONTEXT_REG_XMM6 (ctx) = mctx->fregs [6];
+	UCONTEXT_REG_XMM7 (ctx) = mctx->fregs [7];
+	UCONTEXT_REG_XMM8 (ctx) = mctx->fregs [8];
+	UCONTEXT_REG_XMM9 (ctx) = mctx->fregs [9];
+	UCONTEXT_REG_XMM10 (ctx) = mctx->fregs [10];
+	UCONTEXT_REG_XMM11 (ctx) = mctx->fregs [11];
+	UCONTEXT_REG_XMM12 (ctx) = mctx->fregs [12];
+	UCONTEXT_REG_XMM13 (ctx) = mctx->fregs [13];
+	UCONTEXT_REG_XMM14 (ctx) = mctx->fregs [14];
+	UCONTEXT_REG_XMM15 (ctx) = mctx->fregs [15];
+#endif
+
 #elif defined(HOST_WIN32)
 	CONTEXT *context = (CONTEXT*)sigctx;
 
diff --git a/mono/utils/mono-context.h b/mono/utils/mono-context.h
index 55d75bda44d..188c0b3f7a3 100644
--- a/mono/utils/mono-context.h
+++ b/mono/utils/mono-context.h
@@ -21,6 +21,13 @@
 #include <ucontext.h>
 #endif
 
+#define MONO_CONTEXT_OFFSET(field, index, field_type) \
+    "i" (offsetof (MonoContext, field) + (index) * sizeof (field_type))
+
+#if defined(__APPLE__)
+typedef struct __darwin_xmm_reg MonoContextSimdReg;
+#endif
+
 /*
  * General notes about mono-context.
  * Each arch defines a MonoContext struct with all GPR regs + IP/PC.
@@ -90,6 +97,8 @@ struct sigcontext {
 # define SC_ESI esi
 #endif
 
+#include <mono/arch/x86/x86-codegen.h>
+
 typedef struct {
 	mgreg_t eax;
 	mgreg_t ebx;
@@ -100,6 +109,9 @@ typedef struct {
 	mgreg_t esi;
 	mgreg_t edi;
 	mgreg_t eip;
+#ifdef __APPLE__
+    MonoContextSimdReg fregs [X86_XMM_NREG];
+#endif
 } MonoContext;
 
 #define MONO_CONTEXT_SET_IP(ctx,ip) do { (ctx)->eip = (mgreg_t)(ip); } while (0); 
@@ -132,18 +144,26 @@ typedef struct {
 #else
 #define MONO_CONTEXT_GET_CURRENT(ctx) \
 	__asm__ __volatile__(   \
-	"movl $0x0, 0x00(%0)\n" \
-	"mov %%ebx, 0x04(%0)\n" \
-	"mov %%ecx, 0x08(%0)\n" \
-	"mov %%edx, 0x0c(%0)\n" \
-	"mov %%ebp, 0x10(%0)\n" \
-	"mov %%esp, 0x14(%0)\n" \
-	"mov %%esi, 0x18(%0)\n" \
-	"mov %%edi, 0x1c(%0)\n" \
+	"movl $0x0, %c[eax](%0)\n" \
+	"mov %%ebx, %c[ebx](%0)\n" \
+	"mov %%ecx, %c[ecx](%0)\n" \
+	"mov %%edx, %c[edx](%0)\n" \
+	"mov %%ebp, %c[ebp](%0)\n" \
+	"mov %%esp, %c[esp](%0)\n" \
+	"mov %%esi, %c[esi](%0)\n" \
+	"mov %%edi, %c[edi](%0)\n" \
 	"call 1f\n"     \
 	"1: pop 0x20(%0)\n"     \
 	:	\
-	: "a" (&(ctx))  \
+	: "a" (&(ctx)),	\
+		[eax] MONO_CONTEXT_OFFSET (eax, 0, mgreg_t), \
+		[ebx] MONO_CONTEXT_OFFSET (ebx, 0, mgreg_t), \
+		[ecx] MONO_CONTEXT_OFFSET (ecx, 0, mgreg_t), \
+		[edx] MONO_CONTEXT_OFFSET (edx, 0, mgreg_t), \
+		[ebp] MONO_CONTEXT_OFFSET (ebp, 0, mgreg_t), \
+		[esp] MONO_CONTEXT_OFFSET (esp, 0, mgreg_t), \
+		[esi] MONO_CONTEXT_OFFSET (esi, 0, mgreg_t), \
+		[edi] MONO_CONTEXT_OFFSET (edi, 0, mgreg_t) \
 	: "memory")
 #endif
 
@@ -163,7 +183,11 @@ typedef struct {
 
 typedef struct {
 	mgreg_t gregs [AMD64_NREG];
+#ifdef __APPLE__
+	MonoContextSimdReg fregs [AMD64_XMM_NREG];
+#else
 	double fregs [AMD64_XMM_NREG];
+#endif
 } MonoContext;
 
 #define MONO_CONTEXT_SET_IP(ctx,ip) do { (ctx)->gregs [AMD64_RIP] = (mgreg_t)(ip); } while (0);
@@ -204,30 +228,100 @@ extern void mono_context_get_current (void *);
 		: "a" ((int64_t)&(ctx))	\
 		: "rdx", "memory")
 #else
-#define MONO_CONTEXT_GET_CURRENT(ctx)	\
-	__asm__ __volatile__(	\
-		"movq $0x0,  0x00(%0)\n"	\
-		"movq %%rbx, 0x08(%0)\n"	\
-		"movq %%rcx, 0x10(%0)\n"	\
-		"movq %%rdx, 0x18(%0)\n"	\
-		"movq %%rbp, 0x20(%0)\n"	\
-		"movq %%rsp, 0x28(%0)\n"	\
-		"movq %%rsi, 0x30(%0)\n"	\
-		"movq %%rdi, 0x38(%0)\n"	\
-		"movq %%r8,  0x40(%0)\n"	\
-		"movq %%r9,  0x48(%0)\n"	\
-		"movq %%r10, 0x50(%0)\n"	\
-		"movq %%r11, 0x58(%0)\n"	\
-		"movq %%r12, 0x60(%0)\n"	\
-		"movq %%r13, 0x68(%0)\n"	\
-		"movq %%r14, 0x70(%0)\n"	\
-		"movq %%r15, 0x78(%0)\n"	\
-		/* "leaq (%%rip), %%rdx\n" is not understood by icc */	\
-		".byte 0x48, 0x8d, 0x15, 0x00, 0x00, 0x00, 0x00\n" \
-		"movq %%rdx, 0x80(%0)\n"	\
-		: 	\
-		: "a" (&(ctx))	\
-		: "rdx", "memory")
+
+#define MONO_CONTEXT_GET_CURRENT_GREGS(ctx) \
+	do { \
+		__asm__ __volatile__(	\
+			"movq $0x0,  %c[rax](%0)\n"	\
+			"movq %%rcx, %c[rcx](%0)\n"	\
+			"movq %%rdx, %c[rdx](%0)\n"	\
+			"movq %%rbx, %c[rbx](%0)\n"	\
+			"movq %%rsp, %c[rsp](%0)\n"	\
+			"movq %%rbp, %c[rbp](%0)\n"	\
+			"movq %%rsi, %c[rsi](%0)\n"	\
+			"movq %%rdi, %c[rdi](%0)\n"	\
+			"movq %%r8,  %c[r8](%0)\n"	\
+			"movq %%r9,  %c[r9](%0)\n"	\
+			"movq %%r10, %c[r10](%0)\n"	\
+			"movq %%r11, %c[r11](%0)\n"	\
+			"movq %%r12, %c[r12](%0)\n"	\
+			"movq %%r13, %c[r13](%0)\n"	\
+			"movq %%r14, %c[r14](%0)\n"	\
+			"movq %%r15, %c[r15](%0)\n"	\
+			/* "leaq (%%rip), %%rdx\n" is not understood by icc */	\
+			".byte 0x48, 0x8d, 0x15, 0x00, 0x00, 0x00, 0x00\n" \
+			"movq %%rdx, %c[rip](%0)\n"	\
+			: 	\
+			: "a" (&(ctx)),	\
+				[rax] MONO_CONTEXT_OFFSET (gregs, AMD64_RAX, mgreg_t),	\
+				[rcx] MONO_CONTEXT_OFFSET (gregs, AMD64_RCX, mgreg_t),	\
+				[rdx] MONO_CONTEXT_OFFSET (gregs, AMD64_RDX, mgreg_t),	\
+				[rbx] MONO_CONTEXT_OFFSET (gregs, AMD64_RBX, mgreg_t),	\
+				[rsp] MONO_CONTEXT_OFFSET (gregs, AMD64_RSP, mgreg_t),	\
+				[rbp] MONO_CONTEXT_OFFSET (gregs, AMD64_RBP, mgreg_t),	\
+				[rsi] MONO_CONTEXT_OFFSET (gregs, AMD64_RSI, mgreg_t),	\
+				[rdi] MONO_CONTEXT_OFFSET (gregs, AMD64_RDI, mgreg_t),	\
+				[r8] MONO_CONTEXT_OFFSET (gregs, AMD64_R8, mgreg_t), \
+				[r9] MONO_CONTEXT_OFFSET (gregs, AMD64_R9, mgreg_t), \
+				[r10] MONO_CONTEXT_OFFSET (gregs, AMD64_R10, mgreg_t),	\
+				[r11] MONO_CONTEXT_OFFSET (gregs, AMD64_R11, mgreg_t),	\
+				[r12] MONO_CONTEXT_OFFSET (gregs, AMD64_R12, mgreg_t),	\
+				[r13] MONO_CONTEXT_OFFSET (gregs, AMD64_R13, mgreg_t),	\
+				[r14] MONO_CONTEXT_OFFSET (gregs, AMD64_R14, mgreg_t),	\
+				[r15] MONO_CONTEXT_OFFSET (gregs, AMD64_R15, mgreg_t),	\
+				[rip] MONO_CONTEXT_OFFSET (gregs, AMD64_RIP, mgreg_t)	\
+			: "rdx", "memory");	\
+	} while (0)
+
+#ifdef UCONTEXT_REG_XMM
+#define MONO_CONTEXT_GET_CURRENT_FREGS(ctx) \
+	do { \
+		__asm__ __volatile__ ( \
+			"movups %%xmm0, %c[xmm0](%0)\n"	\
+			"movups %%xmm1, %c[xmm1](%0)\n"	\
+			"movups %%xmm2, %c[xmm2](%0)\n"	\
+			"movups %%xmm3, %c[xmm3](%0)\n"	\
+			"movups %%xmm4, %c[xmm4](%0)\n"	\
+			"movups %%xmm5, %c[xmm5](%0)\n"	\
+			"movups %%xmm6, %c[xmm6](%0)\n"	\
+			"movups %%xmm7, %c[xmm7](%0)\n"	\
+			"movups %%xmm8, %c[xmm8](%0)\n"	\
+			"movups %%xmm9, %c[xmm9](%0)\n"	\
+			"movups %%xmm10, %c[xmm10](%0)\n"	\
+			"movups %%xmm11, %c[xmm11](%0)\n"	\
+			"movups %%xmm12, %c[xmm12](%0)\n"	\
+			"movups %%xmm12, %c[xmm12](%0)\n"	\
+			"movups %%xmm14, %c[xmm14](%0)\n"	\
+			"movups %%xmm15, %c[xmm15](%0)\n"	\
+			: \
+			: "a" (&(ctx)),	\
+				[xmm0] MONO_CONTEXT_OFFSET (fregs, AMD64_XMM0, MonoContextSimdReg), \
+				[xmm1] MONO_CONTEXT_OFFSET (fregs, AMD64_XMM1, MonoContextSimdReg), \
+				[xmm2] MONO_CONTEXT_OFFSET (fregs, AMD64_XMM2, MonoContextSimdReg), \
+				[xmm3] MONO_CONTEXT_OFFSET (fregs, AMD64_XMM3, MonoContextSimdReg), \
+				[xmm4] MONO_CONTEXT_OFFSET (fregs, AMD64_XMM4, MonoContextSimdReg), \
+				[xmm5] MONO_CONTEXT_OFFSET (fregs, AMD64_XMM5, MonoContextSimdReg), \
+				[xmm6] MONO_CONTEXT_OFFSET (fregs, AMD64_XMM6, MonoContextSimdReg), \
+				[xmm7] MONO_CONTEXT_OFFSET (fregs, AMD64_XMM7, MonoContextSimdReg), \
+				[xmm8] MONO_CONTEXT_OFFSET (fregs, AMD64_XMM8, MonoContextSimdReg), \
+				[xmm9] MONO_CONTEXT_OFFSET (fregs, AMD64_XMM9, MonoContextSimdReg), \
+				[xmm10] MONO_CONTEXT_OFFSET (fregs, AMD64_XMM10, MonoContextSimdReg), \
+				[xmm11] MONO_CONTEXT_OFFSET (fregs, AMD64_XMM11, MonoContextSimdReg), \
+				[xmm12] MONO_CONTEXT_OFFSET (fregs, AMD64_XMM12, MonoContextSimdReg), \
+				[xmm13] MONO_CONTEXT_OFFSET (fregs, AMD64_XMM13, MonoContextSimdReg), \
+				[xmm14] MONO_CONTEXT_OFFSET (fregs, AMD64_XMM14, MonoContextSimdReg), \
+				[xmm15] MONO_CONTEXT_OFFSET (fregs, AMD64_XMM15, MonoContextSimdReg));	\
+	} while (0)
+#else
+#define MONO_CONTEXT_GET_CURRENT_FREGS(ctx)
+#endif
+
+#define MONO_CONTEXT_GET_CURRENT(ctx) \
+    do {	\
+		MONO_CONTEXT_GET_CURRENT_GREGS(ctx);	\
+		MONO_CONTEXT_GET_CURRENT_FREGS(ctx);	\
+	} while (0)
+
 #endif
 
 #define MONO_ARCH_HAS_MONO_CONTEXT 1
diff --git a/mono/utils/mono-sigcontext.h b/mono/utils/mono-sigcontext.h
index de4f6c45b75..6d47813802b 100644
--- a/mono/utils/mono-sigcontext.h
+++ b/mono/utils/mono-sigcontext.h
@@ -173,6 +173,23 @@ typedef struct ucontext {
 	#define UCONTEXT_REG_R13(ctx) (((ucontext_t*)(ctx))->uc_mcontext->__ss.__r13)
 	#define UCONTEXT_REG_R14(ctx) (((ucontext_t*)(ctx))->uc_mcontext->__ss.__r14)
 	#define UCONTEXT_REG_R15(ctx) (((ucontext_t*)(ctx))->uc_mcontext->__ss.__r15)
+	#define UCONTEXT_REG_XMM
+	#define UCONTEXT_REG_XMM0(ctx) (((ucontext_t*)(ctx))->uc_mcontext->__fs.__fpu_xmm0)
+	#define UCONTEXT_REG_XMM1(ctx) (((ucontext_t*)(ctx))->uc_mcontext->__fs.__fpu_xmm1)
+	#define UCONTEXT_REG_XMM2(ctx) (((ucontext_t*)(ctx))->uc_mcontext->__fs.__fpu_xmm2)
+	#define UCONTEXT_REG_XMM3(ctx) (((ucontext_t*)(ctx))->uc_mcontext->__fs.__fpu_xmm3)
+	#define UCONTEXT_REG_XMM4(ctx) (((ucontext_t*)(ctx))->uc_mcontext->__fs.__fpu_xmm4)
+	#define UCONTEXT_REG_XMM5(ctx) (((ucontext_t*)(ctx))->uc_mcontext->__fs.__fpu_xmm5)
+	#define UCONTEXT_REG_XMM6(ctx) (((ucontext_t*)(ctx))->uc_mcontext->__fs.__fpu_xmm6)
+	#define UCONTEXT_REG_XMM7(ctx) (((ucontext_t*)(ctx))->uc_mcontext->__fs.__fpu_xmm7)
+	#define UCONTEXT_REG_XMM8(ctx) (((ucontext_t*)(ctx))->uc_mcontext->__fs.__fpu_xmm8)
+	#define UCONTEXT_REG_XMM9(ctx) (((ucontext_t*)(ctx))->uc_mcontext->__fs.__fpu_xmm9)
+	#define UCONTEXT_REG_XMM10(ctx) (((ucontext_t*)(ctx))->uc_mcontext->__fs.__fpu_xmm10)
+	#define UCONTEXT_REG_XMM11(ctx) (((ucontext_t*)(ctx))->uc_mcontext->__fs.__fpu_xmm11)
+	#define UCONTEXT_REG_XMM12(ctx) (((ucontext_t*)(ctx))->uc_mcontext->__fs.__fpu_xmm12)
+	#define UCONTEXT_REG_XMM13(ctx) (((ucontext_t*)(ctx))->uc_mcontext->__fs.__fpu_xmm13)
+	#define UCONTEXT_REG_XMM14(ctx) (((ucontext_t*)(ctx))->uc_mcontext->__fs.__fpu_xmm14)
+	#define UCONTEXT_REG_XMM15(ctx) (((ucontext_t*)(ctx))->uc_mcontext->__fs.__fpu_xmm15)
 #elif defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
 	#define UCONTEXT_REG_RAX(ctx) (((ucontext_t*)(ctx))->uc_mcontext.mc_rax)
 	#define UCONTEXT_REG_RBX(ctx) (((ucontext_t*)(ctx))->uc_mcontext.mc_rbx)
@@ -224,8 +241,8 @@ typedef struct ucontext {
 	#define UCONTEXT_REG_R13(ctx) (((ucontext_t*)(ctx))->sc_r13)
 	#define UCONTEXT_REG_R14(ctx) (((ucontext_t*)(ctx))->sc_r14)
 	#define UCONTEXT_REG_R15(ctx) (((ucontext_t*)(ctx))->sc_r15)
-#else
-#define UCONTEXT_GREGS(ctx)	((guint64*)&(((ucontext_t*)(ctx))->uc_mcontext.gregs))
+#elif !defined(HOST_WIN32)
+	#define UCONTEXT_GREGS(ctx)	((guint64*)&(((ucontext_t*)(ctx))->uc_mcontext.gregs))
 #endif
 
 #ifdef UCONTEXT_GREGS
diff --git a/mono/utils/mono-threads-mach.c b/mono/utils/mono-threads-mach.c
index fc884e7a055..66af94c24b1 100644
--- a/mono/utils/mono-threads-mach.c
+++ b/mono/utils/mono-threads-mach.c
@@ -116,8 +116,8 @@ mono_threads_core_begin_async_resume (MonoThreadInfo *info)
 
 	if (info->async_target) {
 		MonoContext tmp = info->thread_saved_state [ASYNC_SUSPEND_STATE_INDEX].ctx;
-		mach_msg_type_number_t num_state;
-		thread_state_t state;
+		mach_msg_type_number_t num_state, num_fpstate;
+		thread_state_t state, fpstate;
 		ucontext_t uctx;
 		mcontext_t mctx;
 
@@ -126,13 +126,17 @@ mono_threads_core_begin_async_resume (MonoThreadInfo *info)
 		info->async_target = (void (*)(void *)) info->user_data;
 
 		state = (thread_state_t) alloca (mono_mach_arch_get_thread_state_size ());
+		fpstate = (thread_state_t) alloca (mono_mach_arch_get_thread_fpstate_size ());
 		mctx = (mcontext_t) alloca (mono_mach_arch_get_mcontext_size ());
 
-		ret = mono_mach_arch_get_thread_state (info->native_handle, state, &num_state);
+		do {
+			ret = mono_mach_arch_get_thread_states (info->native_handle, state, &num_state, fpstate, &num_fpstate);
+		} while (ret == KERN_ABORTED);
+
 		if (ret != KERN_SUCCESS)
 			return FALSE;
 
-		mono_mach_arch_thread_state_to_mcontext (state, mctx);
+		mono_mach_arch_thread_states_to_mcontext (state, fpstate, mctx);
 #ifdef TARGET_ARM64
 		g_assert_not_reached ();
 #else
@@ -140,9 +144,12 @@ mono_threads_core_begin_async_resume (MonoThreadInfo *info)
 #endif
 		mono_monoctx_to_sigctx (&tmp, &uctx);
 
-		mono_mach_arch_mcontext_to_thread_state (mctx, state);
+		mono_mach_arch_mcontext_to_thread_states (mctx, state, fpstate);
+
+		do {
+			ret = mono_mach_arch_set_thread_states (info->native_handle, state, num_state, fpstate, num_fpstate);
+		} while (ret == KERN_ABORTED);
 
-		ret = mono_mach_arch_set_thread_state (info->native_handle, state, num_state);
 		if (ret != KERN_SUCCESS)
 			return FALSE;
 	}