diff --git a/dlls/ntdll/relay.c b/dlls/ntdll/relay.c
index ee7d9951c59..be2dc833377 100644
--- a/dlls/ntdll/relay.c
+++ b/dlls/ntdll/relay.c
@@ -545,7 +545,6 @@ DECLSPEC_HIDDEN void WINAPI relay_trace_exit( struct relay_descr *descr, unsigne
 
 extern LONGLONG WINAPI relay_call( struct relay_descr *descr, unsigned int idx, const DWORD *stack );
 __ASM_GLOBAL_FUNC( relay_call,
-                   ".arm\n\t"
                    "push {r4-r8,lr}\n\t"
                    "sub sp, #16\n\t"
                    "mov r6, r2\n\t"
@@ -559,8 +558,9 @@ __ASM_GLOBAL_FUNC( relay_call,
                    "lsl r3, r1, #2\n\t"
                    "subs r3, #16\n\t"   /* first 4 args are in registers */
                    "ble 2f\n\t"
+                   "add r3, #7\n\t"
+                   "and r3, #~7\n"
                    "sub sp, r3\n\t"
-                   "and sp, #~7\n"
                    "add r2, r6, #16\n\t"   /* skip r0-r3 */
                    "1:\tsubs r3, r3, #4\n\t"
                    "ldr r0, [r2, r3]\n\t"
@@ -570,6 +570,7 @@ __ASM_GLOBAL_FUNC( relay_call,
 #ifndef __SOFTFP__
                    "tst r1, #0x80000000\n\t"
                    "ldm r6, {r0-r3}\n\t"
+                   "it ne\n\t"
                    "vldmdbne r6!, {s0-s15}\n\t"
 #else
                    "ldm r6, {r0-r3}\n\t"
diff --git a/dlls/ntdll/signal_arm.c b/dlls/ntdll/signal_arm.c
index 512462b7066..0461183460d 100644
--- a/dlls/ntdll/signal_arm.c
+++ b/dlls/ntdll/signal_arm.c
@@ -61,8 +61,8 @@ __ASM_GLOBAL_FUNC( __chkstk, "lsl r4, r4, #2\n\t"
  *		RtlCaptureContext (NTDLL.@)
  */
 __ASM_STDCALL_FUNC( RtlCaptureContext, 4,
-                    ".arm\n\t"
-                    "stmib r0, {r0-r12}\n\t"   /* context->R0..R12 */
+                    "str r0, [r0, #0x4]\n\t"   /* context->R0 */
+                    "str r1, [r0, #0x8]\n\t"   /* context->R1 */
                     "mov r1, #0x0200000\n\t"   /* CONTEXT_ARM */
                     "add r1, r1, #0x3\n\t"     /* CONTEXT_FULL */
                     "str r1, [r0]\n\t"         /* context->ContextFlags */
@@ -70,7 +70,13 @@ __ASM_STDCALL_FUNC( RtlCaptureContext, 4,
                     "str LR, [r0, #0x3c]\n\t"  /* context->Lr */
                     "str LR, [r0, #0x40]\n\t"  /* context->Pc */
                     "mrs r1, CPSR\n\t"
+                    "tst lr, #1\n\t"           /* Thumb? */
+                    "ite ne\n\t"
+                    "orrne r1, r1, #0x20\n\t"
+                    "biceq r1, r1, #0x20\n\t"
                     "str r1, [r0, #0x44]\n\t"  /* context->Cpsr */
+                    "add r0, #0x0c\n\t"
+                    "stm r0, {r2-r12}\n\t"     /* context->R2..R12 */
                     "bx lr"
                     )
 
@@ -257,6 +263,12 @@ __ASM_STDCALL_FUNC( RtlRaiseException, 4,
                     "ldr r0, [sp, #0x1a0]\n\t" /* rec */
                     "ldr r1, [sp, #0x1a4]\n\t"
                     "str r1, [sp, #0x40]\n\t"  /* context->Pc */
+                    "ldr r2, [sp, #0x44]\n\t"  /* context->Cpsr */
+                    "tst r1, #1\n\t"           /* Thumb? */
+                    "ite ne\n\t"
+                    "orrne r2, r2, #0x20\n\t"
+                    "biceq r2, r2, #0x20\n\t"
+                    "str r2, [sp, #0x44]\n\t"  /* context->Cpsr */
                     "str r1, [r0, #12]\n\t"    /* rec->ExceptionAddress */
                     "add r1, sp, #0x1a8\n\t"
                     "str r1, [sp, #0x38]\n\t"  /* context->Sp */
diff --git a/dlls/ntdll/unix/signal_arm.c b/dlls/ntdll/unix/signal_arm.c
index 487c8501cc7..6fa37b81415 100644
--- a/dlls/ntdll/unix/signal_arm.c
+++ b/dlls/ntdll/unix/signal_arm.c
@@ -316,17 +316,18 @@ static void restore_context( const CONTEXT *context, ucontext_t *sigcontext )
  */
 void DECLSPEC_HIDDEN set_cpu_context( const CONTEXT *context );
 __ASM_GLOBAL_FUNC( set_cpu_context,
-                   ".arm\n\t"
                    "ldr r2, [r0, #0x44]\n\t"  /* context->Cpsr */
                    "tst r2, #0x20\n\t"        /* thumb? */
                    "ldr r1, [r0, #0x40]\n\t"  /* context->Pc */
+                   "ite ne\n\t"
                    "orrne r1, r1, #1\n\t"     /* Adjust PC according to thumb */
                    "biceq r1, r1, #1\n\t"     /* Adjust PC according to arm */
                    "msr CPSR_f, r2\n\t"
                    "ldr lr, [r0, #0x3c]\n\t"  /* context->Lr */
                    "ldr sp, [r0, #0x38]\n\t"  /* context->Sp */
                    "push {r1}\n\t"
-                   "ldmib r0, {r0-r12}\n\t"   /* context->R0..R12 */
+                   "add r0, #0x4\n\t"
+                   "ldm r0, {r0-r12}\n\t"     /* context->R0..R12 */
                    "pop {pc}" )
 
 
@@ -611,7 +612,8 @@ __ASM_GLOBAL_FUNC( call_user_apc_dispatcher,
                    "mov sp, r0\n\t"
                    "b 2f\n"
                    "1:\tldr r0, [r10]\n\t"
-                   "sub sp, r0, #0x1a0\n\t"
+                   "sub r0, #0x1a0\n\t"
+                   "mov sp, r0\n\t"
                    "mov r0, #3\n\t"
                    "movt r0, #32\n\t"
                    "str r0, [sp]\n\t"         /* context.ContextFlags = CONTEXT_FULL */
@@ -658,7 +660,8 @@ __ASM_GLOBAL_FUNC( call_user_exception_dispatcher,
                    "ldm r3, {r5-r11}\n\t"
                    "ldr r4, [r3, #32]\n\t"
                    "ldr lr, [r3, #36]\n\t"
-                   "add sp, r3, #40\n\t"
+                   "add r3, #40\n\t"
+                   "mov sp, r3\n\t"
                    "bx r2" )
 
 
@@ -1031,7 +1034,6 @@ PCONTEXT DECLSPEC_HIDDEN get_initial_context( LPTHREAD_START_ROUTINE entry, void
  *           signal_start_thread
  */
 __ASM_GLOBAL_FUNC( signal_start_thread,
-                   ".arm\n\t"
                    "push {r4-r12,lr}\n\t"
                    "mov r5, r3\n\t"           /* thunk */
                    /* store exit frame */
@@ -1039,7 +1041,8 @@ __ASM_GLOBAL_FUNC( signal_start_thread,
                    "str sp, [r3, #0x1d4]\n\t" /* arm_thread_data()->exit_frame */
                    /* switch to thread stack */
                    "ldr r4, [r3, #4]\n\t"     /* teb->Tib.StackBase */
-                   "sub sp, r4, #0x1000\n\t"
+                   "sub r4, #0x1000\n\t"
+                   "mov sp, r4\n\t"
                    /* attach dlls */
                    "bl " __ASM_NAME("get_initial_context") "\n\t"
                    "mov lr, #0\n\t"
@@ -1048,11 +1051,11 @@ __ASM_GLOBAL_FUNC( signal_start_thread,
 
 extern void DECLSPEC_NORETURN call_thread_exit_func( int status, void (*func)(int), TEB *teb );
 __ASM_GLOBAL_FUNC( call_thread_exit_func,
-                   ".arm\n\t"
                    "ldr r3, [r2, #0x1d4]\n\t"  /* arm_thread_data()->exit_frame */
                    "mov ip, #0\n\t"
                    "str ip, [r2, #0x1d4]\n\t"
                    "cmp r3, ip\n\t"
+                   "it ne\n\t"
                    "movne sp, r3\n\t"
                    "blx r1" )
 
diff --git a/include/wine/asm.h b/include/wine/asm.h
index 236b35cf8fe..b9dae95dc6d 100644
--- a/include/wine/asm.h
+++ b/include/wine/asm.h
@@ -59,7 +59,9 @@
 # define __ASM_FUNC_TYPE(name) ".def " name "\n\t.scl 2\n\t.type 32\n\t.endef"
 #elif defined(__APPLE__)
 # define __ASM_FUNC_TYPE(name) ""
-#elif defined(__arm__) || defined(__arm64__)
+#elif defined(__arm__) && defined(__thumb__)
+# define __ASM_FUNC_TYPE(name) ".type " name ",%function\n\t.thumb_func"
+#elif defined(__arm__) || defined(__aarch64__)
 # define __ASM_FUNC_TYPE(name) ".type " name ",%function"
 #else
 # define __ASM_FUNC_TYPE(name) ".type " name ",@function"