1.0.22.15: 16 byte align assembly routines on x86 and x86-64

[sbcl.git] / src / runtime / x86-assem.S

diff --git a/src/runtime/x86-assem.S b/src/runtime/x86-assem.S
index e90f972..315643b 100644 (file)
--- a/src/runtime/x86-assem.S
+++ b/src/runtime/x86-assem.S
@@ -127,7 +127,7 @@ GNAME(call_into_c):
 
 /* Check for a return FP value. */
        fxam
-       fnstsw  %eax
+       fnstsw  %ax
        andl    $0x4500,%eax
        cmpl    $0x4100,%eax
        jne     Lfp_rtn_value
@@ -185,10 +185,8 @@ GNAME(call_into_lisp_first_time):
 #ifndef LISP_FEATURE_WIN32
        movl    %esp,ALIEN_STACK + SYMBOL_VALUE_OFFSET
        movl    GNAME(all_threads),%eax
-       movl    THREAD_CONTROL_STACK_START_OFFSET(%eax) ,%esp
-       /* don't think too hard about what happens if we get interrupted
-       * here */
-       addl    $(THREAD_CONTROL_STACK_SIZE),%esp
+       /* pthread machinery takes care of this for other threads */
+       movl    THREAD_CONTROL_STACK_END_OFFSET(%eax) ,%esp
 #else
 /* Win32 -really- doesn't like you switching stacks out from under it. */
        movl    GNAME(all_threads),%eax
@@ -337,7 +335,7 @@ GNAME(fpu_restore):
  * the undefined-function trampoline
  */
        .text
-       .align  align_4byte,0x90
+       .align  align_16byte,0x90
        .globl GNAME(undefined_tramp)
        TYPE(GNAME(undefined_tramp))
         .byte   0, 0, 0, SIMPLE_FUN_HEADER_WIDETAG
@@ -354,7 +352,7 @@ GNAME(undefined_tramp):
  * the closure trampoline
  */
        .text
-       .align  align_4byte,0x90
+       .align  align_16byte,0x90
        .globl GNAME(closure_tramp)
        TYPE(GNAME(closure_tramp))
         .byte   0, 0, 0, SIMPLE_FUN_HEADER_WIDETAG
@@ -370,7 +368,7 @@ GNAME(closure_tramp):
        SIZE(GNAME(closure_tramp))
 
        .text
-       .align  align_4byte,0x90
+       .align  align_16byte,0x90
        .globl GNAME(funcallable_instance_tramp)
        TYPE(GNAME(funcallable_instance_tramp))
 GNAME(funcallable_instance_tramp):
@@ -385,7 +383,7 @@ GNAME(funcallable_instance_tramp):
  */
        .text
        .globl  GNAME(fun_end_breakpoint_guts)
-       .align  align_4byte
+       .align  align_16byte
 GNAME(fun_end_breakpoint_guts):
        /* Multiple Value return */
        jc      multiple_value_return
@@ -412,7 +410,7 @@ GNAME(fun_end_breakpoint_end):
 \f
        .globl  GNAME(do_pending_interrupt)
        TYPE(GNAME(do_pending_interrupt))
-       .align  align_4byte,0x90
+       .align  align_16byte,0x90
 GNAME(do_pending_interrupt):
        TRAP
        .byte   trap_PendingInterrupt
@@ -437,9 +435,9 @@ GNAME(do_pending_interrupt):
 #define ALLOC(size)                                             \
         pushl   %ebp;              /* Save EBP               */ \
         movl    %esp,%ebp;         /* Save ESP to EBP        */ \
-        andl    $0xfffffff0,%esp;  /* Align stack            */ \
-        pushl   $0;                /* Padding                */ \
-        pushl   size;              /* Argument to alloc      */ \
+        pushl   $0;                /* Reserve space for arg  */ \
+        andl    $0xfffffff0,%esp;  /* Align stack to 16bytes */ \
+        movl    size, (%esp);      /* Argument to alloc      */ \
         call    GNAME(alloc);                                   \
         movl    %ebp,%esp;         /* Restore ESP from EBP   */ \
         popl    %ebp;              /* Restore EBP            */
@@ -453,7 +451,7 @@ GNAME(do_pending_interrupt):
 #define DEFINE_ALLOC_TO_EAX(name,size)                          \
         .globl  GNAME(name);                                    \
         TYPE(GNAME(name));                                      \
-        .align  align_4byte,0x90;                               \
+        .align  align_16byte,0x90;                              \
 GNAME(name):                                                    \
         pushl   %ecx;              /* Save ECX and EDX       */ \
         pushl   %edx;                                           \
@@ -466,7 +464,7 @@ GNAME(name):                                                    \
 #define DEFINE_ALLOC_TO_ECX(name,size)                          \
         .globl  GNAME(name);                                    \
         TYPE(GNAME(name));                                      \
-        .align  align_4byte,0x90;                               \
+        .align  align_16byte,0x90;                              \
 GNAME(name):                                                    \
         pushl   %eax;              /* Save EAX and EDX       */ \
         pushl   %edx;                                           \
@@ -480,7 +478,7 @@ GNAME(name):                                                    \
 #define DEFINE_ALLOC_TO_EDX(name,size)                          \
         .globl  GNAME(name);                                    \
         TYPE(GNAME(name));                                      \
-        .align  align_4byte,0x90;                               \
+        .align  align_16byte,0x90;                              \
 GNAME(name):                                                    \
         pushl   %eax;               /* Save EAX and ECX      */ \
         pushl   %ecx;                                           \
@@ -494,7 +492,7 @@ GNAME(name):                                                    \
 #define DEFINE_ALLOC_TO_REG(name,reg,size)                      \
         .globl  GNAME(name);                                    \
         TYPE(GNAME(name));                                      \
-        .align  align_4byte,0x90;                               \
+        .align  align_16byte,0x90;                              \
 GNAME(name):                                                    \
         pushl   %eax;              /* Save EAX, ECX, and EDX */ \
         pushl   %ecx;                                           \
@@ -549,7 +547,7 @@ DEFINE_ALLOC_TO_REG(alloc_16_to_edi,%edi,$16)
 
 /* This routine handles an overflow with eax=crfp+size. So the
    size=eax-crfp. */
-        .align  align_4byte
+        .align  align_16byte
         .globl  GNAME(alloc_overflow_eax)
        TYPE(GNAME(alloc_overflow_eax))
 GNAME(alloc_overflow_eax):
@@ -561,7 +559,7 @@ GNAME(alloc_overflow_eax):
         ret
         SIZE(GNAME(alloc_overflow_eax))
 
-        .align  align_4byte
+        .align  align_16byte
         .globl  GNAME(alloc_overflow_ecx)
        TYPE(GNAME(alloc_overflow_ecx))
 GNAME(alloc_overflow_ecx):
@@ -574,7 +572,7 @@ GNAME(alloc_overflow_ecx):
         ret
         SIZE(GNAME(alloc_overflow_ecx))
 
-        .align  align_4byte
+        .align  align_16byte
         .globl  GNAME(alloc_overflow_edx)
         TYPE(GNAME(alloc_overflow_edx))
 GNAME(alloc_overflow_edx):
@@ -589,7 +587,7 @@ GNAME(alloc_overflow_edx):
 
 /* This routine handles an overflow with ebx=crfp+size. So the
    size=ebx-crfp. */
-        .align  align_4byte
+        .align  align_16byte
         .globl  GNAME(alloc_overflow_ebx)
         TYPE(GNAME(alloc_overflow_ebx))
 GNAME(alloc_overflow_ebx):
@@ -606,7 +604,7 @@ GNAME(alloc_overflow_ebx):
 
 /* This routine handles an overflow with esi=crfp+size. So the
    size=esi-crfp. */
-        .align  align_4byte
+        .align  align_16byte
         .globl  GNAME(alloc_overflow_esi)
         TYPE(GNAME(alloc_overflow_esi))
 GNAME(alloc_overflow_esi):
@@ -621,7 +619,7 @@ GNAME(alloc_overflow_esi):
         ret
         SIZE(GNAME(alloc_overflow_esi))
 
-        .align  align_4byte
+        .align  align_16byte
         .globl  GNAME(alloc_overflow_edi)
         TYPE(GNAME(alloc_overflow_edi))
 GNAME(alloc_overflow_edi):
@@ -644,7 +642,7 @@ GNAME(alloc_overflow_edi):
         * and EIP from the exception context and use it to fake
         * up a stack frame which will skip over the system SEH
         * code. */
-       .align  align_4byte
+       .align  align_16byte
        .globl  GNAME(exception_handler_wrapper)
        TYPE(GNAME(exception_handler_wrapper))
 GNAME(exception_handler_wrapper):
@@ -675,7 +673,7 @@ GNAME(exception_handler_wrapper):
 #endif
 
 #ifdef LISP_FEATURE_DARWIN
-        .align align_4byte
+        .align align_16byte
         .globl GNAME(call_into_lisp_tramp)
         TYPE(GNAME(call_into_lisp_tramp))
 GNAME(call_into_lisp_tramp):
@@ -724,7 +722,7 @@ GNAME(call_into_lisp_tramp):
         SIZE(call_into_lisp_tramp)
 #endif
         
-       .align  align_4byte,0x90
+       .align  align_16byte,0x90
        .globl  GNAME(post_signal_tramp)
        TYPE(GNAME(post_signal_tramp))
 GNAME(post_signal_tramp):
@@ -749,7 +747,7 @@ GNAME(post_signal_tramp):
 \f
        .globl GNAME(fast_bzero_pointer)
        .data
-        .align  align_4byte
+        .align  align_16byte
 GNAME(fast_bzero_pointer):
         /* Variable containing a pointer to the bzero function to use.
          * Initially points to a basic function.  Change this variable
@@ -757,7 +755,7 @@ GNAME(fast_bzero_pointer):
         .long GNAME(fast_bzero_base)
 \f
        .text
-       .align  align_8byte,0x90
+       .align  align_16byte,0x90
        .globl GNAME(fast_bzero)
        TYPE(GNAME(fast_bzero))
 GNAME(fast_bzero):        
@@ -767,7 +765,7 @@ GNAME(fast_bzero):
         
 \f      
        .text
-       .align  align_8byte,0x90
+       .align  align_16byte,0x90
        .globl GNAME(fast_bzero_detect)
        TYPE(GNAME(fast_bzero_detect))
 GNAME(fast_bzero_detect):
@@ -805,7 +803,7 @@ Lrestore:
         
 \f
        .text
-       .align  align_8byte,0x90
+       .align  align_16byte,0x90
        .globl GNAME(fast_bzero_sse)
        TYPE(GNAME(fast_bzero_sse))
         
@@ -851,7 +849,7 @@ Lend_sse:
                 
 \f
        .text
-       .align  align_8byte,0x90
+       .align  align_16byte,0x90
        .globl GNAME(fast_bzero_base)
        TYPE(GNAME(fast_bzero_base))