Stackless issue python#97: Fix stack switching for optimized builds

This change removes the "static" declaration from the function slp_switch().
This forces the compiler to adhere to the ABI specification.

On i386 and amd64 it is now save to configure stackless with
--enable-stacklessfewerregisters, except if you compile for darwin. The flag
disables explicit saving of %ebp/%rbp.

On Unix-like systems the source file slp_transfer.c now gets compiled
with the additional flag -fno-inline-functions instead of -O2.

https://bitbucket.org/stackless-dev/stackless/issues/98
(grafted from 4f7698499ad53e5fad61fb415fbfe2c036672a9c)

Author: Anselm Kruis

Stackless/changelog.txt

@@ -9,6 +9,14 @@ What's New in Stackless 3.X.X?
 
 *Release date: 20XX-XX-XX*
 
+- https://bitbucket.org/stackless-dev/stackless/issues/97
+  Fix the stack switching for optimized builds for all Unix-like architectures
+  except amd64, where it was already OK. This change removes the "static"
+  declaration from the function slp_switch(). This forces the compiler
+  to adhere to the ABI specification. On i386 and amd64 it is now save to
+  configure stackless with --enable-stacklessfewerregisters, except if you
+  compile for darwin.
+
 - https://bitbucket.org/stackless-dev/stackless/issues/94
   Calls to __init__(self, ...) are now stackless, if
   soft-switching is enabled.

Stackless/platf/switch_amd64_unix.h

@@ -104,12 +104,20 @@ int
 slp_switch(void)
 {
     register long *stackref, stsizediff;
-    void * rbp; int mxcsr; short x87cw;
+    int mxcsr; short x87cw;
+#if STACKLESS_FRHACK
+    __asm__ volatile (
+        "fstcw %0\n\t"
+        "stmxcsr %1\n\t"
+        : "=m" (x87cw), "=m" (mxcsr) : : REGS_CLOBBERED );
+#else
+    void * rbp;
     __asm__ volatile (
         "fstcw %0\n\t"
         "stmxcsr %1\n\t"
         "movq %%rbp, %2\n\t"
         : "=m" (x87cw), "=m" (mxcsr), "=m" (rbp) : : REGS_CLOBBERED );
+#endif
     __asm__ ("movq %%rsp, %0" : "=g" (stackref));
     {
         SLP_SAVE_STATE(stackref, stsizediff);
@@ -120,11 +128,18 @@ slp_switch(void)
             : "r" (stsizediff)
             );
         SLP_RESTORE_STATE();
+#if STACKLESS_FRHACK
+        __asm__ volatile (
+	    "ldmxcsr %1\n\t"
+            "fldcw %0\n\t"
+            : : "m" (x87cw), "m" (mxcsr));
+#else
         __asm__ volatile (
             "movq %2, %%rbp\n\t"
 	    "ldmxcsr %1\n\t"
             "fldcw %0\n\t"
             : : "m" (x87cw), "m" (mxcsr), "m" (rbp));
+#endif
         return 0;
     }
 }

Stackless/platf/switch_arm32_gcc.h

@@ -27,7 +27,15 @@
 #define STACK_MAGIC 0
 #define REGS_TO_SAVE /*"r1", "r2", "r3", "r4",*/ "r5", "r6", "fp", "ip", "lr"
 
-static int
+/*
+ * You may want to make the function static enable optimizations.
+ * However, the ABI SPEC does not apply to static functions. Therefore
+ * I make slp_switch a regular global function.
+ */
+#if 0
+static
+#endif
+int
 slp_switch(void)
 {
     register int *stackref, stsizediff;
@@ -44,7 +52,6 @@ slp_switch(void)
         SLP_RESTORE_STATE();
         return 0;
     }
-    __asm__ volatile ("" : : : REGS_TO_SAVE);
 }
 
 #endif

Stackless/platf/switch_mips_unix.h

@@ -19,7 +19,16 @@
 #define REGS_TO_SAVE "$16", "$17", "$18", "$19", "$20", "$21", "$22", \
        "$23", "$30"
 #endif
-static int
+    __asm__ volatile ("" : : : REGS_TO_SAVE);
+/*
+ * You may want to make the function static enable optimizations.
+ * However, the ABI SPEC does not apply to static functions. Therefore
+ * I make slp_switch a regular global function.
+ */
+#if 0
+static
+#endif
+int
 slp_switch(void)
 {
     register int *stackref, stsizediff;
@@ -38,7 +47,6 @@ slp_switch(void)
             );
         SLP_RESTORE_STATE();
     }
-    /* __asm__ __volatile__ ("" : : : REGS_TO_SAVE); */
     return 0;
 }
 

Stackless/platf/switch_ppc_macosx.h

@@ -42,7 +42,15 @@
        "cr2", "cr3", "cr4"
 #endif
 
-static int
+/*
+ * You may want to make the function static enable optimizations.
+ * However, the ABI SPEC does not apply to static functions. Therefore
+ * I make slp_switch a regular global function.
+ */
+#if 0
+static
+#endif
+int
 slp_switch(void)
 {
     static int x = 0;

Stackless/platf/switch_ppc_unix.h

@@ -37,7 +37,15 @@
 #define REGS_TO_SAVE "r13", "r14", "r15", "r16", "r17", "r18", "r19", "r20", \
        "r21", "r22", "r23", "r24", "r25", "r26", "r27", "r28", "r29", "r31", \
        "cr2", "cr3", "cr4"
-static int
+/*
+ * You may want to make the function static enable optimizations.
+ * However, the ABI SPEC does not apply to static functions. Therefore
+ * I make slp_switch a regular global function.
+ */
+#if 0
+static
+#endif
+int
 slp_switch(void)
 {
     register int *stackref, stsizediff;

Stackless/platf/switch_s390_unix.h

@@ -21,7 +21,15 @@
              "f0", "f1", "f2", "f3", "f4", "f5", "f6", "f7", \
              "f8", "f9", "f10", "f11", "f12", "f13", "f14", "f15"
 
-static int
+/*
+ * You may want to make the function static enable optimizations.
+ * However, the ABI SPEC does not apply to static functions. Therefore
+ * I make slp_switch a regular global function.
+ */
+#if 0
+static
+#endif
+int
 slp_switch(void)
 {
     register int *stackref, stsizediff;
@@ -37,7 +45,6 @@ slp_switch(void)
         SLP_RESTORE_STATE();
         return 0;
     }
-    __asm__ volatile ("" : : : REGS_TO_SAVE);
 }
 
 #endif

Stackless/platf/switch_sparc_sun_gcc.h

@@ -23,7 +23,15 @@
 
 #define STACK_MAGIC 0
 
-static int
+/*
+ * You may want to make the function static enable optimizations.
+ * However, the ABI SPEC does not apply to static functions. Therefore
+ * I make slp_switch a regular global function.
+ */
+#if 0
+static
+#endif
+int
 slp_switch(void)
 {
     register int *stackref, stsizediff;

Stackless/platf/switch_x86_unix.h

@@ -2,6 +2,8 @@
  * this is the internal transfer function.
  *
  * HISTORY
+ * 06-Nov-16  Anselm Kruis <[email protected]>
+ *      Reworked based on the i386 ABI spec.
  * 24-Nov-02  Christian Tismer  <[email protected]>
  *      needed to add another magic constant to insure
  *      that f in slp_eval_frame(PyFrameObject *f)
@@ -23,36 +25,84 @@
 
 #ifdef SLP_EVAL
 
-/* #define STACK_MAGIC 3 */
-/* the above works fine with gcc 2.96, but 2.95.3 wants this */
 #define STACK_MAGIC 0
 
-static int
+/*
+ * In order to switch the stack, we use the fact, that the compiler
+ * already knows how to preserve registers accross function calls.
+ *
+ * The relevant i386 ABI specifigation pecisely defines which registers
+ * must be preserved and which registers may be modified.
+ * We use a gcc inline assembly feature to pretend that the inline
+ * assembly block modifies the registers to be preserved. As a result,
+ * the compiler emits code to preserve those registers.
+ *
+ * The "System V Application Binary Interface Intel386 Architecture Processor Supplment, Fourth Edition"
+ * Section 3 chapter "Function Calling Sequence" states:
+ *   All registers on the Intel386 are global and thus visible to both a calling and a
+ *   called function. Registers %ebp, %ebx, %edi, %esi, and %esp "belong" to the cal-
+ *   ling function. In other words, a called function must preserve these registers’
+ *   values for its caller. Remaining registers ‘‘belong’’ to the called function.
+ *
+ * The compiler always preserves the %esp register accros a function call.
+ *
+ * Depending on the usage of a frame pointer, which is optional
+ * for the i386 ABI, the compiler already preserves the %ebp
+ * register. Unfortunately, we must not add "ebp" to the clobber list, if
+ * ebp is used as a frame pointer (won't compile). Therefore we save
+ * ebp manually.
+ *
+ * For the other registers we tell the compiler,
+ * that we are going to clobber the registers. The compiler will then save the registers
+ * for us. (However the compiler gives no guarantee, when it will restore
+ * the registers.) And the compiler only preserves those registers, that must
+ * be preserved according to the calling convention. It does not preserve any other
+ * register, that may be modified during a function call. Therefore specifying additional
+ * registers has no effect at all. Take a look at the generated assembly code!
+ */
+
+/* Registers marked as clobbered, minimum set according to the ABI spec. */
+#define REGS_CLOBBERED "ebx", "edi", "esi"
+
+/*
+ * You may want to make the function static enable optimizations.
+ * However, the ABI SPEC does not apply to static functions. Therefore
+ * I make slp_switch a regular global function.
+ */
+#if 0
+static
+#endif
+int
 slp_switch(void)
 {
     register int *stackref, stsizediff;
 #if STACKLESS_FRHACK
-    __asm__ volatile ("" : : : "esi", "edi");
+    __asm__ volatile (
+	""
+        : : : REGS_CLOBBERED );
 #else
-    __asm__ volatile ("" : : : "ebx", "esi", "edi");
+    void * ebp;
+    __asm__ volatile (
+	"movl %%ebp, %0\n\t"
+        : "=m" (ebp) : : REGS_CLOBBERED );
 #endif
     __asm__ ("movl %%esp, %0" : "=g" (stackref));
     {
         SLP_SAVE_STATE(stackref, stsizediff);
         __asm__ volatile (
-            "addl %0, %%esp\n"
-            "addl %0, %%ebp\n"
+            "addl %0, %%esp\n\t"
+            "addl %0, %%ebp\n\t"
             :
             : "r" (stsizediff)
             );
         SLP_RESTORE_STATE();
+#if ! STACKLESS_FRHACK
+        __asm__ volatile (
+            "movl %0, %%ebp\n\t"
+            : : "m" (ebp) );
+#endif
         return 0;
     }
-#if STACKLESS_FRHACK
-    __asm__ volatile ("" : : : "esi", "edi");
-#else
-    __asm__ volatile ("" : : : "ebx", "esi", "edi");
-#endif
 }