add --enable-jitframe for configure on x86_64

Using a frame pointer for the ABI of internal helper functions
should make the stack friendlier to tools like `perf`. There
may be a small performance cost, though.

racket/src/configure

@@ -832,6 +832,7 @@ enable_gprof
 enable_gcov
 enable_noopt
 enable_ubsan
+enable_jitframe
 '
       ac_precious_vars='build_alias
 host_alias
@@ -1489,6 +1490,7 @@ Optional Features:
   --enable-strip          strip debug on install (usually enabled by default)
   --enable-noopt          drop -O C flags (useful for low-level debugging)
   --enable-ubsan          compile with -fsanitize=undefined)
+  --enable-jitframe       x86_64: use frame pointer for internal calls
 
 Some influential environment variables:
   CC          C compiler command
@@ -2843,6 +2845,11 @@ if test "${enable_ubsan+set}" = set; then :
   enableval=$enable_ubsan;
 fi
 
+# Check whether --enable-jitframe was given.
+if test "${enable_jitframe+set}" = set; then :
+  enableval=$enable_jitframe;
+fi
+
 
 ###### Some flags imply other flags #######
 
@@ -3096,6 +3103,8 @@ show_explicitly_enabled "${enable_smalloskit}" "OSKit small mode"
 show_explicitly_enabled "${enable_gprof}" "gprof"
 show_explicitly_enabled "${enable_gcov}" "gcov"
 
+show_explicitly_enabled "${enable_jitframe}" "jitframe"
+
 show_explicitly_enabled "${enable_noopt}" "No-optimization" "Note that this mode is intended only for debugging purposes"
 show_explicitly_enabled "${enable_strip}" "Debug-symbol stripping"
 show_explicitly_disabled "${enable_strip}" "Debug-symbol stripping"
@@ -4579,6 +4588,10 @@ if test "${enable_gcov}" = "yes" ; then
   PROFFLAGS="$PROFFLAGS -fprofile-arcs -ftest-coverage"
 fi
 
+if test "${enable_jitframe}" = "yes" ; then
+  MZOPTIONS="$MZOPTIONS -DMZ_PROLOG_CREATE_FULL_STACK_FRAME"
+fi
+
 ORIG_CC_FOR_BUILD="${CC_FOR_BUILD}"
 if test "$CC_FOR_BUILD" = "" ; then
   CC_FOR_BUILD="$CC"

racket/src/racket/configure.ac

@@ -89,6 +89,7 @@ AC_ARG_ENABLE(gcov,    [  --enable-gcov           compile to gather gcov statist
 AC_ARG_ENABLE(noopt,   [  --enable-strip          strip debug on install (usually enabled by default)])
 AC_ARG_ENABLE(noopt,   [  --enable-noopt          drop -O C flags (useful for low-level debugging)])
 AC_ARG_ENABLE(ubsan,   [  --enable-ubsan          compile with -fsanitize=undefined)])
+AC_ARG_ENABLE(jitframe,[  --enable-jitframe       x86_64: use frame pointer for internal calls])
 
 ###### Some flags imply other flags #######
 
@@ -342,6 +343,8 @@ show_explicitly_enabled "${enable_smalloskit}" "OSKit small mode"
 show_explicitly_enabled "${enable_gprof}" "gprof"
 show_explicitly_enabled "${enable_gcov}" "gcov"
 
+show_explicitly_enabled "${enable_jitframe}" "jitframe"
+
 show_explicitly_enabled "${enable_noopt}" "No-optimization" "Note that this mode is intended only for debugging purposes"
 show_explicitly_enabled "${enable_strip}" "Debug-symbol stripping"
 show_explicitly_disabled "${enable_strip}" "Debug-symbol stripping"
@@ -678,6 +681,10 @@ if test "${enable_gcov}" = "yes" ; then
   PROFFLAGS="$PROFFLAGS -fprofile-arcs -ftest-coverage"
 fi
 
+if test "${enable_jitframe}" = "yes" ; then
+  MZOPTIONS="$MZOPTIONS -DMZ_PROLOG_CREATE_FULL_STACK_FRAME"
+fi
+
 ORIG_CC_FOR_BUILD="${CC_FOR_BUILD}"
 if test "$CC_FOR_BUILD" = "" ; then 
   CC_FOR_BUILD="$CC"

racket/src/racket/src/jit.h

@@ -820,7 +820,7 @@ static void *top;
       LOCAL1 (which is a cont_mark_stack offset, if anything)
       LOCAL2 (some pointer, never to stack or runstack)
       LOCAL3 (temp space for misc uses; not saved across calls that might capture LWC)
-      LOCAL4 (x86_64: = saved R14 otherwise when THREAD_LOCAL
+      LOCAL4 (x86_64: = saved R14 when THREAD_LOCAL
               x86: = RUNSTACK_BASE or THREAD_LOCAL)
       [some empty slots, maybe, depending on alignment]
       [space for "flostack" --- local unboxed values, such as flonums]
@@ -944,9 +944,6 @@ void scheme_jit_prolog_again(mz_jit_state *jitter, int n, int ret_addr_reg)
 #  else
 #   define STACK_ALIGN_WORDS 3
 #  endif
-#  define mz_prolog(x) (ADDQiBr(-(STACK_ALIGN_WORDS * JIT_WORD_SIZE), JIT_SP))
-#  define mz_epilog_without_jmp() ADDQiBr((STACK_ALIGN_WORDS + 1) * JIT_WORD_SIZE, JIT_SP)
-#  define mz_epilog(x) (ADDQiBr(STACK_ALIGN_WORDS * JIT_WORD_SIZE, JIT_SP), RET_())
 #  define JIT_LOCAL3 -(JIT_WORD_SIZE * 6)
 #  ifdef NEED_LOCAL4
 #   ifdef JIT_X86_64
@@ -958,10 +955,10 @@ void scheme_jit_prolog_again(mz_jit_state *jitter, int n, int ret_addr_reg)
 #  else
 #   define LOCAL_FRAME_SIZE 3
 #  endif
+#  define _mz_prolog(x) (ADDQiBr(-(STACK_ALIGN_WORDS * JIT_WORD_SIZE), JIT_SP))
+#  define _mz_epilog_without_jmp() ADDQiBr((STACK_ALIGN_WORDS + 1) * JIT_WORD_SIZE, JIT_SP)
+#  define _mz_epilog(x) (ADDQiBr(STACK_ALIGN_WORDS * JIT_WORD_SIZE, JIT_SP), RET_())
 # else
-#  define mz_prolog(x) /* empty */
-#  define mz_epilog(x) RET_()
-#  define mz_epilog_without_jmp() ADDQir(JIT_WORD_SIZE, JIT_SP)
 #  define JIT_LOCAL3 JIT_LOCAL2
 #  ifdef NEED_LOCAL4
 #   define LOCAL_FRAME_SIZE 3
@@ -969,10 +966,29 @@ void scheme_jit_prolog_again(mz_jit_state *jitter, int n, int ret_addr_reg)
 #  else
 #   define LOCAL_FRAME_SIZE 2
 #  endif
+#  define _mz_prolog(x) /* empty */
+#  define _mz_epilog(x) RET_()
+#  define _mz_epilog_without_jmp() ADDQir(JIT_WORD_SIZE, JIT_SP)
 # endif
 # ifdef NEED_LOCAL4
 #   define JIT_LOCAL4 -(JIT_WORD_SIZE * JIT_LOCAL4_OFFSET)
 # endif
+# ifdef MZ_PROLOG_CREATE_FULL_STACK_FRAME
+  /* Make the internal ABI the same as the main call ABI */
+#  define MZ_LOCAL_FRAME_SIZE 8
+#  define mz_prolog(x) (PUSHQr(_EBP),                                  \
+                        mz_get_local_p((x), JIT_LOCAL3),               \
+                        MOVQrr(_ESP, _EBP),                            \
+                        ADDQiBr(-(MZ_LOCAL_FRAME_SIZE * JIT_WORD_SIZE), JIT_SP), \
+                        mz_set_local_p((x), JIT_LOCAL3))
+#  define mz_epilog_without_jmp() (ADDQiBr(MZ_LOCAL_FRAME_SIZE * JIT_WORD_SIZE, JIT_SP), POPQr(_EBP), ADDQiBr(JIT_WORD_SIZE, JIT_SP))
+#  define mz_epilog(x) (ADDQiBr(MZ_LOCAL_FRAME_SIZE * JIT_WORD_SIZE, JIT_SP), POPQr(_EBP), RET_())
+# else
+  /* Normal internal ABI */
+#  define mz_prolog(x) _mz_prolog(x)
+#  define mz_epilog_without_jmp(x) _mz_epilog_without_jmp(x)
+#  define mz_epilog(x) _mz_epilog(x)
+# endif
 # define mz_push_locals() SUBQir((LOCAL_FRAME_SIZE << JIT_LOG_WORD_SIZE), JIT_SP)
 # define mz_pop_locals() ADDQir((LOCAL_FRAME_SIZE << JIT_LOG_WORD_SIZE), JIT_SP)
 # define JIT_FRAME_FLOSTACK_OFFSET (-(JIT_WORD_SIZE * (LOCAL_FRAME_SIZE + 3)))

racket/src/racket/src/jitstack.c

@@ -380,17 +380,23 @@ Scheme_Object *scheme_native_stack_trace(void)
 
       if (STK_COMP((uintptr_t)np, real_stack_end)
 	  && STK_COMP(stack_start, (uintptr_t)np)) {
-	if (SCHEME_VOIDP(name)) {
+ 	if (SCHEME_VOIDP(name)) {
 	  /* JIT_LOCAL2 has the next return address (always) */
 	  q = ((void **)np)[JIT_LOCAL2 >> JIT_LOG_WORD_SIZE];
 	} else {
-#ifdef MZ_USE_JIT_I386
+#ifdef MZ_PROLOG_CREATE_FULL_STACK_FRAME
+          /* np is the actual stack frame */
+          p = np;
+          q = ((void **)p)[RETURN_ADDRESS_OFFSET];
+#else
+# ifdef MZ_USE_JIT_I386
 	  /* Push after local stack of return-address proc
 	     has the next return address */
 	  q = ((void **)np)[-(3 + LOCAL_FRAME_SIZE + 1)];
-#else
+# else
           /* JIT_LOCAL2 has the next return address */
           q = ((void **)np)[JIT_LOCAL2 >> JIT_LOG_WORD_SIZE];
+# endif
 #endif
 	}
       } else {