Implement jit_flush

* include/lightning/jit_private.h, lib/jit_aarch64.c,
	lib/jit_alpha.c, lib/jit_arm.c, lib/jit_hppa.c,
	lib/jit_ia64.c, lib/jit_mips.c, lib/jit_ppc.c,
	lib/jit_s390x.c, lib/jit_sparc.c, lib/jit_x86.c:
	Implement a private jit_flush call, that flushes
	the cache, if applicable, aligning down to the
	previous and up to the next page boundary.

ChangeLog

@@ -1,3 +1,13 @@
+2014-11-08 Paulo Andrade <pcpa@gnu.org>
+
+	* include/lightning/jit_private.h, lib/jit_aarch64.c,
+	lib/jit_alpha.c, lib/jit_arm.c, lib/jit_hppa.c,
+	lib/jit_ia64.c, lib/jit_mips.c, lib/jit_ppc.c,
+	lib/jit_s390x.c, lib/jit_sparc.c, lib/jit_x86.c:
+	Implement a private jit_flush call, that flushes
+	the cache, if applicable, aligning down to the
+	previous and up to the next page boundary.
+
 2014-11-08 Paulo Andrade <pcpa@gnu.org>
 
 	* check/ctramp.c: New file. It just repeats the test

include/lightning/jit_private.h

@@ -589,6 +589,9 @@ _jit_regarg_p(jit_state_t*, jit_node_t*, jit_int32_t);
 extern jit_pointer_t
 _emit_code(jit_state_t*);
 
+extern void
+jit_flush(void *fptr, void *tptr);
+
 #define emit_ldxi(r0, r1, i0)	_emit_ldxi(_jit, r0, r1, i0)
 extern void
 _emit_ldxi(jit_state_t*, jit_int32_t, jit_int32_t, jit_word_t);

lib/jit_aarch64.c

@@ -1170,8 +1170,7 @@ _emit_code(jit_state_t *_jit)
 	patch_at(word, value);
     }
 
-    word = sysconf(_SC_PAGE_SIZE);
+    jit_flush(_jit->code.ptr, _jit->pc.uc);
-    __clear_cache(_jit->code.ptr, (void *)((_jit->pc.w + word) & -word));
 
     return (_jit->code.ptr);
 }
@@ -1181,6 +1180,19 @@ _emit_code(jit_state_t *_jit)
 #  include "jit_aarch64-fpu.c"
 #undef CODE
 
+void
+jit_flush(void *fptr, void *tptr)
+{
+#if defined(__GNUC__)
+    jit_word_t		f, t, s;
+
+    s = sysconf(_SC_PAGE_SIZE);
+    f = (jit_word_t)fptr & -s;
+    t = (((jit_word_t)tptr) + s - 1) & -s;
+    __clear_cache((void *)f, (void *)t);
+#endif
+}
+
 void
 _emit_ldxi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
 {

lib/jit_alpha.c

@@ -1213,6 +1213,8 @@ _emit_code(jit_state_t *_jit)
 	patch_at(_jitc->patches.ptr[offset].inst, word);
     }
 
+    jit_flush(_jit->code.ptr, _jit->pc.uc);
+
     return (_jit->code.ptr);
 }
 
@@ -1221,6 +1223,11 @@ _emit_code(jit_state_t *_jit)
 #  include "jit_alpha-fpu.c"
 #undef CODE
 
+void
+jit_flush(void *fptr, void *tptr)
+{
+}
+
 void
 _emit_ldxi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
 {

lib/jit_arm.c

@@ -1589,10 +1589,7 @@ _emit_code(jit_state_t *_jit)
 	patch_at(_jitc->patches.ptr[offset].kind & ~arm_patch_node, word, value);
     }
 
-#if defined(__GNUC__)
+    jit_flush(_jit->code.ptr, _jit->pc.uc);
-    word = sysconf(_SC_PAGE_SIZE);
-    __clear_cache(_jit->code.ptr, (void *)((_jit->pc.w + word) & -word));
-#endif
 
     return (_jit->code.ptr);
 }
@@ -1603,6 +1600,19 @@ _emit_code(jit_state_t *_jit)
 #  include "jit_arm-vfp.c"
 #undef CODE
 
+void
+jit_flush(void *fptr, void *tptr)
+{
+#if defined(__GNUC__)
+    jit_word_t		f, t, s;
+
+    s = sysconf(_SC_PAGE_SIZE);
+    f = (jit_word_t)fptr & -s;
+    t = (((jit_word_t)tptr) + s - 1) & -s;
+    __clear_cache((void *)f, (void *)t);
+#endif
+}
+
 void
 _emit_ldxi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
 {

lib/jit_hppa.c

@@ -1162,6 +1162,23 @@ _emit_code(jit_state_t *_jit)
 	patch_at(_jitc->patches.ptr[offset].inst, word);
     }
 
+    jit_flush(_jit->code.ptr, _jit->pc.uc);
+
+    return (_jit->code.ptr);
+}
+
+#define CODE				1
+#  include "jit_hppa-cpu.c"
+#  include "jit_hppa-fpu.c"
+#undef CODE
+
+void
+jit_flush(void *fptr, void *tptr)
+{
+    jit_word_t		f, t, s;
+    s = sysconf(_SC_PAGE_SIZE);
+    f = (jit_word_t)fptr & -s;
+    t = (((jit_word_t)tptr) + s - 1) & -s;
 #if defined(__hppa)
 /* --- parisc2.0.pdf ---
 		Programming Note
@@ -1203,9 +1220,7 @@ at most two cachelines.
      * on this software.
      */
     {
-	jit_word_t	f = (jit_word_t)_jit->code.ptr;
 	jit_word_t	n = f + 32;
-	jit_word_t	t = f + _jit->code.length;
 	register int	u, v;
 	for (; f <= t; n = f + 32, f += 64) {
 	    asm volatile ("fdc 0(0,%0)"
@@ -1234,18 +1249,10 @@ at most two cachelines.
     }
 #else
     /* This is supposed to work but appears to fail on multiprocessor systems */
-    word = sysconf(_SC_PAGE_SIZE);
+    __clear_cache((void *)f, (void *)t);
-    __clear_cache(_jit->code.ptr, (void *)((_jit->pc.w + word) & -word));
 #endif
-
-    return (_jit->code.ptr);
 }
 
-#define CODE				1
-#  include "jit_hppa-cpu.c"
-#  include "jit_hppa-fpu.c"
-#undef CODE
-
 void
 _emit_ldxi(jit_state_t *_jit, jit_gpr_t r0, jit_gpr_t r1, jit_word_t i0)
 {

lib/jit_ia64.c

@@ -1385,10 +1385,7 @@ _emit_code(jit_state_t *_jit)
 	patch_at(node->code, _jitc->patches.ptr[offset].inst, word);
     }
 
-#if defined(__GNUC__)
+    jit_flush(_jit->code.ptr, _jit->pc.uc);
-    word = sysconf(_SC_PAGE_SIZE);
-    __clear_cache(_jit->code.ptr, (void *)((_jit->pc.w + word) & -word));
-#endif
 
     return (_jit->code.ptr);
 }
@@ -1398,6 +1395,19 @@ _emit_code(jit_state_t *_jit)
 #  include "jit_ia64-fpu.c"
 #undef CODE
 
+void
+jit_flush(void *fptr, void *tptr)
+{
+#if defined(__GNUC__)
+    jit_word_t		f, t, s;
+
+    s = sysconf(_SC_PAGE_SIZE);
+    f = (jit_word_t)fptr & -s;
+    t = (((jit_word_t)tptr) + s - 1) & -s;
+    __clear_cache((void *)f, (void *)t);
+#endif
+}
+
 /* Use r2 that is reserved to not require a jit_get_reg call, also note
  * that addil needs a register that first in 2 bits, so, if using a
  * register other than r2 must be less than r8 (or check for a smaller

lib/jit_mips.c

@@ -1488,9 +1488,7 @@ _emit_code(jit_state_t *_jit)
 	patch_at(_jitc->patches.ptr[offset].inst, word);
     }
 
-#if defined(__linux__)
+    jit_flush(_jit->code.ptr, _jit->pc.uc);
-    _flush_cache((char *)_jit->code.ptr, _jit->pc.uc - _jit->code.ptr, ICACHE);
-#endif
 
     return (_jit->code.ptr);
 }
@@ -1500,6 +1498,19 @@ _emit_code(jit_state_t *_jit)
 #  include "jit_mips-fpu.c"
 #undef CODE
 
+void
+jit_flush(void *fptr, void *tptr)
+{
+#if defined(__linux__)
+    jit_word_t		f, t, s;
+
+    s = sysconf(_SC_PAGE_SIZE);
+    f = (jit_word_t)fptr & -s;
+    t = (((jit_word_t)tptr) + s - 1) & -s;
+    _flush_cache((void *)f, t - f, ICACHE);
+#endif
+}
+
 void
 _emit_ldxi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
 {

lib/jit_ppc.c

@@ -1372,8 +1372,7 @@ _emit_code(jit_state_t *_jit)
 	patch_at(_jitc->patches.ptr[offset].inst, word);
     }
 
-    word = sysconf(_SC_PAGE_SIZE);
+    jit_flush(_jit->code.ptr, _jit->pc.uc);
-    __clear_cache(_jit->code.ptr, (void *)((_jit->pc.w + word) & -word));
 
     return (_jit->code.ptr);
 }
@@ -1383,6 +1382,19 @@ _emit_code(jit_state_t *_jit)
 #  include "jit_ppc-fpu.c"
 #undef CODE
 
+void
+jit_flush(void *fptr, void *tptr)
+{
+#if defined(__GNUC__)
+    jit_word_t		f, t, s;
+
+    s = sysconf(_SC_PAGE_SIZE);
+    f = (jit_word_t)fptr & -s;
+    t = (((jit_word_t)tptr) + s - 1) & -s;
+    __clear_cache((void *)f, (void *)t);
+#endif
+}
+
 void
 _emit_ldxi(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1, jit_word_t i0)
 {

lib/jit_s390x.c

@@ -1154,8 +1154,7 @@ _emit_code(jit_state_t *_jit)
 	patch_at(_jitc->patches.ptr[offset].inst, word);
     }
 
-    word = sysconf(_SC_PAGE_SIZE);
+    jit_flush(_jit->code.ptr, _jit->pc.uc);
-    __clear_cache(_jit->code.ptr, (void *)((_jit->pc.w + word) & -word));
 
     return (_jit->code.ptr);
 }
@@ -1165,6 +1164,19 @@ _emit_code(jit_state_t *_jit)
 #  include "jit_s390x-fpu.c"
 #undef CODE
 
+void
+jit_flush(void *fptr, void *tptr)
+{
+#if defined(__GNUC__)
+    jit_word_t		f, t, s;
+
+    s = sysconf(_SC_PAGE_SIZE);
+    f = (jit_word_t)fptr & -s;
+    t = (((jit_word_t)tptr) + s - 1) & -s;
+    __clear_cache((void *)f, (void *)t);
+#endif
+}
+
 void
 _emit_ldxi(jit_state_t *_jit, jit_gpr_t r0, jit_gpr_t r1, jit_word_t i0)
 {

lib/jit_sparc.c

@@ -1156,6 +1156,8 @@ _emit_code(jit_state_t *_jit)
 	patch_at(_jitc->patches.ptr[offset].inst, word);
     }
 
+    jit_flush(_jit->code.ptr, _jit->pc.uc);
+
     return (_jit->code.ptr);
 }
 
@@ -1164,6 +1166,11 @@ _emit_code(jit_state_t *_jit)
 #  include "jit_sparc-fpu.c"
 #undef CODE
 
+void
+jit_flush(void *fptr, void *tptr)
+{
+}
+
 void
 _emit_ldxi(jit_state_t *_jit, jit_gpr_t r0, jit_gpr_t r1, jit_word_t i0)
 {

lib/jit_x86.c

@@ -1794,6 +1794,8 @@ _emit_code(jit_state_t *_jit)
 	patch_at(node, _jitc->patches.ptr[offset].inst, word);
     }
 
+    jit_flush(_jit->code.ptr, _jit->pc.uc);
+
     return (_jit->code.ptr);
 }
 
@@ -1803,6 +1805,11 @@ _emit_code(jit_state_t *_jit)
 #  include "jit_x86-x87.c"
 #undef CODE
 
+void
+jit_flush(void *fptr, void *tptr)
+{
+}
+
 void
 _emit_ldxi(jit_state_t *_jit, jit_gpr_t r0, jit_gpr_t r1, jit_word_t i0)
 {