Update lightening

* libguile/lightening: Merge gitlab.com/wingo/lightening/ from
  3260f7deeb.

libguile/lightening/.gitlab-ci.yml

@@ -1,4 +1,4 @@
-image: debian:testing
+image: debian:stable
 
 before_script:
   - dpkg --add-architecture i386
@@ -30,4 +30,9 @@ aarch64:
 armhf:
   stage: test
   script:
-    - make -C tests test-armv7 CC_ARMv7=arm-linux-gnueabihf-gcc
+    - make -C tests test-armv7 CC_ARMv7="arm-linux-gnueabihf-gcc -marm"
+
+armhf-thumb:
+  stage: test
+  script:
+    - make -C tests test-armv7 CC_ARMv7="arm-linux-gnueabihf-gcc -mthumb"

libguile/lightening/lightening.h

@@ -91,6 +91,8 @@ enum jit_reloc_kind
   JIT_RELOC_JCC_WITH_VENEER,
   JIT_RELOC_LOAD_FROM_POOL,
 #endif
+  JIT_RELOC_MASK = 15,
+  JIT_RELOC_FLAG_0 = 16,
 };
 
 typedef struct jit_reloc

libguile/lightening/lightening/aarch64.c

@@ -90,9 +90,9 @@ DEFINE_ENCODER(size, 2, 22, unsigned, uint32_t)
   {                                                                     \
     return read_signed_bitfield(*loc, kind##_width, kind##_shift);      \
   }                                                                     \
-  static int offset_in_##name##_range(ptrdiff_t diff) maybe_unused;     \
+  static int offset_in_##name##_range(ptrdiff_t diff, int flags) maybe_unused; \
   static int                                                            \
-  offset_in_##name##_range(ptrdiff_t diff)                              \
+  offset_in_##name##_range(ptrdiff_t diff, int flags)                   \
   {                                                                     \
     return in_signed_range(diff, kind##_width);                         \
   }                                                                     \
@@ -114,8 +114,12 @@ DEFINE_ENCODER(size, 2, 22, unsigned, uint32_t)
     }                                                                   \
   }
 
-DEFINE_PATCHABLE_INSTRUCTION(jmp, simm26, JMP_WITH_VENEER, 2);
+#define DEFINE_PATCHABLE_INSTRUCTIONS(name, kind, RELOC, rsh)           \
-DEFINE_PATCHABLE_INSTRUCTION(jcc, simm19, JCC_WITH_VENEER, 2);
+    DEFINE_PATCHABLE_INSTRUCTION(name, kind, RELOC, rsh);               \
+    DEFINE_PATCHABLE_INSTRUCTION(veneer_##name, kind, RELOC, rsh);
+
+DEFINE_PATCHABLE_INSTRUCTIONS(jmp, simm26, JMP_WITH_VENEER, 2);
+DEFINE_PATCHABLE_INSTRUCTIONS(jcc, simm19, JCC_WITH_VENEER, 2);
 DEFINE_PATCHABLE_INSTRUCTION(load_from_pool, simm19, LOAD_FROM_POOL, 2);
 
 struct veneer

libguile/lightening/lightening/arm-cpu.c

@@ -186,6 +186,8 @@
 
 #define _NOREG (jit_gpr_regno(_PC))
 
+#define JIT_RELOC_B JIT_RELOC_FLAG_0
+
 static void
 emit_wide_thumb(jit_state_t *_jit, uint32_t inst)
 {
@@ -265,9 +267,12 @@ write_wide_thumb(uint32_t *loc, uint32_t v)
 }
 
 static int
-offset_in_jmp_range(int32_t offset)
+offset_in_jmp_range(int32_t offset, int flags)
 {
-  return -0x800000 <= offset && offset <= 0x7fffff;
+  if (!(offset & 1) && flags | JIT_RELOC_B)
+    return 0;
+  else
+    return -0x1000000 <= offset && offset <= 0xffffff;
 }
 
 static int32_t
@@ -287,7 +292,7 @@ decode_thumb_jump(uint32_t v)
   ret |= i2 << 21;
   ret |= hi << 11;
   ret |= lo;
-  return ret;
+  return ret << 1;
 }
 
 static const uint32_t thumb_jump_mask = 0xf800d000;
@@ -295,14 +300,15 @@ static const uint32_t thumb_jump_mask = 0xf800d000;
 static uint32_t
 encode_thumb_jump(int32_t v)
 {
-  ASSERT(offset_in_jmp_range(v));
+  ASSERT(offset_in_jmp_range(v, 0));
+  v >>= 1;
   uint32_t s  = !!(v & 0x800000);
   uint32_t i1 = !!(v & 0x400000);
   uint32_t i2 = !!(v & 0x200000);
   uint32_t j1 = s ? i1 : !i1;
   uint32_t j2 = s ? i2 : !i2;
   uint32_t ret = (s<<26)|((v&0x1ff800)<<5)|(j1<<13)|(j2<<11)|(v&0x7ff);
-  ASSERT(decode_thumb_jump(ret) == v);
+  ASSERT(decode_thumb_jump(ret) == v << 1);
   ASSERT((ret & thumb_jump_mask) == 0);
   return ret;
 }
@@ -310,7 +316,13 @@ encode_thumb_jump(int32_t v)
 static uint32_t
 patch_thumb_jump(uint32_t inst, int32_t v)
 {
-  return (inst & thumb_jump_mask) | encode_thumb_jump(v);
+  inst &= thumb_jump_mask;
+  if (!(v & 1)) {
+    ASSERT(inst == THUMB2_BLI || inst == THUMB2_BLXI);
+    v = (v + 2) & ~2;
+    inst = THUMB2_BLXI;
+  }
+  return inst | encode_thumb_jump(v);
 }
 
 static int32_t
@@ -325,15 +337,23 @@ patch_jmp_offset(uint32_t *loc, int32_t v)
   write_wide_thumb(loc, patch_thumb_jump(read_wide_thumb(loc), v));
 }
 
+static void
+patch_veneer_jmp_offset(uint32_t *loc, int32_t v)
+{
+  ASSERT(!(v & 1));
+  patch_jmp_offset(loc, v | 1);
+}
+
 static jit_reloc_t
 emit_thumb_jump(jit_state_t *_jit, uint32_t inst)
 {
   while (1) {
     uint8_t *pc_base = _jit->pc.uc + 4;
-    uint8_t rsh = 1;
+    int32_t off = (uint8_t*)jit_address(_jit) - pc_base;
-    int32_t off = (_jit->pc.uc - pc_base) >> rsh;
+    enum jit_reloc_kind kind = JIT_RELOC_JMP_WITH_VENEER;
-    jit_reloc_t ret =
+    if (inst == THUMB2_B)
-      jit_reloc (_jit, JIT_RELOC_JMP_WITH_VENEER, 0, _jit->pc.uc, pc_base, rsh);
+      kind |= JIT_RELOC_B;
+    jit_reloc_t ret = jit_reloc (_jit, kind, 0, _jit->pc.uc, pc_base, 0);
     uint8_t thumb_jump_width = 24;
     if (add_pending_literal(_jit, ret, thumb_jump_width - 1)) {
       emit_wide_thumb(_jit, patch_thumb_jump(inst, off));
@@ -343,9 +363,12 @@ emit_thumb_jump(jit_state_t *_jit, uint32_t inst)
 }
 
 static int
-offset_in_jcc_range(int32_t v)
+offset_in_jcc_range(int32_t v, int flags)
 {
-  return -0x80000 <= v && v <= 0x7ffff;
+  if (!(v & 1))
+    return 0;
+  else
+    return -0x100000 <= v && v <= 0xfffff;
 }
 
 static int32_t
@@ -363,7 +386,7 @@ decode_thumb_cc_jump(uint32_t v)
   ret |= j1 << 17;
   ret |= hi << 11;
   ret |= lo;
-  return ret;
+  return ret << 1;
 }
 
 static const uint32_t thumb_cc_jump_mask = 0xfbc0d000;
@@ -371,14 +394,15 @@ static const uint32_t thumb_cc_jump_mask = 0xfbc0d000;
 static uint32_t
 encode_thumb_cc_jump(int32_t v)
 {
-  ASSERT(offset_in_jcc_range(v));
+  ASSERT(offset_in_jcc_range(v, 0));
+  v >>= 1;
   uint32_t s  = !!(v & 0x80000);
   uint32_t j2 = !!(v & 0x40000);
   uint32_t j1 = !!(v & 0x20000);
   uint32_t hi = (v >> 11) & 0x3f;
   uint32_t lo = v & 0x7ff;
   uint32_t ret = (s<<26)|(hi << 16)|(j1<<13)|(j2<<11)|lo;
-  ASSERT(decode_thumb_cc_jump(ret) == v);
+  ASSERT(decode_thumb_cc_jump(ret) == v << 1);
   ASSERT((ret & thumb_cc_jump_mask) == 0);
   return ret;
 }
@@ -401,15 +425,21 @@ patch_jcc_offset(uint32_t *loc, int32_t v)
   write_wide_thumb(loc, patch_thumb_cc_jump(read_wide_thumb(loc), v));
 }
 
+static void
+patch_veneer_jcc_offset(uint32_t *loc, int32_t v)
+{
+  ASSERT(!(v & 1));
+  patch_jcc_offset(loc, v | 1);
+}
+
 static jit_reloc_t
 emit_thumb_cc_jump(jit_state_t *_jit, uint32_t inst)
 {
   while (1) {
     uint8_t *pc_base = _jit->pc.uc + 4;
-    uint8_t rsh = 1;
+    int32_t off = (uint8_t*)jit_address(_jit) - pc_base;
-    int32_t off = (_jit->pc.uc - pc_base) >> rsh;
     jit_reloc_t ret =
-      jit_reloc (_jit, JIT_RELOC_JCC_WITH_VENEER, 0, _jit->pc.uc, pc_base, rsh);
+      jit_reloc (_jit, JIT_RELOC_JCC_WITH_VENEER, 0, _jit->pc.uc, pc_base, 0);
     uint8_t thumb_cc_jump_width = 20;
     if (add_pending_literal(_jit, ret, thumb_cc_jump_width - 1)) {
       emit_wide_thumb(_jit, patch_thumb_cc_jump(inst, off));
@@ -1028,12 +1058,6 @@ T2_BLI(jit_state_t *_jit)
   return tb(_jit, THUMB2_BLI);
 }
 
-static jit_reloc_t
-T2_BLXI(jit_state_t *_jit)
-{
-  return tb(_jit, THUMB2_BLXI);
-}
-
 enum dmb_option { DMB_ISH = 0xb };
 static void
 T1_DMB(jit_state_t *_jit, enum dmb_option option)
@@ -2052,7 +2076,7 @@ rshi_u(jit_state_t *_jit, int32_t r0, int32_t r1, jit_word_t i0)
 static void
 jmpr(jit_state_t *_jit, int32_t r0)
 {
-  T1_MOV(_jit, jit_gpr_regno(_PC), r0);
+  T1_BX(_jit, r0);
 }
 
 static jit_reloc_t
@@ -2064,8 +2088,6 @@ jmp(jit_state_t *_jit)
 static void
 jmpi(jit_state_t *_jit, jit_word_t i0)
 {
-  /* Strip thumb bit, if any.  */
-  i0 &= ~1;
   return jit_patch_there(_jit, jmp(_jit), (void*)i0);
 }
 
@@ -2905,10 +2927,7 @@ callr(jit_state_t *_jit, int32_t r0)
 static void
 calli(jit_state_t *_jit, jit_word_t i0)
 {
-  if (i0 & 1)
+  jit_patch_there(_jit, T2_BLI(_jit), (void*)i0);
-    jit_patch_there(_jit, T2_BLI(_jit), (void*)(i0 & ~1));
-  else
-    jit_patch_there(_jit, T2_BLXI(_jit), (void*)i0);
 }
 
 static void
@@ -2990,8 +3009,7 @@ static void
 patch_jmp_without_veneer(jit_state_t *_jit, uint32_t *loc)
 {
   uint8_t *pc_base = ((uint8_t *)loc) + 4;
-  uint8_t rsh = 1;
+  int32_t off = (uint8_t*)jit_address(_jit) - pc_base;
-  int32_t off = (_jit->pc.uc - pc_base) >> rsh;
   write_wide_thumb(loc, THUMB2_B | encode_thumb_jump(off));
 }
 
@@ -3014,11 +3032,10 @@ emit_veneer(jit_state_t *_jit, jit_pointer_t target)
 {
   uint16_t thumb1_ldr = 0x4800;
   int32_t tmp = jit_gpr_regno(JIT_TMP1);
-  int32_t rd = jit_gpr_regno(_PC);
   ASSERT(tmp < 8);
   // Loaded addr is 4 bytes after the LDR, which is aligned, so offset is 0.
   emit_u16(_jit, thumb1_ldr | (tmp << 8));
-  emit_u16(_jit, THUMB_MOV|((_u4(rd)&8)<<4)|(_u4(tmp)<<3)|(rd&7));
+  emit_u16(_jit, THUMB_BX|(_u4(tmp)<<3));
   emit_u32(_jit, (uint32_t) target);
 }
 

libguile/lightening/lightening/lightening.c

@@ -105,11 +105,13 @@ enum guard_pool { GUARD_NEEDED, NO_GUARD_NEEDED };
 static void emit_literal_pool(jit_state_t *_jit, enum guard_pool guard);
 
 static int32_t read_jmp_offset(uint32_t *loc);
-static int offset_in_jmp_range(ptrdiff_t offset);
+static int offset_in_jmp_range(ptrdiff_t offset, int flags);
 static void patch_jmp_offset(uint32_t *loc, ptrdiff_t offset);
+static void patch_veneer_jmp_offset(uint32_t *loc, ptrdiff_t offset);
 static int32_t read_jcc_offset(uint32_t *loc);
-static int offset_in_jcc_range(ptrdiff_t offset);
+static int offset_in_jcc_range(ptrdiff_t offset, int flags);
 static void patch_jcc_offset(uint32_t *loc, ptrdiff_t offset);
+static void patch_veneer_jcc_offset(uint32_t *loc, ptrdiff_t offset);
 static void patch_veneer(uint32_t *loc, jit_pointer_t addr);
 static int32_t read_load_from_pool_offset(uint32_t *loc);
 #endif
@@ -169,7 +171,7 @@ jit_pointer_t
 jit_address(jit_state_t *_jit)
 {
   ASSERT (_jit->start);
-  return _jit->pc.uc;
+  return jit_address_to_function_pointer (_jit->pc.uc);
 }
 
 void
@@ -378,8 +380,9 @@ jit_patch_there(jit_state_t* _jit, jit_reloc_t reloc, jit_pointer_t addr)
   ptrdiff_t diff = (uint8_t*)addr - pc_base;
   ASSERT((diff & ((1 << reloc.rsh) - 1)) == 0);
   diff >>= reloc.rsh;
+  int flags = reloc.kind & ~JIT_RELOC_MASK;
 
-  switch (reloc.kind)
+  switch (reloc.kind & JIT_RELOC_MASK)
     {
     case JIT_RELOC_ABSOLUTE:
       if (sizeof(diff) == 4)
@@ -404,7 +407,7 @@ jit_patch_there(jit_state_t* _jit, jit_reloc_t reloc, jit_pointer_t addr)
       uint8_t *target = pc_base + (voff << reloc.rsh);
       if (target == loc.uc) {
         // PC still in range to reify direct branch.
-        if (offset_in_jmp_range(diff)) {
+        if (offset_in_jmp_range(diff, flags)) {
           // Target also in range: reify direct branch.
           patch_jmp_offset(loc.ui, diff);
           remove_pending_literal(_jit, reloc);
@@ -423,7 +426,7 @@ jit_patch_there(jit_state_t* _jit, jit_reloc_t reloc, jit_pointer_t addr)
       int32_t voff = read_jcc_offset(loc.ui);
       uint8_t *target = pc_base + (voff << reloc.rsh);
       if (target == loc.uc) {
-        if (offset_in_jcc_range(diff)) {
+        if (offset_in_jcc_range(diff, flags)) {
           patch_jcc_offset(loc.ui, diff);
           remove_pending_literal(_jit, reloc);
         } else {
@@ -1365,13 +1368,13 @@ emit_literal_pool(jit_state_t *_jit, enum guard_pool guard)
     if (_jit->overflow)
       return;
 
-    switch (entry->reloc.kind) {
+    switch (entry->reloc.kind & JIT_RELOC_MASK) {
     case JIT_RELOC_JMP_WITH_VENEER:
-      patch_jmp_offset((uint32_t*) loc, diff);
+      patch_veneer_jmp_offset((uint32_t*) loc, diff);
       emit_veneer(_jit, (void*) (uintptr_t) entry->value);
       break;
     case JIT_RELOC_JCC_WITH_VENEER:
-      patch_jcc_offset((uint32_t*) loc, diff);
+      patch_veneer_jcc_offset((uint32_t*) loc, diff);
       emit_veneer(_jit, (void*) (uintptr_t) entry->value);
       break;
     case JIT_RELOC_LOAD_FROM_POOL:

libguile/lightening/tests/jmpi_local.c

@@ -0,0 +1,25 @@
+#include "test.h"
+
+static void
+run_test(jit_state_t *j, uint8_t *arena_base, size_t arena_size)
+{
+  jit_begin(j, arena_base, arena_size);
+
+  jit_reloc_t r = jit_jmp (j);
+  jit_reti (j, 0);
+  jit_pointer_t addr = jit_address (j);
+  jit_reti (j, 1);
+  jit_patch_here (j, r);
+  jit_jmpi (j, addr);
+  jit_reti (j, 2);
+
+  int (*f)(void) = jit_end(j, NULL);
+
+  ASSERT(f() == 1);
+}
+
+int
+main (int argc, char *argv[])
+{
+  return main_helper(argc, argv, run_test);
+}