Pack operands into the instruction word

On 64-bit machines where the C code is always at address below 4Gb,
pack one or more operands into the instruction word.

3 files changed, 67 insertions, 21 deletions

diff --git a/erts/emulator/beam/beam_debug.c b/erts/emulator/beam/beam_debug.c
index d8303a651a..70078c8c59 100644
--- a/erts/emulator/beam/beam_debug.c
+++ b/erts/emulator/beam/beam_debug.c
@@ -201,7 +201,7 @@ void debug_dump_code(BeamInstr *I, int num)
 	for (i = 0; i < NUM_SPECIFIC_OPS; i++) {
 	    if (BeamIsOpCode(instr, i) && opc[i].name[0] != '\0') {
 		code_ptr += print_op(ERTS_PRINT_DSBUF, (void *) dsbufp,
-				     i, opc[i].sz-1, code_ptr+1) + 1;
+				     i, opc[i].sz-1, code_ptr) + 1;
 		break;
 	    }
 	}
@@ -321,7 +321,7 @@ erts_debug_disassemble_1(BIF_ALIST_1)
     for (i = 0; i < NUM_SPECIFIC_OPS; i++) {
 	if (BeamIsOpCode(instr, i) && opc[i].name[0] != '\0') {
 	    code_ptr += print_op(ERTS_PRINT_DSBUF, (void *) dsbufp,
-				 i, opc[i].sz-1, code_ptr+1) + 1;
+				 i, opc[i].sz-1, code_ptr) + 1;
 	    break;
 	}
     }
@@ -405,8 +405,11 @@ print_op(fmtfn_t to, void *to_arg, int op, int size, BeamInstr* addr)
 	 * Avoid copying because instructions containing bignum operands
 	 * are bigger than actually declared.
 	 */
-	ap = (BeamInstr *) addr;
+        addr++;
+        ap = addr;
     } else {
+        BeamInstr instr_word = addr++[0];
+
 	/*
 	 * Copy all arguments to a local buffer for the unpacking.
 	 */
@@ -431,23 +434,22 @@ print_op(fmtfn_t to, void *to_arg, int op, int size, BeamInstr* addr)
 	    case 'q':
 		*ap++ = *--sp;
 		break;
-	    case 'i':		/* Initialize packing accumulator. */
-		*ap++ = packed;
-		break;
-	    case 's':
-		*ap++ = packed & 0x3ff;
-		packed >>= 10;
+#ifdef ARCH_64
+	    case '1':		/* Tightest shift */
+		*ap++ = (packed & BEAM_TIGHTEST_MASK) << 3;
+		packed >>= BEAM_TIGHTEST_SHIFT;
 		break;
-	    case '0':		/* Tight shift */
+#endif
+	    case '2':		/* Tight shift */
 		*ap++ = packed & BEAM_TIGHT_MASK;
 		packed >>= BEAM_TIGHT_SHIFT;
 		break;
-	    case '6':		/* Shift 16 steps */
+	    case '3':		/* Loose shift */
 		*ap++ = packed & BEAM_LOOSE_MASK;
 		packed >>= BEAM_LOOSE_SHIFT;
 		break;
 #ifdef ARCH_64
-	    case 'w':		/* Shift 32 steps */
+	    case '4':		/* Shift 32 steps */
 		*ap++ = packed & BEAM_WIDE_MASK;
 		packed >>= BEAM_WIDE_SHIFT;
 		break;
@@ -458,8 +460,18 @@ print_op(fmtfn_t to, void *to_arg, int op, int size, BeamInstr* addr)
 	    case 'P':
 		packed = *--sp;
 		break;
+#if defined(ARCH_64) && defined(CODE_MODEL_SMALL)
+            case '#':       /* -1 */
+            case '$':       /* -2 */
+            case '%':       /* -3 */
+            case '&':       /* -4 */
+            case '\'':      /* -5 */
+            case '(':       /* -6 */
+                packed = (packed << BEAM_WIDE_SHIFT) | BeamExtraData(instr_word);
+		break;
+#endif
 	    default:
-		ASSERT(0);
+                erts_exit(ERTS_ERROR_EXIT, "beam_debug: invalid packing op: %c\n", *prog);
 	    }
 	}
 	ap = args;
diff --git a/erts/emulator/beam/beam_emu.c b/erts/emulator/beam/beam_emu.c
index 6a45087a34..73c4e3532b 100644
--- a/erts/emulator/beam/beam_emu.c
+++ b/erts/emulator/beam/beam_emu.c
@@ -234,15 +234,18 @@ void** beam_ops;
 #define fb(N) ((Sint)(Sint32)(N))
 #define jb(N) ((Sint)(Sint32)(N))
 #define tb(N) (N)
-#define xb(N) (*(Eterm *) (((unsigned char *)reg) + (N)))
-#define yb(N) (*(Eterm *) (((unsigned char *)E) + (N)))
+#define xb(N) (*ADD_BYTE_OFFSET(reg, N))
+#define yb(N) (*ADD_BYTE_OFFSET(E, N))
 #define Sb(N) (*REG_TARGET_PTR(N))
 #define lb(N) (*(double *) (((unsigned char *)&(freg[0].fd)) + (N)))
 #define Qb(N) (N)
 #define Ib(N) (N)
+
 #define x(N) reg[N]
 #define y(N) E[N]
 #define r(N) x(N)
+#define Q(N) (N*sizeof(Eterm *))
+#define l(N) (freg[N].fd)
 
 /*
  * Check that we haven't used the reductions and jump to function pointed to by
diff --git a/erts/emulator/beam/beam_load.c b/erts/emulator/beam/beam_load.c
index 9835b1c096..00dd28b26c 100644
--- a/erts/emulator/beam/beam_load.c
+++ b/erts/emulator/beam/beam_load.c
@@ -2578,23 +2578,31 @@ load_code(LoaderState* stp)
                         sp++;
                     }
                     break;
-		case 'i':	/* Initialize packing accumulator. */
-		    packed = code[--ci];
+#ifdef ARCH_64
+		case '1':	/* Tightest shift (always 10 bits) */
+                    ci--;
+                    ASSERT((code[ci] & ~0x1FF8ull) == 0); /* Fits in 10 bits */
+                    packed = (packed << BEAM_TIGHTEST_SHIFT);
+                    packed |= code[ci] >> 3;
+                    if (packed_label) {
+                        packed_label->packed++;
+                    }
 		    break;
-		case '0':	/* Tight shift */
+#endif
+		case '2':	/* Tight shift (10 or 16 bits) */
 		    packed = (packed << BEAM_TIGHT_SHIFT) | code[--ci];
                     if (packed_label) {
                         packed_label->packed++;
                     }
 		    break;
-		case '6':	/* Shift 16 steps */
+		case '3':	/* Loose shift (16 bits) */
 		    packed = (packed << BEAM_LOOSE_SHIFT) | code[--ci];
                     if (packed_label) {
                         packed_label->packed++;
                     }
 		    break;
 #ifdef ARCH_64
-		case 'w':	/* Shift 32 steps */
+		case '4':	/* Wide shift (32 bits) */
                     {
                         Uint w = code[--ci];
 
@@ -2646,8 +2654,31 @@ load_code(LoaderState* stp)
                     sp++;
 		    packed = 0;
 		    break;
+#if defined(ARCH_64) && defined(CODE_MODEL_SMALL)
+                case '#':       /* -1 */
+                case '$':       /* -2 */
+                case '%':       /* -3 */
+                case '&':       /* -4 */
+                case '\'':      /* -5 */
+                case '(':       /* -6 */
+                    /* Pack accumulator contents into instruction word. */
+                    {
+                        Sint pos = ci - (*prog - '#' + 1);
+                        /* Are the high 32 bits of the instruction word zero? */
+                        ASSERT((code[pos] & ~((1ull << BEAM_WIDE_SHIFT)-1)) == 0);
+                        code[pos] |= packed << BEAM_WIDE_SHIFT;
+                        if (packed_label) {
+                            ASSERT(packed_label->packed == 1);
+                            packed_label->pos = pos;
+                            packed_label->packed = 2;
+                            packed_label = 0;
+                        }
+                        packed >>= BEAM_WIDE_SHIFT;
+                    }
+		    break;
+#endif
 		default:
-		    ASSERT(0);
+                    erts_exit(ERTS_ERROR_EXIT, "beam_load: invalid packing op: %c\n", *prog);
 		}
 	    }
 	    ASSERT(sp == stack); /* Incorrect program? */