JIT compiler update: Make fast_call a separate call type. Allows call optimizations on MIPS.

git-svn-id: svn://vcs.exim.org/pcre/code/trunk@722 2f5784b3-3f2a-0410-8824-cb99058d5e15

9 files changed, 38 insertions, 42 deletions

diff --git a/sljit/sljitLir.c b/sljit/sljitLir.c
index 844029e..dcc82c4 100644
--- a/sljit/sljitLir.c
+++ b/sljit/sljitLir.c
@@ -662,7 +662,7 @@ static char* jump_names[] = {
 	(char*)"c_float_less", (char*)"c_float_greater_equal",
 	(char*)"c_float_greater", (char*)"c_float_less_equal",
 	(char*)"c_float_nan", (char*)"c_float_not_nan",
-	(char*)"jump",
+	(char*)"jump", (char*)"fast_call",
 	(char*)"call0", (char*)"call1", (char*)"call2", (char*)"call3"
 };
 
diff --git a/sljit/sljitLir.h b/sljit/sljitLir.h
index ddf775f..9502f14 100644
--- a/sljit/sljitLir.h
+++ b/sljit/sljitLir.h
@@ -573,13 +573,13 @@ struct sljit_label* sljit_emit_label(struct sljit_compiler *compiler);
 #define SLJIT_C_FLOAT_NOT_NAN		21
 
 #define SLJIT_JUMP			22
-#define SLJIT_CALL0			23
-#define SLJIT_CALL1			24
-#define SLJIT_CALL2			25
-#define SLJIT_CALL3			26
+#define SLJIT_FAST_CALL			23
+#define SLJIT_CALL0			24
+#define SLJIT_CALL1			25
+#define SLJIT_CALL2			26
+#define SLJIT_CALL3			27
 
 /* Fast calling method. See sljit_emit_fast_enter / sljit_emit_fast_return. */
-#define SLJIT_FAST_CALL			SLJIT_CALL0
 
 /* The target can be changed during runtime (see: sljit_set_jump_addr). */
 #define SLJIT_REWRITABLE_JUMP		0x1000
diff --git a/sljit/sljitNativeARM_Thumb2.c b/sljit/sljitNativeARM_Thumb2.c
index de68d7b..9a7d6b9 100644
--- a/sljit/sljitNativeARM_Thumb2.c
+++ b/sljit/sljitNativeARM_Thumb2.c
@@ -1736,26 +1736,19 @@ int sljit_emit_ijump(struct sljit_compiler *compiler, int type, int src, sljit_w
 	if (src & SLJIT_IMM) {
 		jump = (struct sljit_jump*)ensure_abuf(compiler, sizeof(struct sljit_jump));
 		FAIL_IF(!jump);
-		set_jump(jump, compiler, JUMP_ADDR | ((type >= SLJIT_CALL0) ? IS_BL : 0));
+		set_jump(jump, compiler, JUMP_ADDR | ((type >= SLJIT_FAST_CALL) ? IS_BL : 0));
 		jump->u.target = srcw;
 
 		FAIL_IF(emit_imm32_const(compiler, TMP_REG1, 0));
 		jump->addr = compiler->size;
-		if (type <= SLJIT_JUMP)
-			FAIL_IF(push_inst16(compiler, BX | RN3(TMP_REG1)));
-		else
-			FAIL_IF(push_inst16(compiler, BLX | RN3(TMP_REG1)));
+		FAIL_IF(push_inst16(compiler, (type <= SLJIT_JUMP ? BX : BLX) | RN3(TMP_REG1)));
 	}
 	else {
-		if (src >= SLJIT_TEMPORARY_REG1 && src <= SLJIT_NO_REGISTERS) {
-			if (type <= SLJIT_JUMP)
-				return push_inst16(compiler, BX | RN3(src));
-			else
-				return push_inst16(compiler, BLX | RN3(src));
-		}
+		if (src >= SLJIT_TEMPORARY_REG1 && src <= SLJIT_NO_REGISTERS)
+			return push_inst16(compiler, (type <= SLJIT_JUMP ? BX : BLX) | RN3(src));
 
 		FAIL_IF(emit_op_mem(compiler, WORD_SIZE, type <= SLJIT_JUMP ? TMP_PC : TMP_REG1, src, srcw));
-		if (type >= SLJIT_CALL0)
+		if (type >= SLJIT_FAST_CALL)
 			return push_inst16(compiler, BLX | RN3(TMP_REG1));
 	}
 	return SLJIT_SUCCESS;
diff --git a/sljit/sljitNativeARM_v5.c b/sljit/sljitNativeARM_v5.c
index b1551df..075fefe 100644
--- a/sljit/sljitNativeARM_v5.c
+++ b/sljit/sljitNativeARM_v5.c
@@ -400,7 +400,7 @@ static SLJIT_INLINE int detect_jump_type(struct sljit_jump *jump, sljit_uw *code
 		diff = ((sljit_w)(code + jump->u.label->size) - (sljit_w)(code_ptr + 2));
 	}
 
-	/* Branch to Thumb code has not optimized yet. */
+	/* Branch to Thumb code has not been optimized yet. */
 	if (diff & 0x3)
 		return 0;
 
@@ -426,7 +426,7 @@ static SLJIT_INLINE int detect_jump_type(struct sljit_jump *jump, sljit_uw *code
 		diff = ((sljit_w)(code + jump->u.label->size) - (sljit_w)code_ptr);
 	}
 
-	/* Branch to Thumb code has not optimized yet. */
+	/* Branch to Thumb code has not been optimized yet. */
 	if (diff & 0x3)
 		return 0;
 
@@ -2220,7 +2220,7 @@ struct sljit_jump* sljit_emit_jump(struct sljit_compiler *compiler, int type)
 
 	/* In ARM, we don't need to touch the arguments. */
 #if (defined SLJIT_CONFIG_ARM_V5 && SLJIT_CONFIG_ARM_V5)
-	if (type >= SLJIT_CALL0)
+	if (type >= SLJIT_FAST_CALL)
 		PTR_FAIL_IF(prepare_blx(compiler));
 	PTR_FAIL_IF(push_inst_with_unique_literal(compiler, ((EMIT_DATA_TRANSFER(WORD_DATA | LOAD_DATA, 1, 0,
 		type <= SLJIT_JUMP ? TMP_PC : TMP_REG1, TMP_PC, 0)) & ~COND_MASK) | get_cc(type), 0));
@@ -2230,7 +2230,7 @@ struct sljit_jump* sljit_emit_jump(struct sljit_compiler *compiler, int type)
 		compiler->patches++;
 	}
 
-	if (type >= SLJIT_CALL0) {
+	if (type >= SLJIT_FAST_CALL) {
 		jump->flags |= IS_BL;
 		PTR_FAIL_IF(emit_blx(compiler));
 	}
@@ -2238,10 +2238,10 @@ struct sljit_jump* sljit_emit_jump(struct sljit_compiler *compiler, int type)
 	if (!(jump->flags & SLJIT_REWRITABLE_JUMP))
 		jump->addr = compiler->size;
 #else
-	if (type >= SLJIT_CALL0)
+	if (type >= SLJIT_FAST_CALL)
 		jump->flags |= IS_BL;
 	PTR_FAIL_IF(emit_imm(compiler, TMP_REG1, 0));
-	PTR_FAIL_IF(push_inst(compiler, (((type < SLJIT_CALL0 ? BX : BLX) | RM(TMP_REG1)) & ~COND_MASK) | get_cc(type)));
+	PTR_FAIL_IF(push_inst(compiler, (((type <= SLJIT_JUMP ? BX : BLX) | RM(TMP_REG1)) & ~COND_MASK) | get_cc(type)));
 	jump->addr = compiler->size;
 #endif
 	return jump;
@@ -2258,18 +2258,18 @@ int sljit_emit_ijump(struct sljit_compiler *compiler, int type, int src, sljit_w
 	if (src & SLJIT_IMM) {
 		jump = (struct sljit_jump*)ensure_abuf(compiler, sizeof(struct sljit_jump));
 		FAIL_IF(!jump);
-		set_jump(jump, compiler, JUMP_ADDR | ((type >= SLJIT_CALL0) ? IS_BL : 0));
+		set_jump(jump, compiler, JUMP_ADDR | ((type >= SLJIT_FAST_CALL) ? IS_BL : 0));
 		jump->u.target = srcw;
 
 #if (defined SLJIT_CONFIG_ARM_V5 && SLJIT_CONFIG_ARM_V5)
-		if (type >= SLJIT_CALL0)
+		if (type >= SLJIT_FAST_CALL)
 			FAIL_IF(prepare_blx(compiler));
 		FAIL_IF(push_inst_with_unique_literal(compiler, EMIT_DATA_TRANSFER(WORD_DATA | LOAD_DATA, 1, 0, type <= SLJIT_JUMP ? TMP_PC : TMP_REG1, TMP_PC, 0), 0));
-		if (type >= SLJIT_CALL0)
+		if (type >= SLJIT_FAST_CALL)
 			FAIL_IF(emit_blx(compiler));
 #else
 		FAIL_IF(emit_imm(compiler, TMP_REG1, 0));
-		FAIL_IF(push_inst(compiler, (type < SLJIT_CALL0 ? BX : BLX) | RM(TMP_REG1)));
+		FAIL_IF(push_inst(compiler, (type <= SLJIT_JUMP ? BX : BLX) | RM(TMP_REG1)));
 #endif
 		jump->addr = compiler->size;
 	}
diff --git a/sljit/sljitNativeMIPS_common.c b/sljit/sljitNativeMIPS_common.c
index 4a34222..7c7df5f 100644
--- a/sljit/sljitNativeMIPS_common.c
+++ b/sljit/sljitNativeMIPS_common.c
@@ -97,6 +97,7 @@ typedef sljit_ui sljit_ins;
 #define AND		(HI(0) | LO(36))
 #define ANDI		(HI(12))
 #define B		(HI(4))
+#define BAL		(HI(1) | (17 << 16))
 #define BC1F		(HI(17) | (8 << 21))
 #define BC1T		(HI(17) | (8 << 21) | (1 << 16))
 #define BEQ		(HI(4))
@@ -111,6 +112,7 @@ typedef sljit_ui sljit_ins;
 #define C_ULT_D		(HI(17) | FMT_D | LO(53))
 #define DIV_D		(HI(17) | FMT_D | LO(3))
 #define J		(HI(2))
+#define JAL		(HI(3))
 #define JALR		(HI(0) | LO(9))
 #define JR		(HI(0) | LO(8))
 #define LD		(HI(55))
@@ -199,7 +201,7 @@ static SLJIT_INLINE sljit_ins* optimize_jump(struct sljit_jump *jump, sljit_ins
 	sljit_ins *inst;
 	sljit_ins saved_inst;
 
-	if (jump->flags & (SLJIT_REWRITABLE_JUMP | IS_JAL))
+	if (jump->flags & SLJIT_REWRITABLE_JUMP)
 		return code_ptr;
 
 	if (jump->flags & JUMP_ADDR)
@@ -220,7 +222,7 @@ static SLJIT_INLINE sljit_ins* optimize_jump(struct sljit_jump *jump, sljit_ins
 
 			if (!(jump->flags & IS_COND)) {
 				inst[0] = inst[-1];
-				inst[-1] = B;
+				inst[-1] = (jump->flags & IS_JAL) ? BAL : B;
 				jump->addr -= sizeof(sljit_ins);
 				return inst;
 			}
@@ -237,7 +239,7 @@ static SLJIT_INLINE sljit_ins* optimize_jump(struct sljit_jump *jump, sljit_ins
 		jump->flags |= PATCH_B;
 
 		if (!(jump->flags & IS_COND)) {
-			inst[0] = B;
+			inst[0] = (jump->flags & IS_JAL) ? BAL : B;
 			inst[1] = NOP;
 			return inst + 1;
 		}
@@ -265,7 +267,7 @@ static SLJIT_INLINE sljit_ins* optimize_jump(struct sljit_jump *jump, sljit_ins
 		if ((target_addr & ~0xfffffff) == (jump->addr & ~0xfffffff)) {
 			jump->flags |= PATCH_J;
 			inst[0] = inst[-1];
-			inst[-1] = J;
+			inst[-1] = (jump->flags & IS_JAL) ? JAL : J;
 			jump->addr -= sizeof(sljit_ins);
 			return inst;
 		}
@@ -273,7 +275,7 @@ static SLJIT_INLINE sljit_ins* optimize_jump(struct sljit_jump *jump, sljit_ins
 
 	if ((target_addr & ~0xfffffff) == ((jump->addr + sizeof(sljit_ins)) & ~0xfffffff)) {
 		jump->flags |= PATCH_J;
-		inst[0] = J;
+		inst[0] = (jump->flags & IS_JAL) ? JAL : J;
 		inst[1] = NOP;
 		return inst + 1;
 	}
@@ -1396,11 +1398,12 @@ struct sljit_jump* sljit_emit_jump(struct sljit_compiler *compiler, int type)
 		jump->addr = compiler->size;
 		PTR_FAIL_IF(push_inst(compiler, NOP, UNMOVABLE_INS));
 	} else {
-		/* Cannot be optimized out. */
 		SLJIT_ASSERT(DR(PIC_ADDR_REG) == 25 && PIC_ADDR_REG == TMP_REG2);
-		jump->flags |= IS_JAL;
+		/* Cannot be optimized out if type is >= CALL0. */
+		jump->flags |= IS_JAL | (type >= SLJIT_CALL0 ? SLJIT_REWRITABLE_JUMP : 0);
 		PTR_FAIL_IF(push_inst(compiler, JALR | S(TMP_REG2) | DA(RETURN_ADDR_REG), UNMOVABLE_INS));
 		jump->addr = compiler->size;
+		/* A NOP if type < CALL1. */
 		PTR_FAIL_IF(push_inst(compiler, ADDU_W | S(SLJIT_TEMPORARY_REG1) | TA(0) | DA(4), UNMOVABLE_INS));
 	}
 	return jump;
@@ -1596,7 +1599,7 @@ int sljit_emit_ijump(struct sljit_compiler *compiler, int type, int src, sljit_w
 	if (src & SLJIT_IMM) {
 		jump = (struct sljit_jump*)ensure_abuf(compiler, sizeof(struct sljit_jump));
 		FAIL_IF(!jump);
-		set_jump(jump, compiler, JUMP_ADDR);
+		set_jump(jump, compiler, JUMP_ADDR | ((type >= SLJIT_FAST_CALL) ? IS_JAL : 0));
 		jump->u.target = srcw;
 
 		if (compiler->delay_slot != UNMOVABLE_INS)
diff --git a/sljit/sljitNativePPC_common.c b/sljit/sljitNativePPC_common.c
index 0ad1a35..d356af2 100644
--- a/sljit/sljitNativePPC_common.c
+++ b/sljit/sljitNativePPC_common.c
@@ -1612,7 +1612,7 @@ struct sljit_jump* sljit_emit_jump(struct sljit_compiler *compiler, int type)
 	PTR_FAIL_IF(emit_const(compiler, TMP_REG1, 0));
 	PTR_FAIL_IF(push_inst(compiler, MTCTR | S(TMP_REG1)));
 	jump->addr = compiler->size;
-	PTR_FAIL_IF(push_inst(compiler, BCCTR | bo_bi_flags | (type >= SLJIT_CALL0 ? 1 : 0)));
+	PTR_FAIL_IF(push_inst(compiler, BCCTR | bo_bi_flags | (type >= SLJIT_FAST_CALL ? 1 : 0)));
 	return jump;
 }
 
@@ -1647,7 +1647,7 @@ int sljit_emit_ijump(struct sljit_compiler *compiler, int type, int src, sljit_w
 	FAIL_IF(push_inst(compiler, MTCTR | S(src_r)));
 	if (jump)
 		jump->addr = compiler->size;
-	return push_inst(compiler, BCCTR | bo_bi_flags | (type >= SLJIT_CALL0 ? 1 : 0));
+	return push_inst(compiler, BCCTR | bo_bi_flags | (type >= SLJIT_FAST_CALL ? 1 : 0));
 }
 
 /* Get a bit from CR, all other bits are zeroed. */
diff --git a/sljit/sljitNativeX86_32.c b/sljit/sljitNativeX86_32.c
index a1123cb..97e6ade 100644
--- a/sljit/sljitNativeX86_32.c
+++ b/sljit/sljitNativeX86_32.c
@@ -44,7 +44,7 @@ static sljit_ub* generate_far_jump_code(struct sljit_jump *jump, sljit_ub *code_
 		*code_ptr++ = 0xe9;
 		jump->addr++;
 	}
-	else if (type >= SLJIT_CALL0) {
+	else if (type >= SLJIT_FAST_CALL) {
 		*code_ptr++ = 0xe8;
 		jump->addr++;
 	}
diff --git a/sljit/sljitNativeX86_64.c b/sljit/sljitNativeX86_64.c
index c41d15c..0ee34f0 100644
--- a/sljit/sljitNativeX86_64.c
+++ b/sljit/sljitNativeX86_64.c
@@ -59,7 +59,7 @@ static sljit_ub* generate_far_jump_code(struct sljit_jump *jump, sljit_ub *code_
 	code_ptr += sizeof(sljit_w);
 	*code_ptr++ = REX_B;
 	*code_ptr++ = 0xff;
-	*code_ptr++ = (type >= SLJIT_CALL0) ? 0xd1 /* call */ : 0xe1 /* jmp */;
+	*code_ptr++ = (type >= SLJIT_FAST_CALL) ? 0xd1 /* call */ : 0xe1 /* jmp */;
 
 	return code_ptr;
 }
diff --git a/sljit/sljitNativeX86_common.c b/sljit/sljitNativeX86_common.c
index dd993e0..f86546e 100644
--- a/sljit/sljitNativeX86_common.c
+++ b/sljit/sljitNativeX86_common.c
@@ -245,7 +245,7 @@ static sljit_ub* generate_near_jump_code(struct sljit_jump *jump, sljit_ub *code
 			*code_ptr++ = 0xe9;
 		jump->addr++;
 	}
-	else if (type >= SLJIT_CALL0) {
+	else if (type >= SLJIT_FAST_CALL) {
 		short_jump = 0;
 		*code_ptr++ = 0xe8;
 		jump->addr++;
@@ -2436,7 +2436,7 @@ int sljit_emit_ijump(struct sljit_compiler *compiler, int type, int src, sljit_w
 		code = emit_x86_instruction(compiler, 1, 0, 0, src, srcw);
 		FAIL_IF(!code);
 		*code++ = 0xff;
-		*code |= (type >= SLJIT_CALL0) ? (2 << 3) : (4 << 3);
+		*code |= (type >= SLJIT_FAST_CALL) ? (2 << 3) : (4 << 3);
 	}
 	return SLJIT_SUCCESS;
 }