JIT compiler update.

git-svn-id: svn://vcs.exim.org/pcre2/code/trunk@768 6239d852-aaf2-0410-a92c-79f79f948069

1 files changed, 73 insertions, 51 deletions

diff --git a/src/sljit/sljitNativeMIPS_common.c b/src/sljit/sljitNativeMIPS_common.c
index 339bdbb..ee207fe 100644
--- a/src/sljit/sljitNativeMIPS_common.c
+++ b/src/sljit/sljitNativeMIPS_common.c
@@ -178,6 +178,8 @@ static const sljit_u8 reg_map[SLJIT_NUMBER_OF_REGISTERS + 5] = {
 #define MOVT		(HI(0) | (1 << 16) | LO(1))
 #define MOVZ		(HI(0) | LO(10))
 #define MUL		(HI(28) | LO(2))
+#define PREF		(HI(51))
+#define PREFX		(HI(19) | LO(15))
 #define SEB		(HI(31) | (16 << 6) | LO(32))
 #define SEH		(HI(31) | (24 << 6) | LO(32))
 #endif
@@ -920,10 +922,8 @@ static sljit_s32 emit_op(struct sljit_compiler *compiler, sljit_s32 op, sljit_s3
 	}
 
 	if (SLJIT_UNLIKELY(dst == SLJIT_UNUSED)) {
-		if (op >= SLJIT_MOV && op <= SLJIT_MOVU_S32 && !(src2 & SLJIT_MEM))
-			return SLJIT_SUCCESS;
-		if (HAS_FLAGS(op))
-			flags |= UNUSED_DEST;
+		SLJIT_ASSERT(HAS_FLAGS(op));
+		flags |= UNUSED_DEST;
 	}
 	else if (FAST_IS_REG(dst)) {
 		dst_r = dst;
@@ -1085,6 +1085,29 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op0(struct sljit_compiler *compile
 	return SLJIT_SUCCESS;
 }
 
+#if (defined SLJIT_MIPS_R1 && SLJIT_MIPS_R1)
+static sljit_s32 emit_prefetch(struct sljit_compiler *compiler,
+        sljit_s32 src, sljit_sw srcw)
+{
+	if (!(src & OFFS_REG_MASK)) {
+		if (srcw <= SIMM_MAX && srcw >= SIMM_MIN)
+			return push_inst(compiler, PREF | S(src & REG_MASK) | IMM(srcw), MOVABLE_INS);
+
+		FAIL_IF(load_immediate(compiler, DR(TMP_REG1), srcw));
+		return push_inst(compiler, PREFX | S(src & REG_MASK) | T(TMP_REG1), MOVABLE_INS);
+	}
+
+	srcw &= 0x3;
+
+	if (SLJIT_UNLIKELY(srcw != 0)) {
+		FAIL_IF(push_inst(compiler, SLL_W | T(OFFS_REG(src)) | D(TMP_REG1) | SH_IMM(srcw), DR(TMP_REG1)));
+		return push_inst(compiler, PREFX | S(src & REG_MASK) | T(TMP_REG1), MOVABLE_INS);
+	}
+
+	return push_inst(compiler, PREFX | S(src & REG_MASK) | T(OFFS_REG(src)), MOVABLE_INS);
+}
+#endif
+
 SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op1(struct sljit_compiler *compiler, sljit_s32 op,
 	sljit_s32 dst, sljit_sw dstw,
 	sljit_s32 src, sljit_sw srcw)
@@ -1100,6 +1123,14 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op1(struct sljit_compiler *compile
 	ADJUST_LOCAL_OFFSET(dst, dstw);
 	ADJUST_LOCAL_OFFSET(src, srcw);
 
+	if (dst == SLJIT_UNUSED && !HAS_FLAGS(op)) {
+#if (defined SLJIT_MIPS_R1 && SLJIT_MIPS_R1)
+		if (op <= SLJIT_MOV_P && (src & SLJIT_MEM))
+			return emit_prefetch(compiler, src, srcw);
+#endif
+		return SLJIT_SUCCESS;
+	}
+
 #if (defined SLJIT_CONFIG_MIPS_64 && SLJIT_CONFIG_MIPS_64)
 	if ((op & SLJIT_I32_OP) && GET_OPCODE(op) >= SLJIT_NOT) {
 		flags |= INT_DATA | SIGNED_DATA;
@@ -1203,6 +1234,9 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op2(struct sljit_compiler *compile
 	ADJUST_LOCAL_OFFSET(src1, src1w);
 	ADJUST_LOCAL_OFFSET(src2, src2w);
 
+	if (dst == SLJIT_UNUSED && !HAS_FLAGS(op))
+		return SLJIT_SUCCESS;
+
 #if (defined SLJIT_CONFIG_MIPS_64 && SLJIT_CONFIG_MIPS_64)
 	if (op & SLJIT_I32_OP) {
 		flags |= INT_DATA | SIGNED_DATA;
@@ -1301,9 +1335,6 @@ static SLJIT_INLINE sljit_s32 sljit_emit_fop1_conv_sw_from_f64(struct sljit_comp
 
 	FAIL_IF(push_inst(compiler, (TRUNC_W_S ^ (flags >> 19)) | FMT(op) | FS(src) | FD(TMP_FREG1), MOVABLE_INS));
 
-	if (dst == SLJIT_UNUSED)
-		return SLJIT_SUCCESS;
-
 	if (FAST_IS_REG(dst))
 		return push_inst(compiler, MFC1 | flags | T(dst) | FS(TMP_FREG1), MOVABLE_INS);
 
@@ -1538,10 +1569,6 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_enter(struct sljit_compiler *
 	CHECK(check_sljit_emit_fast_enter(compiler, dst, dstw));
 	ADJUST_LOCAL_OFFSET(dst, dstw);
 
-	/* For UNUSED dst. Uncommon, but possible. */
-	if (dst == SLJIT_UNUSED)
-		return SLJIT_SUCCESS;
-
 	if (FAST_IS_REG(dst))
 		return push_inst(compiler, ADDU_W | SA(RETURN_ADDR_REG) | TA(0) | D(dst), DR(dst));
 
@@ -1903,50 +1930,43 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_ijump(struct sljit_compiler *compi
 
 SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op_flags(struct sljit_compiler *compiler, sljit_s32 op,
 	sljit_s32 dst, sljit_sw dstw,
-	sljit_s32 src, sljit_sw srcw,
 	sljit_s32 type)
 {
-	sljit_s32 sugg_dst_ar, dst_ar;
-	sljit_s32 flags = GET_ALL_FLAGS(op);
+	sljit_s32 src_ar, dst_ar;
+	sljit_s32 saved_op = op;
 #if (defined SLJIT_CONFIG_MIPS_32 && SLJIT_CONFIG_MIPS_32)
-#	define mem_type WORD_DATA
+	sljit_s32 mem_type = WORD_DATA;
 #else
 	sljit_s32 mem_type = (op & SLJIT_I32_OP) ? (INT_DATA | SIGNED_DATA) : WORD_DATA;
 #endif
 
 	CHECK_ERROR();
-	CHECK(check_sljit_emit_op_flags(compiler, op, dst, dstw, src, srcw, type));
+	CHECK(check_sljit_emit_op_flags(compiler, op, dst, dstw, type));
 	ADJUST_LOCAL_OFFSET(dst, dstw);
 
-	if (dst == SLJIT_UNUSED)
-		return SLJIT_SUCCESS;
-
 	op = GET_OPCODE(op);
 #if (defined SLJIT_CONFIG_MIPS_64 && SLJIT_CONFIG_MIPS_64)
-	if (op == SLJIT_MOV_S32 || op == SLJIT_MOV_U32)
+	if (op == SLJIT_MOV_S32)
 		mem_type = INT_DATA | SIGNED_DATA;
 #endif
-	sugg_dst_ar = DR((op < SLJIT_ADD && FAST_IS_REG(dst)) ? dst : TMP_REG2);
+	dst_ar = DR((op < SLJIT_ADD && FAST_IS_REG(dst)) ? dst : TMP_REG2);
 
 	compiler->cache_arg = 0;
 	compiler->cache_argw = 0;
-	if (op >= SLJIT_ADD && (src & SLJIT_MEM)) {
-		ADJUST_LOCAL_OFFSET(src, srcw);
-		FAIL_IF(emit_op_mem2(compiler, mem_type | LOAD_DATA, DR(TMP_REG1), src, srcw, dst, dstw));
-		src = TMP_REG1;
-		srcw = 0;
-	}
+
+	if (op >= SLJIT_ADD && (dst & SLJIT_MEM))
+		FAIL_IF(emit_op_mem2(compiler, mem_type | LOAD_DATA, DR(TMP_REG1), dst, dstw, dst, dstw));
 
 	switch (type & 0xff) {
 	case SLJIT_EQUAL:
 	case SLJIT_NOT_EQUAL:
-		FAIL_IF(push_inst(compiler, SLTIU | SA(EQUAL_FLAG) | TA(sugg_dst_ar) | IMM(1), sugg_dst_ar));
-		dst_ar = sugg_dst_ar;
+		FAIL_IF(push_inst(compiler, SLTIU | SA(EQUAL_FLAG) | TA(dst_ar) | IMM(1), dst_ar));
+		src_ar = dst_ar;
 		break;
 	case SLJIT_MUL_OVERFLOW:
 	case SLJIT_MUL_NOT_OVERFLOW:
-		FAIL_IF(push_inst(compiler, SLTIU | SA(OTHER_FLAG) | TA(sugg_dst_ar) | IMM(1), sugg_dst_ar));
-		dst_ar = sugg_dst_ar;
+		FAIL_IF(push_inst(compiler, SLTIU | SA(OTHER_FLAG) | TA(dst_ar) | IMM(1), dst_ar));
+		src_ar = dst_ar;
 		type ^= 0x1; /* Flip type bit for the XORI below. */
 		break;
 	case SLJIT_GREATER_F64:
@@ -1958,38 +1978,40 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op_flags(struct sljit_compiler *co
 	case SLJIT_GREATER_EQUAL_F64:
 	case SLJIT_UNORDERED_F64:
 	case SLJIT_ORDERED_F64:
-		FAIL_IF(push_inst(compiler, CFC1 | TA(sugg_dst_ar) | DA(FCSR_REG), sugg_dst_ar));
-		FAIL_IF(push_inst(compiler, SRL | TA(sugg_dst_ar) | DA(sugg_dst_ar) | SH_IMM(23), sugg_dst_ar));
-		FAIL_IF(push_inst(compiler, ANDI | SA(sugg_dst_ar) | TA(sugg_dst_ar) | IMM(1), sugg_dst_ar));
-		dst_ar = sugg_dst_ar;
+		FAIL_IF(push_inst(compiler, CFC1 | TA(dst_ar) | DA(FCSR_REG), dst_ar));
+		FAIL_IF(push_inst(compiler, SRL | TA(dst_ar) | DA(dst_ar) | SH_IMM(23), dst_ar));
+		FAIL_IF(push_inst(compiler, ANDI | SA(dst_ar) | TA(dst_ar) | IMM(1), dst_ar));
+		src_ar = dst_ar;
 		break;
 
 	default:
-		dst_ar = OTHER_FLAG;
+		src_ar = OTHER_FLAG;
 		break;
 	}
 
 	if (type & 0x1) {
-		FAIL_IF(push_inst(compiler, XORI | SA(dst_ar) | TA(sugg_dst_ar) | IMM(1), sugg_dst_ar));
-		dst_ar = sugg_dst_ar;
+		FAIL_IF(push_inst(compiler, XORI | SA(src_ar) | TA(dst_ar) | IMM(1), dst_ar));
+		src_ar = dst_ar;
 	}
 
-	if (op >= SLJIT_ADD) {
-		if (DR(TMP_REG2) != dst_ar)
-			FAIL_IF(push_inst(compiler, ADDU_W | SA(dst_ar) | TA(0) | D(TMP_REG2), DR(TMP_REG2)));
-		return emit_op(compiler, op | flags, mem_type | CUMULATIVE_OP | LOGICAL_OP | IMM_OP | ALT_KEEP_CACHE, dst, dstw, src, srcw, TMP_REG2, 0);
+	if (op < SLJIT_ADD) {
+		if (dst & SLJIT_MEM)
+			return emit_op_mem(compiler, mem_type, src_ar, dst, dstw);
+
+		if (src_ar != dst_ar)
+			return push_inst(compiler, ADDU_W | SA(src_ar) | TA(0) | DA(dst_ar), dst_ar);
+		return SLJIT_SUCCESS;
 	}
 
-	if (dst & SLJIT_MEM)
-		return emit_op_mem(compiler, mem_type, dst_ar, dst, dstw);
+	/* OTHER_FLAG cannot be specified as src2 argument at the moment. */
+	if (DR(TMP_REG2) != src_ar)
+		FAIL_IF(push_inst(compiler, ADDU_W | SA(src_ar) | TA(0) | D(TMP_REG2), DR(TMP_REG2)));
 
-	if (sugg_dst_ar != dst_ar)
-		return push_inst(compiler, ADDU_W | SA(dst_ar) | TA(0) | DA(sugg_dst_ar), sugg_dst_ar);
-	return SLJIT_SUCCESS;
+	mem_type |= CUMULATIVE_OP | LOGICAL_OP | IMM_OP | ALT_KEEP_CACHE;
 
-#if (defined SLJIT_CONFIG_MIPS_32 && SLJIT_CONFIG_MIPS_32)
-#	undef mem_type
-#endif
+	if (dst & SLJIT_MEM)
+		return emit_op(compiler, saved_op, mem_type, dst, dstw, TMP_REG1, 0, TMP_REG2, 0);
+	return emit_op(compiler, saved_op, mem_type, dst, dstw, dst, dstw, TMP_REG2, 0);
 }
 
 SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_cmov(struct sljit_compiler *compiler, sljit_s32 type,
@@ -2078,7 +2100,7 @@ SLJIT_API_FUNC_ATTRIBUTE struct sljit_const* sljit_emit_const(struct sljit_compi
 	PTR_FAIL_IF(!const_);
 	set_const(const_, compiler);
 
-	reg = SLOW_IS_REG(dst) ? dst : TMP_REG2;
+	reg = FAST_IS_REG(dst) ? dst : TMP_REG2;
 
 	PTR_FAIL_IF(emit_const(compiler, reg, init_value));