Fix unaligned accesses on x86. Patch by Marc Mutz.

git-svn-id: svn://vcs.exim.org/pcre/code/trunk@1662 2f5784b3-3f2a-0410-8824-cb99058d5e15

5 files changed, 71 insertions, 37 deletions

diff --git a/ChangeLog b/ChangeLog
index 732d199..c8f3642 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -17,6 +17,8 @@ Version 8.40 17-June-2016
 
 4.  Ignore "show all captures" (/=) for DFA matching.
 
+5.  Fix unaligned accesses on x86. Patch by Marc Mutz.
+
 
 Version 8.39 14-June-2016
 -------------------------
diff --git a/sljit/sljitConfigInternal.h b/sljit/sljitConfigInternal.h
index 9275b14..1456bd5 100644
--- a/sljit/sljitConfigInternal.h
+++ b/sljit/sljitConfigInternal.h
@@ -214,6 +214,10 @@
 #define SLJIT_MEMMOVE(dest, src, len) memmove(dest, src, len)
 #endif
 
+#ifndef SLJIT_MEMCPY
+#define SLJIT_MEMCPY(dest, src, len) memcpy(dest, src, len)
+#endif
+
 #ifndef SLJIT_ZEROMEM
 #define SLJIT_ZEROMEM(dest, len) memset(dest, 0, len)
 #endif
diff --git a/sljit/sljitNativeX86_32.c b/sljit/sljitNativeX86_32.c
index cd3c656..78f3dcb 100644
--- a/sljit/sljitNativeX86_32.c
+++ b/sljit/sljitNativeX86_32.c
@@ -34,7 +34,7 @@ static sljit_s32 emit_do_imm(struct sljit_compiler *compiler, sljit_u8 opcode, s
 	FAIL_IF(!inst);
 	INC_SIZE(1 + sizeof(sljit_sw));
 	*inst++ = opcode;
-	*(sljit_sw*)inst = imm;
+	sljit_unaligned_store_sw(inst, imm);
 	return SLJIT_SUCCESS;
 }
 
@@ -57,7 +57,7 @@ static sljit_u8* generate_far_jump_code(struct sljit_jump *jump, sljit_u8 *code_
 	if (jump->flags & JUMP_LABEL)
 		jump->flags |= PATCH_MW;
 	else
-		*(sljit_sw*)code_ptr = jump->u.target - (jump->addr + 4);
+		sljit_unaligned_store_sw(code_ptr, jump->u.target - (jump->addr + 4));
 	code_ptr += 4;
 
 	return code_ptr;
@@ -151,12 +151,12 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_enter(struct sljit_compiler *compi
 		inst[1] = MOD_REG | (reg_map[TMP_REG1] << 3) | reg_map[SLJIT_SP];
 		inst[2] = GROUP_F7;
 		inst[3] = MOD_REG | (0 << 3) | reg_map[SLJIT_SP];
-		*(sljit_sw*)(inst + 4) = 0x4;
+		sljit_unaligned_store_sw(inst + 4, 0x4);
 		inst[8] = JNE_i8;
 		inst[9] = 6;
 		inst[10] = GROUP_BINARY_81;
 		inst[11] = MOD_REG | (5 << 3) | reg_map[SLJIT_SP];
-		*(sljit_sw*)(inst + 12) = 0x4;
+		sljit_unaligned_store_sw(inst + 12, 0x4);
 		inst[16] = PUSH_r + reg_map[TMP_REG1];
 	}
 	else
@@ -406,7 +406,7 @@ static sljit_u8* emit_x86_instruction(struct sljit_compiler *compiler, sljit_s32
 				if (immb <= 127 && immb >= -128)
 					*buf_ptr++ = immb; /* 8 bit displacement. */
 				else {
-					*(sljit_sw*)buf_ptr = immb; /* 32 bit displacement. */
+					sljit_unaligned_store_sw(buf_ptr, immb); /* 32 bit displacement. */
 					buf_ptr += sizeof(sljit_sw);
 				}
 			}
@@ -418,7 +418,7 @@ static sljit_u8* emit_x86_instruction(struct sljit_compiler *compiler, sljit_s32
 	}
 	else {
 		*buf_ptr++ |= 0x05;
-		*(sljit_sw*)buf_ptr = immb; /* 32 bit displacement. */
+		sljit_unaligned_store_sw(buf_ptr, immb); /* 32 bit displacement. */
 		buf_ptr += sizeof(sljit_sw);
 	}
 
@@ -426,9 +426,9 @@ static sljit_u8* emit_x86_instruction(struct sljit_compiler *compiler, sljit_s32
 		if (flags & EX86_BYTE_ARG)
 			*buf_ptr = imma;
 		else if (flags & EX86_HALF_ARG)
-			*(short*)buf_ptr = imma;
+			sljit_unaligned_store_s16(buf_ptr, imma);
 		else if (!(flags & EX86_SHIFT_INS))
-			*(sljit_sw*)buf_ptr = imma;
+			sljit_unaligned_store_sw(buf_ptr, imma);
 	}
 
 	return !(flags & EX86_SHIFT_INS) ? inst : (inst + 1);
@@ -541,7 +541,7 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_return(struct sljit_compiler
 
 		INC_SIZE(5 + 1);
 		*inst++ = PUSH_i32;
-		*(sljit_sw*)inst = srcw;
+		sljit_unaligned_store_sw(inst, srcw);
 		inst += sizeof(sljit_sw);
 	}
 
diff --git a/sljit/sljitNativeX86_64.c b/sljit/sljitNativeX86_64.c
index 6ea27e5..e88dded 100644
--- a/sljit/sljitNativeX86_64.c
+++ b/sljit/sljitNativeX86_64.c
@@ -35,7 +35,7 @@ static sljit_s32 emit_load_imm64(struct sljit_compiler *compiler, sljit_s32 reg,
 	INC_SIZE(2 + sizeof(sljit_sw));
 	*inst++ = REX_W | ((reg_map[reg] <= 7) ? 0 : REX_B);
 	*inst++ = MOV_r_i32 + (reg_map[reg] & 0x7);
-	*(sljit_sw*)inst = imm;
+	sljit_unaligned_store_sw(inst, imm);
 	return SLJIT_SUCCESS;
 }
 
@@ -55,7 +55,7 @@ static sljit_u8* generate_far_jump_code(struct sljit_jump *jump, sljit_u8 *code_
 	if (jump->flags & JUMP_LABEL)
 		jump->flags |= PATCH_MD;
 	else
-		*(sljit_sw*)code_ptr = jump->u.target;
+		sljit_unaligned_store_sw(code_ptr, jump->u.target);
 
 	code_ptr += sizeof(sljit_sw);
 	*code_ptr++ = REX_B;
@@ -71,13 +71,13 @@ static sljit_u8* generate_fixed_jump(sljit_u8 *code_ptr, sljit_sw addr, sljit_s3
 
 	if (delta <= HALFWORD_MAX && delta >= HALFWORD_MIN) {
 		*code_ptr++ = (type == 2) ? CALL_i32 : JMP_i32;
-		*(sljit_sw*)code_ptr = delta;
+		sljit_unaligned_store_sw(code_ptr, delta);
 	}
 	else {
 		SLJIT_COMPILE_ASSERT(reg_map[TMP_REG3] == 9, tmp3_is_9_second);
 		*code_ptr++ = REX_W | REX_B;
 		*code_ptr++ = MOV_r_i32 + 1;
-		*(sljit_sw*)code_ptr = addr;
+		sljit_unaligned_store_sw(code_ptr, addr);
 		code_ptr += sizeof(sljit_sw);
 		*code_ptr++ = REX_B;
 		*code_ptr++ = GROUP_FF;
@@ -193,7 +193,7 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_enter(struct sljit_compiler *compi
 		*inst++ = REX_W;
 		*inst++ = MOV_rm_i32;
 		*inst++ = MOD_REG | reg_lmap[SLJIT_R0];
-		*(sljit_s32*)inst = local_size;
+		sljit_unaligned_store_s32(inst, local_size);
 #if (defined SLJIT_VERBOSE && SLJIT_VERBOSE) \
 			|| (defined SLJIT_ARGUMENT_CHECKS && SLJIT_ARGUMENT_CHECKS)
 		compiler->skip_checks = 1;
@@ -219,7 +219,7 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_enter(struct sljit_compiler *compi
 		*inst++ = REX_W;
 		*inst++ = GROUP_BINARY_81;
 		*inst++ = MOD_REG | SUB | 4;
-		*(sljit_s32*)inst = local_size;
+		sljit_unaligned_store_s32(inst, local_size);
 		inst += sizeof(sljit_s32);
 	}
 
@@ -230,7 +230,7 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_enter(struct sljit_compiler *compi
 		FAIL_IF(!inst);
 		INC_SIZE(5);
 		*inst++ = GROUP_0F;
-		*(sljit_s32*)inst = 0x20247429;
+		sljit_unaligned_store_s32(inst, 0x20247429);
 	}
 #endif
 
@@ -271,7 +271,7 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_return(struct sljit_compiler *comp
 		FAIL_IF(!inst);
 		INC_SIZE(5);
 		*inst++ = GROUP_0F;
-		*(sljit_s32*)inst = 0x20247428;
+		sljit_unaligned_store_s32(inst, 0x20247428);
 	}
 #endif
 
@@ -292,7 +292,7 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_return(struct sljit_compiler *comp
 		*inst++ = REX_W;
 		*inst++ = GROUP_BINARY_81;
 		*inst++ = MOD_REG | ADD | 4;
-		*(sljit_s32*)inst = compiler->local_size;
+		sljit_unaligned_store_s32(inst, compiler->local_size);
 	}
 
 	tmp = compiler->scratches;
@@ -339,7 +339,7 @@ static sljit_s32 emit_do_imm32(struct sljit_compiler *compiler, sljit_u8 rex, sl
 	if (rex)
 		*inst++ = rex;
 	*inst++ = opcode;
-	*(sljit_s32*)inst = imm;
+	sljit_unaligned_store_s32(inst, imm);
 	return SLJIT_SUCCESS;
 }
 
@@ -516,7 +516,7 @@ static sljit_u8* emit_x86_instruction(struct sljit_compiler *compiler, sljit_s32
 				if (immb <= 127 && immb >= -128)
 					*buf_ptr++ = immb; /* 8 bit displacement. */
 				else {
-					*(sljit_s32*)buf_ptr = immb; /* 32 bit displacement. */
+					sljit_unaligned_store_s32(buf_ptr, immb); /* 32 bit displacement. */
 					buf_ptr += sizeof(sljit_s32);
 				}
 			}
@@ -533,7 +533,7 @@ static sljit_u8* emit_x86_instruction(struct sljit_compiler *compiler, sljit_s32
 	else {
 		*buf_ptr++ |= 0x04;
 		*buf_ptr++ = 0x25;
-		*(sljit_s32*)buf_ptr = immb; /* 32 bit displacement. */
+		sljit_unaligned_store_s32(buf_ptr, immb); /* 32 bit displacement. */
 		buf_ptr += sizeof(sljit_s32);
 	}
 
@@ -541,9 +541,9 @@ static sljit_u8* emit_x86_instruction(struct sljit_compiler *compiler, sljit_s32
 		if (flags & EX86_BYTE_ARG)
 			*buf_ptr = imma;
 		else if (flags & EX86_HALF_ARG)
-			*(short*)buf_ptr = imma;
+			sljit_unaligned_store_s16(buf_ptr, imma);
 		else if (!(flags & EX86_SHIFT_INS))
-			*(sljit_s32*)buf_ptr = imma;
+			sljit_unaligned_store_s32(buf_ptr, imma);
 	}
 
 	return !(flags & EX86_SHIFT_INS) ? inst : (inst + 1);
@@ -676,7 +676,7 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_fast_return(struct sljit_compiler
 
 		INC_SIZE(5 + 1);
 		*inst++ = PUSH_i32;
-		*(sljit_s32*)inst = srcw;
+		sljit_unaligned_store_s32(inst, srcw);
 		inst += sizeof(sljit_s32);
 	}
 
diff --git a/sljit/sljitNativeX86_common.c b/sljit/sljitNativeX86_common.c
index 54c3ac7..0d1de96 100644
--- a/sljit/sljitNativeX86_common.c
+++ b/sljit/sljitNativeX86_common.c
@@ -279,6 +279,34 @@ static sljit_s32 cpu_has_cmov = -1;
 #include <intrin.h>
 #endif
 
+/******************************************************/
+/*    Unaligned-store functions                       */
+/******************************************************/
+
+static SLJIT_INLINE void sljit_unaligned_store_s16(void *addr, sljit_s16 value)
+{
+	SLJIT_MEMCPY(addr, &value, sizeof(value));
+}
+
+static SLJIT_INLINE void sljit_unaligned_store_s32(void *addr, sljit_s32 value)
+{
+	SLJIT_MEMCPY(addr, &value, sizeof(value));
+}
+
+static SLJIT_INLINE void sljit_unaligned_store_sw(void *addr, sljit_sw value)
+{
+	SLJIT_MEMCPY(addr, &value, sizeof(value));
+}
+
+static SLJIT_INLINE void sljit_unaligned_store_uw(void *addr, sljit_uw value)
+{
+	SLJIT_MEMCPY(addr, &value, sizeof(value));
+}
+
+/******************************************************/
+/*    Utility functions                               */
+/******************************************************/
+
 static void get_cpu_features(void)
 {
 	sljit_u32 features;
@@ -504,7 +532,7 @@ SLJIT_API_FUNC_ATTRIBUTE void* sljit_generate_code(struct sljit_compiler *compil
 #if (defined SLJIT_CONFIG_X86_32 && SLJIT_CONFIG_X86_32)
 					*code_ptr++ = (*buf_ptr == 2) ? CALL_i32 : JMP_i32;
 					buf_ptr++;
-					*(sljit_sw*)code_ptr = *(sljit_sw*)buf_ptr - ((sljit_sw)code_ptr + sizeof(sljit_sw));
+					sljit_unaligned_store_sw(code_ptr, *(sljit_sw*)buf_ptr - ((sljit_sw)code_ptr + sizeof(sljit_sw)));
 					code_ptr += sizeof(sljit_sw);
 					buf_ptr += sizeof(sljit_sw) - 1;
 #else
@@ -531,24 +559,24 @@ SLJIT_API_FUNC_ATTRIBUTE void* sljit_generate_code(struct sljit_compiler *compil
 		} else if (jump->flags & PATCH_MW) {
 			if (jump->flags & JUMP_LABEL) {
 #if (defined SLJIT_CONFIG_X86_32 && SLJIT_CONFIG_X86_32)
-				*(sljit_sw*)jump->addr = (sljit_sw)(jump->u.label->addr - (jump->addr + sizeof(sljit_sw)));
+				sljit_unaligned_store_sw((void*)jump->addr, (sljit_sw)(jump->u.label->addr - (jump->addr + sizeof(sljit_sw))));
 #else
 				SLJIT_ASSERT((sljit_sw)(jump->u.label->addr - (jump->addr + sizeof(sljit_s32))) >= HALFWORD_MIN && (sljit_sw)(jump->u.label->addr - (jump->addr + sizeof(sljit_s32))) <= HALFWORD_MAX);
-				*(sljit_s32*)jump->addr = (sljit_s32)(jump->u.label->addr - (jump->addr + sizeof(sljit_s32)));
+				sljit_unaligned_store_s32((void*)jump->addr, (sljit_s32)(jump->u.label->addr - (jump->addr + sizeof(sljit_s32))));
 #endif
 			}
 			else {
 #if (defined SLJIT_CONFIG_X86_32 && SLJIT_CONFIG_X86_32)
-				*(sljit_sw*)jump->addr = (sljit_sw)(jump->u.target - (jump->addr + sizeof(sljit_sw)));
+				sljit_unaligned_store_sw((void*)jump->addr, (sljit_sw)(jump->u.target - (jump->addr + sizeof(sljit_sw))));
 #else
 				SLJIT_ASSERT((sljit_sw)(jump->u.target - (jump->addr + sizeof(sljit_s32))) >= HALFWORD_MIN && (sljit_sw)(jump->u.target - (jump->addr + sizeof(sljit_s32))) <= HALFWORD_MAX);
-				*(sljit_s32*)jump->addr = (sljit_s32)(jump->u.target - (jump->addr + sizeof(sljit_s32)));
+				sljit_unaligned_store_s32((void*)jump->addr, (sljit_s32)(jump->u.target - (jump->addr + sizeof(sljit_s32))));
 #endif
 			}
 		}
 #if (defined SLJIT_CONFIG_X86_64 && SLJIT_CONFIG_X86_64)
 		else if (jump->flags & PATCH_MD)
-			*(sljit_sw*)jump->addr = jump->u.label->addr;
+			sljit_unaligned_store_sw((void*)jump->addr, jump->u.label->addr);
 #endif
 
 		jump = jump->next;
@@ -1699,7 +1727,7 @@ static sljit_s32 emit_mul(struct sljit_compiler *compiler,
 			inst = (sljit_u8*)ensure_buf(compiler, 1 + 4);
 			FAIL_IF(!inst);
 			INC_SIZE(4);
-			*(sljit_sw*)inst = src1w;
+			sljit_unaligned_store_sw(inst, src1w);
 		}
 #else
 		else if (IS_HALFWORD(src1w)) {
@@ -1709,7 +1737,7 @@ static sljit_s32 emit_mul(struct sljit_compiler *compiler,
 			inst = (sljit_u8*)ensure_buf(compiler, 1 + 4);
 			FAIL_IF(!inst);
 			INC_SIZE(4);
-			*(sljit_s32*)inst = (sljit_s32)src1w;
+			sljit_unaligned_store_s32(inst, (sljit_s32)src1w);
 		}
 		else {
 			EMIT_MOV(compiler, TMP_REG2, 0, SLJIT_IMM, src1w);
@@ -1742,7 +1770,7 @@ static sljit_s32 emit_mul(struct sljit_compiler *compiler,
 			inst = (sljit_u8*)ensure_buf(compiler, 1 + 4);
 			FAIL_IF(!inst);
 			INC_SIZE(4);
-			*(sljit_sw*)inst = src2w;
+			sljit_unaligned_store_sw(inst, src2w);
 		}
 #else
 		else if (IS_HALFWORD(src2w)) {
@@ -1752,7 +1780,7 @@ static sljit_s32 emit_mul(struct sljit_compiler *compiler,
 			inst = (sljit_u8*)ensure_buf(compiler, 1 + 4);
 			FAIL_IF(!inst);
 			INC_SIZE(4);
-			*(sljit_s32*)inst = (sljit_s32)src2w;
+			sljit_unaligned_store_s32(inst, (sljit_s32)src2w);
 		}
 		else {
 			EMIT_MOV(compiler, TMP_REG2, 0, SLJIT_IMM, src2w);
@@ -2926,15 +2954,15 @@ SLJIT_API_FUNC_ATTRIBUTE struct sljit_const* sljit_emit_const(struct sljit_compi
 SLJIT_API_FUNC_ATTRIBUTE void sljit_set_jump_addr(sljit_uw addr, sljit_uw new_addr)
 {
 #if (defined SLJIT_CONFIG_X86_32 && SLJIT_CONFIG_X86_32)
-	*(sljit_sw*)addr = new_addr - (addr + 4);
+	sljit_unaligned_store_sw((void*)addr, new_addr - (addr + 4));
 #else
-	*(sljit_uw*)addr = new_addr;
+	sljit_unaligned_store_uw((void*)addr, new_addr);
 #endif
 }
 
 SLJIT_API_FUNC_ATTRIBUTE void sljit_set_const(sljit_uw addr, sljit_sw new_constant)
 {
-	*(sljit_sw*)addr = new_constant;
+	sljit_unaligned_store_sw((void*)addr, new_constant);
 }
 
 SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_x86_is_sse2_available(void)