diff --git a/src/pcre2_jit_compile.c b/src/pcre2_jit_compile.c
index 838b7d8..91be12d 100644
--- a/src/pcre2_jit_compile.c
+++ b/src/pcre2_jit_compile.c
@@ -3048,13 +3048,50 @@ else
 static SLJIT_INLINE void reset_fast_fail(compiler_common *common)
 {
 DEFINE_COMPILER;
+sljit_s32 size = common->fast_fail_end_ptr - common->fast_fail_start_ptr;
+sljit_s32 src = SLJIT_IMM;
 sljit_s32 i;
+struct sljit_label *loop;
 
 SLJIT_ASSERT(common->fast_fail_start_ptr < common->fast_fail_end_ptr);
 
-OP2(SLJIT_SUB, TMP1, 0, STR_PTR, 0, SLJIT_IMM, IN_UCHARS(1));
-for (i = common->fast_fail_start_ptr; i < common->fast_fail_end_ptr; i += sizeof(sljit_sw))
-  OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_SP), i, TMP1, 0);
+if (size == sizeof(sljit_sw))
+  {
+  OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_SP), common->fast_fail_start_ptr, SLJIT_IMM, 0);
+  return;
+  }
+
+if (sljit_get_register_index(TMP3) >= 0 && !sljit_has_cpu_feature(SLJIT_HAS_ZERO_REGISTER))
+  {
+  OP1(SLJIT_MOV, TMP3, 0, SLJIT_IMM, 0);
+  src = TMP3;
+  }
+
+if (size <= 6 * sizeof(sljit_sw))
+  {
+  for (i = common->fast_fail_start_ptr; i < common->fast_fail_end_ptr; i += sizeof(sljit_sw))
+    OP1(SLJIT_MOV, SLJIT_MEM1(SLJIT_SP), i, src, 0);
+  return;
+  }
+
+GET_LOCAL_BASE(TMP1, 0, common->fast_fail_start_ptr);
+
+i = ((size / (sljit_s32)sizeof(sljit_sw)) % 3) * sizeof(sljit_sw);
+
+OP2(SLJIT_ADD, TMP2, 0, TMP1, 0, SLJIT_IMM, size - i);
+
+loop = LABEL();
+OP1(SLJIT_MOV, SLJIT_MEM1(TMP1), 0, src, 0);
+OP2(SLJIT_ADD, TMP1, 0, TMP1, 0, SLJIT_IMM, 3 * sizeof(sljit_sw));
+OP1(SLJIT_MOV, SLJIT_MEM1(TMP1), -2 * (sljit_sw)sizeof(sljit_sw), src, 0);
+OP1(SLJIT_MOV, SLJIT_MEM1(TMP1), -1 * (sljit_sw)sizeof(sljit_sw), src, 0);
+CMPTO(SLJIT_LESS, TMP1, 0, TMP2, 0, loop);
+
+if (i >= (sljit_sw)sizeof(sljit_sw))
+  OP1(SLJIT_MOV, SLJIT_MEM1(TMP1), 0, src, 0);
+
+if (i >= 2 * (sljit_sw)sizeof(sljit_sw))
+  OP1(SLJIT_MOV, SLJIT_MEM1(TMP1), sizeof(sljit_sw), src, 0);
 }
 
 static SLJIT_INLINE void do_reset_match(compiler_common *common, int length)
diff --git a/src/sljit/sljitLir.h b/src/sljit/sljitLir.h
index aef90e6..72595bb 100644
--- a/src/sljit/sljitLir.h
+++ b/src/sljit/sljitLir.h
@@ -571,12 +571,14 @@ static SLJIT_INLINE sljit_uw sljit_get_generated_code_size(struct sljit_compiler
 #define SLJIT_HAS_FPU			0
 /* [Limitation] Some registers are virtual registers. */
 #define SLJIT_HAS_VIRTUAL_REGISTERS	1
+/* [Emulated] Has zero register (setting a memory location to zero is efficient). */
+#define SLJIT_HAS_ZERO_REGISTER		2
 /* [Emulated] Count leading zero is supported. */
-#define SLJIT_HAS_CLZ			2
+#define SLJIT_HAS_CLZ			3
 /* [Emulated] Conditional move is supported. */
-#define SLJIT_HAS_CMOV			3
+#define SLJIT_HAS_CMOV			4
 /* [Emulated] Conditional move is supported. */
-#define SLJIT_HAS_PREFETCH		4
+#define SLJIT_HAS_PREFETCH		5
 
 #if (defined SLJIT_CONFIG_X86 && SLJIT_CONFIG_X86)
 /* [Not emulated] SSE2 support is available on x86. */
diff --git a/src/sljit/sljitNativeMIPS_common.c b/src/sljit/sljitNativeMIPS_common.c
index 7628be6..88df904 100644
--- a/src/sljit/sljitNativeMIPS_common.c
+++ b/src/sljit/sljitNativeMIPS_common.c
@@ -684,6 +684,8 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_has_cpu_feature(sljit_s32 feature_type)
 #else
 #error "FIR check is not implemented for this architecture"
 #endif
+	case SLJIT_HAS_ZERO_REGISTER:
+		return 1;
 
 #if (defined SLJIT_MIPS_REV && SLJIT_MIPS_REV >= 1)
 	case SLJIT_HAS_CLZ:
diff --git a/src/sljit/sljitNativePPC_common.c b/src/sljit/sljitNativePPC_common.c
index 17bf9a9..590f91c 100644
--- a/src/sljit/sljitNativePPC_common.c
+++ b/src/sljit/sljitNativePPC_common.c
@@ -626,6 +626,8 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_has_cpu_feature(sljit_s32 feature_type)
 		return 1;
 #endif
 
+	/* A saved register is set to a zero value. */
+	case SLJIT_HAS_ZERO_REGISTER:
 	case SLJIT_HAS_CLZ:
 	case SLJIT_HAS_PREFETCH:
 		return 1;
diff --git a/src/sljit/sljitNativeSPARC_common.c b/src/sljit/sljitNativeSPARC_common.c
index 4c95350..7d6be6c 100644
--- a/src/sljit/sljitNativeSPARC_common.c
+++ b/src/sljit/sljitNativeSPARC_common.c
@@ -451,6 +451,9 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_has_cpu_feature(sljit_s32 feature_type)
 		return 1;
 #endif
 
+	case SLJIT_HAS_ZERO_REGISTER:
+		return 1;
+
 #if (defined SLJIT_CONFIG_SPARC_64 && SLJIT_CONFIG_SPARC_64)
 	case SLJIT_HAS_CMOV:
 		return 1;
diff --git a/src/sljit/sljitNativeX86_common.c b/src/sljit/sljitNativeX86_common.c
index 9ced936..74965e3 100644
--- a/src/sljit/sljitNativeX86_common.c
+++ b/src/sljit/sljitNativeX86_common.c
@@ -2316,6 +2316,10 @@ SLJIT_API_FUNC_ATTRIBUTE sljit_s32 sljit_emit_op2(struct sljit_compiler *compile
 		if (!HAS_FLAGS(op)) {
 			if ((src2 & SLJIT_IMM) && emit_lea_binary(compiler, dst, dstw, src1, src1w, SLJIT_IMM, -src2w) != SLJIT_ERR_UNSUPPORTED)
 				return compiler->error;
+			if (SLOW_IS_REG(dst) && src2 == dst) {
+				FAIL_IF(emit_non_cum_binary(compiler, BINARY_OPCODE(SUB), dst, 0, dst, 0, src1, src1w));
+				return emit_unary(compiler, NEG_rm, dst, 0, dst, 0);
+			}
 		}
 
 		if (dst == SLJIT_UNUSED)