diff --git a/yjit/src/asm/mod.rs b/yjit/src/asm/mod.rs
index 9df246dcbb2f8b..68c09eb1148130 100644
--- a/yjit/src/asm/mod.rs
+++ b/yjit/src/asm/mod.rs
@@ -2,10 +2,6 @@ use std::cell::RefCell;
 use std::fmt;
 use std::mem;
 use std::rc::Rc;
-#[cfg(target_arch = "x86_64")]
-use crate::backend::x86_64::JMP_PTR_BYTES;
-#[cfg(target_arch = "aarch64")]
-use crate::backend::arm64::JMP_PTR_BYTES;
 use crate::core::IseqPayload;
 use crate::core::for_each_off_stack_iseq_payload;
 use crate::core::for_each_on_stack_iseq_payload;
@@ -123,7 +119,7 @@ impl CodeBlock {
             page_size,
             write_pos: 0,
             past_page_bytes: 0,
-            page_end_reserve: JMP_PTR_BYTES,
+            page_end_reserve: 0,
             label_addrs: Vec::new(),
             label_names: Vec::new(),
             label_refs: Vec::new(),
@@ -133,6 +129,7 @@ impl CodeBlock {
             dropped_bytes: false,
             freed_pages,
         };
+        cb.page_end_reserve = cb.jmp_ptr_bytes();
         cb.write_pos = cb.page_start();
         cb
     }
@@ -196,7 +193,7 @@ impl CodeBlock {
             self.write_pos = dst_pos;
             let dst_ptr = self.get_write_ptr();
             self.write_pos = src_pos;
-            self.without_page_end_reserve(|cb| assert!(cb.has_capacity(JMP_PTR_BYTES)));
+            self.without_page_end_reserve(|cb| assert!(cb.has_capacity(cb.jmp_ptr_bytes())));
 
             // Generate jmp_ptr from src_pos to dst_pos
             self.without_page_end_reserve(|cb| {
@@ -242,6 +239,11 @@ impl CodeBlock {
         self.mem_block.borrow().mapped_region_size()
     }
 
+    /// Size of the region in bytes where writes could be attempted.
+    pub fn virtual_region_size(&self) -> usize {
+        self.mem_block.borrow().virtual_region_size()
+    }
+
     /// Return the number of code pages that have been mapped by the VirtualMemory.
     pub fn num_mapped_pages(&self) -> usize {
         // CodeBlock's page size != VirtualMem's page size on Linux,
@@ -287,7 +289,7 @@ impl CodeBlock {
         if cfg!(debug_assertions) && !cfg!(test) {
             // Leave illegal instructions at the beginning of each page to assert
             // we're not accidentally crossing page boundaries.
-            start += JMP_PTR_BYTES;
+            start += self.jmp_ptr_bytes();
         }
         start
     }
diff --git a/yjit/src/backend/arm64/mod.rs b/yjit/src/backend/arm64/mod.rs
index b85c0320765187..a1823c478eca70 100644
--- a/yjit/src/backend/arm64/mod.rs
+++ b/yjit/src/backend/arm64/mod.rs
@@ -2,6 +2,7 @@
 #![allow(unused_variables)]
 #![allow(unused_imports)]
 
+use crate::asm::x86_64::jmp_ptr;
 use crate::asm::{CodeBlock};
 use crate::asm::arm64::*;
 use crate::codegen::{JITState, CodegenGlobals};
@@ -38,8 +39,25 @@ pub const _C_RET_OPND: Opnd = Opnd::Reg(X0_REG);
 pub const C_SP_REG: A64Opnd = X31;
 pub const C_SP_STEP: i32 = 16;
 
-// The number of bytes that are generated by emit_jmp_ptr
-pub const JMP_PTR_BYTES: usize = 20;
+impl CodeBlock {
+    // The maximum number of bytes that can be generated by emit_jmp_ptr.
+    pub fn jmp_ptr_bytes(&self) -> usize {
+        // b instruction's offset is encoded as imm26 times 4. It can jump to
+        // +/-128MiB, so this can be used when --yjit-exec-mem-size <= 128.
+        let num_insns = if b_offset_fits_bits(self.virtual_region_size() as i64 / 4) {
+            1 // b instruction
+        } else {
+            5 // 4 instructions to load a 64-bit absolute address + br instruction
+        };
+        num_insns * 4
+    }
+
+    // The maximum number of instructions that can be generated by emit_conditional_jump.
+    fn conditional_jump_insns(&self) -> i32 {
+        // The worst case is instructions for a jump + bcond.
+        self.jmp_ptr_bytes() as i32 / 4 + 1
+    }
+}
 
 /// Map Opnd to A64Opnd
 impl From<Opnd> for A64Opnd {
@@ -110,7 +128,8 @@ fn emit_jmp_ptr(cb: &mut CodeBlock, dst_ptr: CodePtr, padding: bool) {
         // Make sure it's always a consistent number of
         // instructions in case it gets patched and has to
         // use the other branch.
-        for _ in num_insns..(JMP_PTR_BYTES / 4) {
+        assert!(num_insns * 4 <= cb.jmp_ptr_bytes());
+        for _ in num_insns..(cb.jmp_ptr_bytes() / 4) {
             nop(cb);
         }
     }
@@ -697,6 +716,18 @@ impl Assembler
                         // Here we're going to return 1 because we've only
                         // written out 1 instruction.
                         1
+                    } else if b_offset_fits_bits((dst_addr - (src_addr + 4)) / 4) { // + 4 for bcond
+                        // If the jump offset fits into the unconditional jump as
+                        // an immediate value, we can use inverse b.cond + b.
+                        //
+                        // We're going to write out the inverse condition so
+                        // that if it doesn't match it will skip over the
+                        // instruction used for branching.
+                        bcond(cb, Condition::inverse(CONDITION), 2.into());
+                        b(cb, InstructionOffset::from_bytes((dst_addr - (src_addr + 4)) as i32)); // + 4 for bcond
+
+                        // We've only written out 2 instructions.
+                        2
                     } else {
                         // Otherwise, we need to load the address into a
                         // register and use the branch register instruction.
@@ -720,7 +751,8 @@ impl Assembler
                         // We need to make sure we have at least 6 instructions for
                         // every kind of jump for invalidation purposes, so we're
                         // going to write out padding nop instructions here.
-                        for _ in num_insns..6 { nop(cb); }
+                        assert!(num_insns <= cb.conditional_jump_insns());
+                        for _ in num_insns..cb.conditional_jump_insns() { nop(cb); }
                     }
                 },
                 Target::Label(label_idx) => {
@@ -1063,7 +1095,7 @@ impl Assembler
                 Insn::RegTemps(_) |
                 Insn::SpillTemp(_) => (), // just a reg alloc signal, no code
                 Insn::PadInvalPatch => {
-                    while (cb.get_write_pos().saturating_sub(std::cmp::max(start_write_pos, cb.page_start_pos()))) < JMP_PTR_BYTES && !cb.has_dropped_bytes() {
+                    while (cb.get_write_pos().saturating_sub(std::cmp::max(start_write_pos, cb.page_start_pos()))) < cb.jmp_ptr_bytes() && !cb.has_dropped_bytes() {
                         nop(cb);
                     }
                 }
diff --git a/yjit/src/backend/x86_64/mod.rs b/yjit/src/backend/x86_64/mod.rs
index dffac1758d1f06..170fc888e0656d 100644
--- a/yjit/src/backend/x86_64/mod.rs
+++ b/yjit/src/backend/x86_64/mod.rs
@@ -34,8 +34,10 @@ pub const _C_ARG_OPNDS: [Opnd; 6] = [
 pub const C_RET_REG: Reg = RAX_REG;
 pub const _C_RET_OPND: Opnd = Opnd::Reg(RAX_REG);
 
-// The number of bytes that are generated by jmp_ptr
-pub const JMP_PTR_BYTES: usize = 6;
+impl CodeBlock {
+    // The number of bytes that are generated by jmp_ptr
+    pub fn jmp_ptr_bytes(&self) -> usize { 6 }
+}
 
 /// Map Opnd to X86Opnd
 impl From<Opnd> for X86Opnd {
@@ -718,8 +720,8 @@ impl Assembler
                 Insn::SpillTemp(_) => (), // just a reg alloc signal, no code
                 Insn::PadInvalPatch => {
                     let code_size = cb.get_write_pos().saturating_sub(std::cmp::max(start_write_pos, cb.page_start_pos()));
-                    if code_size < JMP_PTR_BYTES {
-                        nop(cb, (JMP_PTR_BYTES - code_size) as u32);
+                    if code_size < cb.jmp_ptr_bytes() {
+                        nop(cb, (cb.jmp_ptr_bytes() - code_size) as u32);
                     }
                 }