diff --git a/dex2oat/linker/oat_writer.cc b/dex2oat/linker/oat_writer.cc
index 6b23883..3c49634 100644
--- a/dex2oat/linker/oat_writer.cc
+++ b/dex2oat/linker/oat_writer.cc
@@ -2194,7 +2194,8 @@
     size_t adjusted_offset = offset;
 
     #define DO_TRAMPOLINE(field, fn_name)                                   \
-      offset = CompiledCode::AlignCode(offset, instruction_set);            \
+      /* Pad with at least four 0xFFs so we can do DCHECKs in OatQuickMethodHeader */ \
+      offset = CompiledCode::AlignCode(offset + 4, instruction_set);        \
       adjusted_offset = offset + CompiledCode::CodeDelta(instruction_set);  \
       oat_header_->Set ## fn_name ## Offset(adjusted_offset);               \
       (field) = compiler_driver_->Create ## fn_name();                      \
@@ -3070,9 +3071,13 @@
 
     #define DO_TRAMPOLINE(field) \
       do { \
-        uint32_t aligned_offset = CompiledCode::AlignCode(relative_offset, instruction_set); \
+        /* Pad with at least four 0xFFs so we can do DCHECKs in OatQuickMethodHeader */ \
+        uint32_t aligned_offset = CompiledCode::AlignCode(relative_offset + 4, instruction_set); \
         uint32_t alignment_padding = aligned_offset - relative_offset; \
-        out->Seek(alignment_padding, kSeekCurrent); \
+        for (size_t i = 0; i < alignment_padding; i++) { \
+          uint8_t padding = 0xFF; \
+          out->WriteFully(&padding, 1); \
+        } \
         size_trampoline_alignment_ += alignment_padding; \
         if (!out->WriteFully((field)->data(), (field)->size())) { \
           PLOG(ERROR) << "Failed to write " # field " to " << out->GetLocation(); \
diff --git a/runtime/arch/arm/asm_support_arm.S b/runtime/arch/arm/asm_support_arm.S
index eeac743..633591d 100644
--- a/runtime/arch/arm/asm_support_arm.S
+++ b/runtime/arch/arm/asm_support_arm.S
@@ -54,7 +54,7 @@
 // Common ENTRY declaration code for ARM and thumb, an ENTRY should always be paired with an END.
 // Declares the RUNTIME_CURRENT[123] macros that can be used within an ENTRY and will have literals
 // generated at END.
-.macro DEF_ENTRY thumb_or_arm, name
+.macro DEF_ENTRY thumb_or_arm, name, alignment
     \thumb_or_arm
 // Clang ignores .thumb_func and requires an explicit .thumb. Investigate whether we should still
 // carry around the .thumb_func.
@@ -64,8 +64,12 @@
     .type \name, #function
     .hidden \name  // Hide this as a global symbol, so we do not incur plt calls.
     .global \name
+    // ART-compiled functions have OatQuickMethodHeader but assembly funtions do not.
+    // Prefix the assembly code with 0xFFs, which means there is no method header.
+    .byte 0xFF, 0xFF, 0xFF, 0xFF
     // Cache alignment for function entry.
-    .balign 16
+    // NB: 0xFF because there is a bug in balign where 0x00 creates nop instructions.
+    .balign \alignment, 0xFF
 \name:
     .cfi_startproc
     .fnstart
@@ -88,12 +92,15 @@
 
 // A thumb2 style ENTRY.
 .macro ENTRY name
-    DEF_ENTRY .thumb_func, \name
+    DEF_ENTRY .thumb_func, \name, 16
+.endm
+.macro ENTRY_ALIGNED name, alignment
+    DEF_ENTRY .thumb_func, \name, \alignment
 .endm
 
 // A ARM style ENTRY.
 .macro ARM_ENTRY name
-    DEF_ENTRY .arm, \name
+    DEF_ENTRY .arm, \name, 16
 .endm
 
 // Terminate an ENTRY and generate GOT_PREL references.
diff --git a/runtime/arch/arm/quick_entrypoints_arm.S b/runtime/arch/arm/quick_entrypoints_arm.S
index 1153a77..3450444 100644
--- a/runtime/arch/arm/quick_entrypoints_arm.S
+++ b/runtime/arch/arm/quick_entrypoints_arm.S
@@ -2624,8 +2624,7 @@
      *     (6 bytes). Loads the return register and jumps to the runtime call.
      */
 #if defined(USE_READ_BARRIER) && defined(USE_BAKER_READ_BARRIER)
-    .balign 512
-ENTRY art_quick_read_barrier_mark_introspection
+ENTRY_ALIGNED art_quick_read_barrier_mark_introspection, 512
     // At this point, IP contains the reference, rMR is clobbered by the thunk
     // and can be freely used as it will be set back to 1 before returning.
     // For heap poisoning, the reference is poisoned, so unpoison it first.
diff --git a/runtime/arch/arm64/asm_support_arm64.S b/runtime/arch/arm64/asm_support_arm64.S
index 715fc35..3eb0991 100644
--- a/runtime/arch/arm64/asm_support_arm64.S
+++ b/runtime/arch/arm64/asm_support_arm64.S
@@ -40,16 +40,24 @@
 #define wMR w20
 #endif
 
-.macro ENTRY name
+.macro ENTRY_ALIGNED name, alignment
     .type \name, #function
     .hidden \name  // Hide this as a global symbol, so we do not incur plt calls.
     .global \name
-    /* Cache alignment for function entry */
-    .balign 16
+    // ART-compiled functions have OatQuickMethodHeader but assembly funtions do not.
+    // Prefix the assembly code with 0xFFs, which means there is no method header.
+    .byte 0xFF, 0xFF, 0xFF, 0xFF
+    // Cache alignment for function entry.
+    // NB: 0xFF because there is a bug in balign where 0x00 creates nop instructions.
+    .balign \alignment, 0xFF
 \name:
     .cfi_startproc
 .endm
 
+.macro ENTRY name
+    ENTRY_ALIGNED \name, 16
+.endm
+
 .macro END name
     .cfi_endproc
     .size \name, .-\name
diff --git a/runtime/arch/arm64/quick_entrypoints_arm64.S b/runtime/arch/arm64/quick_entrypoints_arm64.S
index 9f3377e..5945c45 100644
--- a/runtime/arch/arm64/quick_entrypoints_arm64.S
+++ b/runtime/arch/arm64/quick_entrypoints_arm64.S
@@ -2790,8 +2790,7 @@
      *   art_quick_read_barrier_mark_introspection_gc_roots:
      *     GC root entrypoint code.
      */
-    .balign 512
-ENTRY art_quick_read_barrier_mark_introspection
+ENTRY_ALIGNED art_quick_read_barrier_mark_introspection, 512
     // At this point, IP0 contains the reference, IP1 can be freely used.
     // For heap poisoning, the reference is poisoned, so unpoison it first.
     UNPOISON_HEAP_REF wIP0
diff --git a/runtime/arch/x86/asm_support_x86.S b/runtime/arch/x86/asm_support_x86.S
index c9514f5..cd5ebd7 100644
--- a/runtime/arch/x86/asm_support_x86.S
+++ b/runtime/arch/x86/asm_support_x86.S
@@ -113,7 +113,11 @@
 
     /* Cache alignment for function entry */
 MACRO0(ALIGN_FUNCTION_ENTRY)
-    .balign 16
+    // ART-compiled functions have OatQuickMethodHeader but assembly funtions do not.
+    // Prefix the assembly code with 0xFFs, which means there is no method header.
+    .byte 0xFF, 0xFF, 0xFF, 0xFF
+    // Cache alignment for function entry.
+    .balign 16, 0xFF
 END_MACRO
 
 MACRO2(DEFINE_FUNCTION_CUSTOM_CFA, c_name, cfa_offset)
diff --git a/runtime/arch/x86_64/asm_support_x86_64.S b/runtime/arch/x86_64/asm_support_x86_64.S
index 28018c5..6b09a6e 100644
--- a/runtime/arch/x86_64/asm_support_x86_64.S
+++ b/runtime/arch/x86_64/asm_support_x86_64.S
@@ -109,7 +109,11 @@
 
     /* Cache alignment for function entry */
 MACRO0(ALIGN_FUNCTION_ENTRY)
-    .balign 16
+    // ART-compiled functions have OatQuickMethodHeader but assembly funtions do not.
+    // Prefix the assembly code with 0xFFs, which means there is no method header.
+    .byte 0xFF, 0xFF, 0xFF, 0xFF
+    // Cache alignment for function entry.
+    .balign 16, 0xFF
 END_MACRO
 
 // TODO: we might need to use SYMBOL() here to add the underscore prefix
diff --git a/runtime/oat_quick_method_header.h b/runtime/oat_quick_method_header.h
index 8798c69..e41c7ee 100644
--- a/runtime/oat_quick_method_header.h
+++ b/runtime/oat_quick_method_header.h
@@ -76,6 +76,10 @@
   }
 
   uint32_t GetCodeSize() const {
+    // ART compiled method are prefixed with header, but we can also easily
+    // accidentally use a function pointer to one of the stubs/trampolines.
+    // We prefix those with 0xFF in the aseembly so that we can do DCHECKs.
+    CHECK_NE(code_size_, 0xFFFFFFFF) << code_;
     return code_size_ & kCodeSizeMask;
   }