diff --git a/sw/device/boot_rom/rom_crt.S b/sw/device/boot_rom/rom_crt.S
index 2009f91..6713de5 100644
--- a/sw/device/boot_rom/rom_crt.S
+++ b/sw/device/boot_rom/rom_crt.S
@@ -54,6 +54,13 @@
   // Set up the stack.
   la  sp, _stack_start
 
+  // Set up the global pointer. This requires that we disable linker relaxations
+  // (or it will be relaxed to `mv gp, gp`).
+  .option push
+  .option norelax
+  la  gp, __global_pointer$
+  .option pop
+
   // Explicit fall-through to |_start|.
 
 /**
diff --git a/sw/device/boot_rom/rom_link.ld b/sw/device/boot_rom/rom_link.ld
index 59f93a3..dfb328c 100644
--- a/sw/device/boot_rom/rom_link.ld
+++ b/sw/device/boot_rom/rom_link.ld
@@ -100,9 +100,19 @@
    */
   .data ORIGIN(ram): AT(_data_init_start) ALIGN(4) {
     _data_start = .;
+    __global_pointer$ = . + 2048;
+
+    /* Small data should come before larger data. This helps to ensure small
+     * globals are within 2048 bytes of the value of `gp`, making their accesses
+     * hopefully only take one instruction. */
+    *(.sdata)
+    *(.sdata.*)
+
+    /* Other data will likely need multiple instructions to load, so we're less
+     * concerned about address materialisation taking more than one instruction.
+     */
     *(.data)
     *(.data.*)
-    *(.sdata)
     _data_end = .;
   } > ram
 
@@ -118,10 +128,12 @@
    */
   .bss : ALIGN(4) {
     _bss_start = .;
-    *(.bss)
-    *(.bss.*)
+    /* Small BSS comes before regular BSS for the same reasons as in the data
+     * section */
     *(.sbss)
     *(.sbss.*)
+    *(.bss)
+    *(.bss.*)
     *(COMMON)
     _bss_end = .;
   } > ram
diff --git a/sw/device/exts/common/flash_crt.S b/sw/device/exts/common/flash_crt.S
index c082356..0ffac3b 100644
--- a/sw/device/exts/common/flash_crt.S
+++ b/sw/device/exts/common/flash_crt.S
@@ -25,6 +25,13 @@
   // jumps here will have the correct stack start linked in.
   la sp, _stack_start
 
+  // Set up the global pointer. This requires that we disable linker relaxations
+  // (or it will be relaxed to `mv gp, gp`).
+  .option push
+  .option norelax
+  la gp, __global_pointer$
+  .option pop
+
   // Set up the new interrupt vector.
   la   t0, _vectors_start
   csrw mtvec, t0
diff --git a/sw/device/exts/common/flash_link.ld b/sw/device/exts/common/flash_link.ld
index 3e41e5c..49c09f3 100644
--- a/sw/device/exts/common/flash_link.ld
+++ b/sw/device/exts/common/flash_link.ld
@@ -92,9 +92,19 @@
    */ 
   .data ORIGIN(ram): AT(_data_init_start) ALIGN(4) {
     _data_start = .;
+    __global_pointer$ = . + 2048;
+
+    /* Small data should come before larger data. This helps to ensure small
+     * globals are within 2048 bytes of the value of `gp`, making their accesses
+     * hopefully only take one instruction. */
+    *(.sdata)
+    *(.sdata.*)
+
+    /* Other data will likely need multiple instructions to load, so we're less
+     * concerned about address materialisation taking more than one instruction.
+     */
     *(.data)
     *(.data.*)
-    *(.sdata)
     _data_end = .;
   } > ram
 
@@ -103,10 +113,12 @@
    */
   .bss : ALIGN(4) {
     _bss_start = .;
-    *(.bss)
-    *(.bss.*)
+    /* Small BSS comes before regular BSS for the same reasons as in the data
+     * section */
     *(.sbss)
     *(.sbss.*)
+    *(.bss)
+    *(.bss.*)
     *(COMMON)
     _bss_end = .;
   } > ram