Move fence inside entry_loop.

This greatly reduces the amount of spinning in the debug loop spike does
when DEBUG_ROM_FLAGS are written by the debug module, because now it's
not getting a cached version over and over again.

Test using riscv-tests/debug. Debugging still works.

debug_rom/debug_rom.S

@@ -20,9 +20,6 @@ exception:
        jal zero, _exception
 
 _entry:
-        // This fence is required because the execution may have written something
-        // into the Abstract Data or Program Buffer registers.
-        fence
         csrw CSR_DSCRATCH0, s0  // Save s0 to allow signaling MHARTID
 
         // We continue to let the hart know that we are halted in order that
@@ -30,6 +27,8 @@ _entry:
         // We keep checking both whether there is something the debugger wants
         // us to do, or whether we should resume.
 entry_loop:
+	// Make sure that we don't access a cached version of DEBUG_ROM_FLAGS
+        fence
         csrr s0, CSR_MHARTID
         sw   s0, DEBUG_ROM_HALTED(zero)
         lbu  s0, DEBUG_ROM_FLAGS(s0) // 1 byte flag per hart. Only one hart advances here.

debug_rom/debug_rom.h

@@ -1,10 +1,10 @@
 static const unsigned char debug_rom_raw[] = {
   0x6f, 0x00, 0xc0, 0x00, 0x6f, 0x00, 0x00, 0x06, 0x6f, 0x00, 0x80, 0x03,
-  0x0f, 0x00, 0xf0, 0x0f, 0x73, 0x10, 0x24, 0x7b, 0x73, 0x24, 0x40, 0xf1,
+  0x73, 0x10, 0x24, 0x7b, 0x0f, 0x00, 0xf0, 0x0f, 0x73, 0x24, 0x40, 0xf1,
   0x23, 0x20, 0x80, 0x10, 0x03, 0x44, 0x04, 0x40, 0x13, 0x74, 0x14, 0x00,
   0x63, 0x14, 0x04, 0x02, 0x73, 0x24, 0x40, 0xf1, 0x03, 0x44, 0x04, 0x40,
   0x13, 0x74, 0x24, 0x00, 0x63, 0x18, 0x04, 0x02, 0x73, 0x00, 0x50, 0x10,
-  0x6f, 0xf0, 0x9f, 0xfd, 0x73, 0x24, 0x20, 0x7b, 0x23, 0x26, 0x00, 0x10,
+  0x6f, 0xf0, 0x5f, 0xfd, 0x73, 0x24, 0x20, 0x7b, 0x23, 0x26, 0x00, 0x10,
   0x73, 0x00, 0x10, 0x00, 0x73, 0x24, 0x40, 0xf1, 0x23, 0x22, 0x80, 0x10,
   0x73, 0x24, 0x20, 0x7b, 0x0f, 0x00, 0xf0, 0x0f, 0x0f, 0x10, 0x00, 0x00,
   0x67, 0x00, 0x00, 0x30, 0x73, 0x24, 0x40, 0xf1, 0x23, 0x24, 0x80, 0x10,