realtek: rt-loader: use -O2 with selective LZMA inlining

Build rt-loader with -O2 and tune the LZMA decoder's inlining decisions
to balance code size and decompression performance.

The existing -O3 setting applies to all loader C sources. In the
standalone decoder comparison used to investigate this change, -O3
substantially increased the generated code size. Simply switching to
-O2 reduced that footprint, but left the frequently used rc_get_bit()
helper out of line, adding function-call and stack-access overhead
throughout range decoding.

Adjust the two helpers independently:

- Declare rc_get_bit() inline to encourage integration into its callers.
  This lets the compiler eliminate call overhead and optimise accesses
  to the range-coder state, probability values and decoded symbols.
- Mark rc_do_normalize() noinline to retain a shared implementation of
  the conditional normalisation path. This limits code duplication as
  the surrounding bit-decoding helpers are inlined.

Use the inline keyword rather than always_inline for rc_get_bit().
With the tested compiler, forcing this helper to inline provided no
additional benefit over the simpler declaration.

A standalone comparison with GCC 13.2.0 targeting MIPS32r2/34Kc used
the same synthetic 1 MiB LZMA stream, producing 4 MiB of output. A
generic single-issue in-order model with a 32 KiB instruction cache
and a 16 KiB data cache produced the following results:

- Original -O2: 4,960 bytes of decoder .text; 548,160,936 model cycles.
- Modified -O2: 4,960 bytes of decoder .text; 389,069,748 model cycles.
- Original -O3: 10,496 bytes of decoder .text; 336,140,240 model cycles.

In that setup, the modified -O2 decoder retains approximately 75% of
the cycle reduction achieved by -O3 while using 52.7% less code.
It requires 15.7% more model cycles than -O3. Compared with unmodified
-O2, it reduces model cycles by 29.0% without increasing code size.
All compared variants produced byte-identical output.

These figures describe the standalone decoder experiment, not measured
Realtek hardware performance. The experiment used -fno-pic and
-mno-abicalls; rt-loader retains -fpic and -mabicalls. Actual loader
size and boot-time effects therefore require validation with the
target build and hardware.

The patch changes the optimisation level for the whole loader and
the inlining policy of these two decoder helpers. It leaves the
decompression algorithm, bitstream format and error-handling logic
unchanged.

Assisted-by: ChatGPT (OpenAI GPT-6 Astra)
Link: openwrt/openwrt#25813
Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>

Author: plappermaul

target/linux/realtek/image/rt-loader/Makefile

@@ -73,7 +73,7 @@ LD		:= $(CROSS_COMPILE)ld
 OBJCOPY		:= $(CROSS_COMPILE)objcopy
 OBJDUMP		:= $(CROSS_COMPILE)objdump
 
-CFLAGS		= -fpic -mabicalls -O3 -fno-builtin-printf -Iinclude
+CFLAGS		= -fpic -mabicalls -O2 -fno-builtin-printf -Iinclude
 CFLAGS		+= -DFLASH_ADDR=$(FLASH_ADDR)
 CFLAGS		+= -DKERNEL_ADDR=$(KERNEL_ADDR)
 

target/linux/realtek/image/rt-loader/src/unlzma.c

@@ -124,7 +124,7 @@ static inline void rc_init_code(struct rc *rc)
 
 
 /* Called twice, but one callsite is in inline'd rc_is_bit_0_helper() */
-static void rc_do_normalize(struct rc *rc)
+static __attribute__((noinline)) void rc_do_normalize(struct rc *rc)
 {
 	if (rc->ptr >= rc->buffer_end)
 		rc_read(rc);
@@ -167,7 +167,7 @@ static inline void rc_update_bit_1(struct rc *rc, uint16_t *p)
 }
 
 /* Called 4 times in unlzma loop */
-static int rc_get_bit(struct rc *rc, uint16_t *p, int *symbol)
+static inline int rc_get_bit(struct rc *rc, uint16_t *p, int *symbol)
 {
 	if (rc_is_bit_0(rc, p)) {
 		rc_update_bit_0(rc, p);