The `-mtune=esp-base` option is identical to the default tuning profile,
except that `slow_unaligned_access` is set to false.
This reduces the instruction count for built-in `memcpy` and improves
performance, since our chips can handle misaligned access with minimal
penalty (without triggering exceptions).
Example:
void load(uint32_t *r, char* x) {
memcpy(r, x, sizeof(uint32_t));
}
void store(char* x, uint32_t v) {
memcpy(x, &v, sizeof(uint32_t));
}
Previously generated code:
load:
lbu a5,2(a1)
lbu a3,0(a1)
lbu a4,1(a1)
sb a5,2(a0)
sb a3,0(a0)
sb a4,1(a0)
lbu a5,3(a1)
sb a5,3(a0)
ret
store:
srli a3,a1,8
srli a4,a1,16
srli a5,a1,24
addi sp,sp,-16
sb a1,0(a0)
sb a3,1(a0)
sb a4,2(a0)
sb a5,3(a0)
addi sp,sp,16
jr ra
With `-mtune=esp-base`:
load:
lw a5,0(a1)
sw a5,0(a0)
ret
store:
sw a1,0(a0)
ret
Inlining behavior
=================
Without `-mtune=esp-base`:
- `memcpy()` is inlined only when the compile-time size is ≤ 12 bytes.
- Maximum cost: ~25 instructions
With `-mtune=esp-base`:
- `memcpy()` is inlined for all compile-time constant sizes.
- Maximum cost: ~14 instructions
As a result, some applications may see reduced code size, while others
may increase slightly. However, performance always improves because
extra `memcpy` calls are eliminated.
Performance results
===================
esp32p4 (Ethernet iperf):
- No noticeable difference
esp32c61 (Wi-Fi iperf):
- ~2 Mb/s increase for TCP and UDP TX (may be within measurement error)
NOTE
====
Applies only to RISC-V chips that do not have the hardware issue marked
by the SOC_CPU_MISALIGNED_ACCESS_ON_PMP_MISMATCH_ISSUE macro.
