Merge remote-tracking branch 'upstream/inline-micros-esp32' into integration

# Conflicts:
#	esphome/core/hal.h

esphome/components/esp8266/core.cpp

@@ -84,7 +84,7 @@ void HOT delay(uint32_t ms) {
     optimistic_yield(1000);
   }
 }
-void IRAM_ATTR HOT delayMicroseconds(uint32_t us) { delay_microseconds_safe(us); }
+// delayMicroseconds(), arch_feed_wdt(), and progmem_read_*() are inlined in hal/hal_esp8266.h.
 void arch_restart() {
   system_restart();
   // restart() doesn't always end execution
@@ -93,17 +93,6 @@ void arch_restart() {
   }
 }
 void arch_init() {}
-void HOT arch_feed_wdt() { system_soft_wdt_feed(); }
-
-uint8_t progmem_read_byte(const uint8_t *addr) {
-  return pgm_read_byte(addr);  // NOLINT
-}
-const char *progmem_read_ptr(const char *const *addr) {
-  return reinterpret_cast<const char *>(pgm_read_ptr(addr));  // NOLINT
-}
-uint16_t progmem_read_uint16(const uint16_t *addr) {
-  return pgm_read_word(addr);  // NOLINT
-}
 uint32_t IRAM_ATTR HOT arch_get_cpu_cycle_count() { return esp_get_cycle_count(); }
 uint32_t arch_get_cpu_freq_hz() { return F_CPU; }
 

esphome/core/hal.h

@@ -6,190 +6,44 @@
 #include "esphome/core/time_64.h"
 #include "esphome/core/time_conversion.h"
 
+// Per-platform HAL bits (IRAM_ATTR / PROGMEM macros, in_isr_context(),
+// inline yield/delay/micros/millis/millis_64 wrappers, ESP8266 progmem
+// helpers) live under esphome/core/hal/ and are dispatched here based on
+// the active USE_* platform define. Each header guards its body with the
+// matching #ifdef USE_<platform> and re-enters namespace esphome {} so it
+// is safe to be re-included.
 #if defined(USE_ESP32)
-#include <esp_attr.h>
-#ifndef PROGMEM
-#define PROGMEM
-#endif
-
+#include "esphome/core/hal/hal_esp32.h"
 #elif defined(USE_ESP8266)
-
-#include <c_types.h>
-#ifndef PROGMEM
-#define PROGMEM ICACHE_RODATA_ATTR
-#endif
-
-#elif defined(USE_RP2040)
-
-#define IRAM_ATTR __attribute__((noinline, long_call, section(".time_critical")))
-#define PROGMEM
-
+#include "esphome/core/hal/hal_esp8266.h"
 #elif defined(USE_LIBRETINY)
-
-// IRAM_ATTR places a function in executable RAM so it is callable from an
-// ISR even while flash is busy (XIP stall, OTA, logger flash write).
-// Each family uses a section its stock linker already routes to RAM:
-// RTL8710B → .image2.ram.text, RTL8720C → .sram.text. LN882H is the
-// exception: its stock linker has no matching glob, so patch_linker.py
-// injects KEEP(*(.sram.text*)) into .flash_copysection at pre-link.
-//
-// BK72xx (all variants) are left as a no-op: their SDK wraps flash
-// operations in GLOBAL_INT_DISABLE() which masks FIQ + IRQ at the CPU for
-// the duration of every write, so no ISR fires while flash is stalled and
-// the race IRAM_ATTR guards against cannot occur. The trade-off is that
-// interrupts are delayed (not dropped) by up to ~20 ms during a sector
-// erase, but that is an SDK-level choice and cannot be changed from this
-// layer.
-#if defined(USE_BK72XX)
-#define IRAM_ATTR
-#elif defined(USE_LIBRETINY_VARIANT_RTL8710B)
-// Stock linker consumes *(.image2.ram.text*) into .ram_image2.text (> BD_RAM).
-#define IRAM_ATTR __attribute__((noinline, section(".image2.ram.text")))
+#include "esphome/core/hal/hal_libretiny.h"
+#elif defined(USE_RP2040)
+#include "esphome/core/hal/hal_rp2040.h"
+#elif defined(USE_HOST)
+#include "esphome/core/hal/hal_host.h"
+#elif defined(USE_ZEPHYR)
+#include "esphome/core/hal/hal_zephyr.h"
 #else
-// RTL8720C: stock linker consumes *(.sram.text*) into .ram.code_text.
-// LN882H: patch_linker.py.script injects *(.sram.text*) into
-// .flash_copysection (> RAM0 AT> FLASH).
-#define IRAM_ATTR __attribute__((noinline, section(".sram.text")))
-#endif
-#define PROGMEM
-
-#else
-
-#define IRAM_ATTR
-#define PROGMEM
-
-#endif
-
-#ifdef USE_ESP32
-#include <freertos/FreeRTOS.h>
-#include <freertos/task.h>
-#endif
-
-#ifdef USE_LIBRETINY
-// For the inline millis() fast paths (xTaskGetTickCount, portTICK_PERIOD_MS).
-#include <FreeRTOS.h>
-#include <task.h>
-#endif
-
-#ifdef USE_BK72XX
-// Declared in the Beken FreeRTOS port (portmacro.h) and built in ARM mode so
-// it is callable from Thumb code via interworking. The MRS CPSR instruction
-// is ARM-only and user code here may be built in Thumb, so in_isr_context()
-// defers to this port helper on BK72xx instead of reading CPSR inline.
-extern "C" uint32_t platform_is_in_interrupt_context(void);
-#endif
-
-// Forward decls from Arduino's <Arduino.h> for the inline wrappers below.
-// NOLINT covers TUs that also include Arduino.h.
-#if defined(USE_ESP8266) || defined(USE_LIBRETINY) || defined(USE_RP2040)
-// NOLINTBEGIN(google-runtime-int,readability-identifier-naming,readability-redundant-declaration)
-extern "C" void yield(void);
-extern "C" void delay(unsigned long ms);
-extern "C" unsigned long micros(void);
-extern "C" unsigned long millis(void);
-// NOLINTEND(google-runtime-int,readability-identifier-naming,readability-redundant-declaration)
-#endif
-
-#ifdef USE_RP2040
-// Forward decl from <pico/time.h>.
-extern "C" uint64_t time_us_64(void);
+#error "hal.h: not implemented for this platform"
 #endif
 
 namespace esphome {
 
-/// Returns true when executing inside an interrupt handler.
-/// always_inline so callers placed in IRAM keep the detection in IRAM.
-__attribute__((always_inline)) inline bool in_isr_context() {
-#if defined(USE_ESP32)
-  return xPortInIsrContext() != 0;
-#elif defined(USE_ESP8266)
-  // ESP8266 has no reliable single-register ISR detection: PS.INTLEVEL is
-  // non-zero both in a real ISR and when user code masks interrupts.  The
-  // ESP8266 wake path is context-agnostic (wake_loop_impl uses esp_schedule
-  // which is ISR-safe) so this helper is unused on this platform.
-  return false;
-#elif defined(USE_RP2040)
-  uint32_t ipsr;
-  __asm__ volatile("mrs %0, ipsr" : "=r"(ipsr));
-  return ipsr != 0;
-#elif defined(USE_BK72XX)
-  // BK72xx is ARM968E-S (ARM9); see extern declaration above.
-  return platform_is_in_interrupt_context() != 0;
-#elif defined(USE_LIBRETINY)
-  // Cortex-M (AmebaZ, AmebaZ2, LN882H). IPSR is the active exception number;
-  // non-zero means we're in a handler.
-  uint32_t ipsr;
-  __asm__ volatile("mrs %0, ipsr" : "=r"(ipsr));
-  return ipsr != 0;
-#else
-  // Host and any future platform without an ISR concept.
-  return false;
-#endif
-}
-
-// yield()/delay()/micros()/millis()/millis_64() are inlined per platform to
-// drop the wrapper call/return — most relevant to runtime_stats and the main
-// loop. ESP8266/LibreTiny/RP2040 share Arduino's ::yield/::delay/::micros.
-#if defined(USE_ESP32)
-// Forward decl from <esp_timer.h>.
-// NOLINTNEXTLINE(readability-redundant-declaration)
-extern "C" int64_t esp_timer_get_time(void);
-__attribute__((always_inline)) inline void yield() { vPortYield(); }
-__attribute__((always_inline)) inline void delay(uint32_t ms) { vTaskDelay(ms / portTICK_PERIOD_MS); }
-__attribute__((always_inline)) inline uint32_t micros() { return static_cast<uint32_t>(esp_timer_get_time()); }
-uint32_t millis();
-__attribute__((always_inline)) inline uint64_t millis_64() {
-  return micros_to_millis<uint64_t>(static_cast<uint64_t>(esp_timer_get_time()));
-}
-#elif defined(USE_ESP8266) || defined(USE_LIBRETINY) || defined(USE_RP2040)
-__attribute__((always_inline)) inline void yield() { ::yield(); }
-__attribute__((always_inline)) inline uint32_t micros() { return static_cast<uint32_t>(::micros()); }
-#if defined(USE_ESP8266)
-void delay(uint32_t ms);
-uint32_t millis();
-__attribute__((always_inline)) inline uint64_t millis_64() { return Millis64Impl::compute(millis()); }
-#elif defined(USE_LIBRETINY)
-__attribute__((always_inline)) inline void delay(uint32_t ms) { ::delay(ms); }
-// Per-variant millis() fast path — matches MillisInternal::get().
-#if defined(USE_RTL87XX) || defined(USE_LN882X)
-static_assert(configTICK_RATE_HZ == 1000, "millis() fast path requires 1 kHz FreeRTOS tick");
-__attribute__((always_inline)) inline uint32_t millis() {
-  // xTaskGetTickCountFromISR is mandatory in interrupt context per the FreeRTOS API contract.
-  return in_isr_context() ? xTaskGetTickCountFromISR() : xTaskGetTickCount();
-}
-#elif defined(USE_BK72XX)
-static_assert(configTICK_RATE_HZ == 500, "BK72xx millis() fast path assumes 500 Hz FreeRTOS tick");
-__attribute__((always_inline)) inline uint32_t millis() { return xTaskGetTickCount() * portTICK_PERIOD_MS; }
-#else
-__attribute__((always_inline)) inline uint32_t millis() { return static_cast<uint32_t>(::millis()); }
-#endif
-__attribute__((always_inline)) inline uint64_t millis_64() { return Millis64Impl::compute(millis()); }
-#else  // USE_RP2040
-__attribute__((always_inline)) inline void delay(uint32_t ms) { ::delay(ms); }
-__attribute__((always_inline)) inline uint32_t millis() { return micros_to_millis(::time_us_64()); }
-__attribute__((always_inline)) inline uint64_t millis_64() { return micros_to_millis<uint64_t>(::time_us_64()); }
-#endif
-#else
-void yield();
-void delay(uint32_t ms);
-uint32_t micros();
-uint32_t millis();
-uint64_t millis_64();
-#endif
+// ESP8266 inlines delayMicroseconds() and arch_feed_wdt() in hal/hal_esp8266.h;
+// every other platform keeps them out-of-line in components/<platform>/core.cpp.
+#ifndef USE_ESP8266
 void delayMicroseconds(uint32_t us);  // NOLINT(readability-identifier-naming)
+void arch_feed_wdt();
+#endif
 void __attribute__((noreturn)) arch_restart();
 void arch_init();
-void arch_feed_wdt();
 uint32_t arch_get_cpu_cycle_count();
 uint32_t arch_get_cpu_freq_hz();
 
-#ifdef USE_ESP8266
-// ESP8266: pgm_read_* does real flash reads on Harvard architecture
-uint8_t progmem_read_byte(const uint8_t *addr);
-const char *progmem_read_ptr(const char *const *addr);
-uint16_t progmem_read_uint16(const uint16_t *addr);
-#else
-// All other platforms: PROGMEM is a no-op, so these are direct dereferences
+#ifndef USE_ESP8266
+// All non-ESP8266 platforms: PROGMEM is a no-op, so these are direct dereferences.
+// ESP8266's out-of-line declarations live in hal/hal_esp8266.h.
 inline uint8_t progmem_read_byte(const uint8_t *addr) { return *addr; }
 inline const char *progmem_read_ptr(const char *const *addr) { return *addr; }
 inline uint16_t progmem_read_uint16(const uint16_t *addr) { return *addr; }

esphome/core/hal/hal_esp32.h

@@ -0,0 +1,35 @@
+#pragma once
+
+#ifdef USE_ESP32
+
+#include <cstdint>
+#include <esp_attr.h>
+#include <freertos/FreeRTOS.h>
+#include <freertos/task.h>
+
+#include "esphome/core/time_conversion.h"
+
+#ifndef PROGMEM
+#define PROGMEM
+#endif
+
+namespace esphome {
+
+/// Returns true when executing inside an interrupt handler.
+__attribute__((always_inline)) inline bool in_isr_context() { return xPortInIsrContext() != 0; }
+
+// Forward decl from <esp_timer.h>.
+// NOLINTNEXTLINE(readability-redundant-declaration)
+extern "C" int64_t esp_timer_get_time(void);
+
+__attribute__((always_inline)) inline void yield() { vPortYield(); }
+__attribute__((always_inline)) inline void delay(uint32_t ms) { vTaskDelay(ms / portTICK_PERIOD_MS); }
+__attribute__((always_inline)) inline uint32_t micros() { return static_cast<uint32_t>(esp_timer_get_time()); }
+uint32_t millis();
+__attribute__((always_inline)) inline uint64_t millis_64() {
+  return micros_to_millis<uint64_t>(static_cast<uint64_t>(esp_timer_get_time()));
+}
+
+}  // namespace esphome
+
+#endif  // USE_ESP32

esphome/core/hal/hal_esp8266.h

@@ -0,0 +1,65 @@
+#pragma once
+
+#ifdef USE_ESP8266
+
+#include <c_types.h>
+#include <cstdint>
+#include <pgmspace.h>
+
+#include "esphome/core/time_64.h"
+
+#ifndef PROGMEM
+#define PROGMEM ICACHE_RODATA_ATTR
+#endif
+
+// Forward decls from Arduino's <Arduino.h> for the inline wrappers below.
+// NOLINTBEGIN(google-runtime-int,readability-identifier-naming,readability-redundant-declaration)
+extern "C" void yield(void);
+extern "C" void delay(unsigned long ms);
+extern "C" unsigned long micros(void);
+extern "C" unsigned long millis(void);
+// NOLINTEND(google-runtime-int,readability-identifier-naming,readability-redundant-declaration)
+
+// Forward decl from <user_interface.h> for arch_feed_wdt() inline below.
+// NOLINTNEXTLINE(readability-redundant-declaration)
+extern "C" void system_soft_wdt_feed(void);
+
+namespace esphome {
+
+// Forward decl from helpers.h so this header stays cheap.
+// NOLINTNEXTLINE(readability-redundant-declaration)
+void delay_microseconds_safe(uint32_t us);
+
+/// Returns true when executing inside an interrupt handler.
+/// ESP8266 has no reliable single-register ISR detection: PS.INTLEVEL is
+/// non-zero both in a real ISR and when user code masks interrupts.  The
+/// ESP8266 wake path is context-agnostic (wake_loop_impl uses esp_schedule
+/// which is ISR-safe) so this helper is unused on this platform.
+__attribute__((always_inline)) inline bool in_isr_context() { return false; }
+
+__attribute__((always_inline)) inline void yield() { ::yield(); }
+__attribute__((always_inline)) inline uint32_t micros() { return static_cast<uint32_t>(::micros()); }
+void delay(uint32_t ms);
+uint32_t millis();
+__attribute__((always_inline)) inline uint64_t millis_64() { return Millis64Impl::compute(millis()); }
+
+// ESP8266: pgm_read_* does aligned 32-bit flash reads on Harvard architecture.
+// Inline-forward to the platform macros so the wrappers themselves don't
+// occupy IRAM/flash on every call site.
+__attribute__((always_inline)) inline uint8_t progmem_read_byte(const uint8_t *addr) {
+  return pgm_read_byte(addr);  // NOLINT
+}
+__attribute__((always_inline)) inline const char *progmem_read_ptr(const char *const *addr) {
+  return reinterpret_cast<const char *>(pgm_read_ptr(addr));  // NOLINT
+}
+__attribute__((always_inline)) inline uint16_t progmem_read_uint16(const uint16_t *addr) {
+  return pgm_read_word(addr);  // NOLINT
+}
+
+// NOLINTNEXTLINE(readability-identifier-naming)
+__attribute__((always_inline)) inline void delayMicroseconds(uint32_t us) { delay_microseconds_safe(us); }
+__attribute__((always_inline)) inline void arch_feed_wdt() { system_soft_wdt_feed(); }
+
+}  // namespace esphome
+
+#endif  // USE_ESP8266

esphome/core/hal/hal_host.h

@@ -0,0 +1,24 @@
+#pragma once
+
+#ifdef USE_HOST
+
+#include <cstdint>
+
+#define IRAM_ATTR
+#define PROGMEM
+
+namespace esphome {
+
+/// Returns true when executing inside an interrupt handler.
+/// Host has no ISR concept.
+__attribute__((always_inline)) inline bool in_isr_context() { return false; }
+
+void yield();
+void delay(uint32_t ms);
+uint32_t micros();
+uint32_t millis();
+uint64_t millis_64();
+
+}  // namespace esphome
+
+#endif  // USE_HOST

esphome/core/hal/hal_libretiny.h

@@ -0,0 +1,93 @@
+#pragma once
+
+#ifdef USE_LIBRETINY
+
+#include <cstdint>
+
+// For the inline millis() fast paths (xTaskGetTickCount, portTICK_PERIOD_MS).
+#include <FreeRTOS.h>
+#include <task.h>
+
+#include "esphome/core/time_64.h"
+
+// IRAM_ATTR places a function in executable RAM so it is callable from an
+// ISR even while flash is busy (XIP stall, OTA, logger flash write).
+// Each family uses a section its stock linker already routes to RAM:
+// RTL8710B → .image2.ram.text, RTL8720C → .sram.text. LN882H is the
+// exception: its stock linker has no matching glob, so patch_linker.py
+// injects KEEP(*(.sram.text*)) into .flash_copysection at pre-link.
+//
+// BK72xx (all variants) are left as a no-op: their SDK wraps flash
+// operations in GLOBAL_INT_DISABLE() which masks FIQ + IRQ at the CPU for
+// the duration of every write, so no ISR fires while flash is stalled and
+// the race IRAM_ATTR guards against cannot occur. The trade-off is that
+// interrupts are delayed (not dropped) by up to ~20 ms during a sector
+// erase, but that is an SDK-level choice and cannot be changed from this
+// layer.
+#if defined(USE_BK72XX)
+#define IRAM_ATTR
+#elif defined(USE_LIBRETINY_VARIANT_RTL8710B)
+// Stock linker consumes *(.image2.ram.text*) into .ram_image2.text (> BD_RAM).
+#define IRAM_ATTR __attribute__((noinline, section(".image2.ram.text")))
+#else
+// RTL8720C: stock linker consumes *(.sram.text*) into .ram.code_text.
+// LN882H: patch_linker.py.script injects *(.sram.text*) into
+// .flash_copysection (> RAM0 AT> FLASH).
+#define IRAM_ATTR __attribute__((noinline, section(".sram.text")))
+#endif
+#define PROGMEM
+
+#ifdef USE_BK72XX
+// Declared in the Beken FreeRTOS port (portmacro.h) and built in ARM mode so
+// it is callable from Thumb code via interworking. The MRS CPSR instruction
+// is ARM-only and user code here may be built in Thumb, so in_isr_context()
+// defers to this port helper on BK72xx instead of reading CPSR inline.
+extern "C" uint32_t platform_is_in_interrupt_context(void);
+#endif
+
+// Forward decls from Arduino's <Arduino.h> for the inline wrappers below.
+// NOLINTBEGIN(google-runtime-int,readability-identifier-naming,readability-redundant-declaration)
+extern "C" void yield(void);
+extern "C" void delay(unsigned long ms);
+extern "C" unsigned long micros(void);
+extern "C" unsigned long millis(void);
+// NOLINTEND(google-runtime-int,readability-identifier-naming,readability-redundant-declaration)
+
+namespace esphome {
+
+/// Returns true when executing inside an interrupt handler.
+__attribute__((always_inline)) inline bool in_isr_context() {
+#if defined(USE_BK72XX)
+  // BK72xx is ARM968E-S (ARM9); see extern declaration above.
+  return platform_is_in_interrupt_context() != 0;
+#else
+  // Cortex-M (AmebaZ, AmebaZ2, LN882H). IPSR is the active exception number;
+  // non-zero means we're in a handler.
+  uint32_t ipsr;
+  __asm__ volatile("mrs %0, ipsr" : "=r"(ipsr));
+  return ipsr != 0;
+#endif
+}
+
+__attribute__((always_inline)) inline void yield() { ::yield(); }
+__attribute__((always_inline)) inline void delay(uint32_t ms) { ::delay(ms); }
+__attribute__((always_inline)) inline uint32_t micros() { return static_cast<uint32_t>(::micros()); }
+
+// Per-variant millis() fast path — matches MillisInternal::get().
+#if defined(USE_RTL87XX) || defined(USE_LN882X)
+static_assert(configTICK_RATE_HZ == 1000, "millis() fast path requires 1 kHz FreeRTOS tick");
+__attribute__((always_inline)) inline uint32_t millis() {
+  // xTaskGetTickCountFromISR is mandatory in interrupt context per the FreeRTOS API contract.
+  return in_isr_context() ? xTaskGetTickCountFromISR() : xTaskGetTickCount();
+}
+#elif defined(USE_BK72XX)
+static_assert(configTICK_RATE_HZ == 500, "BK72xx millis() fast path assumes 500 Hz FreeRTOS tick");
+__attribute__((always_inline)) inline uint32_t millis() { return xTaskGetTickCount() * portTICK_PERIOD_MS; }
+#else
+__attribute__((always_inline)) inline uint32_t millis() { return static_cast<uint32_t>(::millis()); }
+#endif
+__attribute__((always_inline)) inline uint64_t millis_64() { return Millis64Impl::compute(millis()); }
+
+}  // namespace esphome
+
+#endif  // USE_LIBRETINY

esphome/core/hal/hal_rp2040.h

@@ -0,0 +1,40 @@
+#pragma once
+
+#ifdef USE_RP2040
+
+#include <cstdint>
+
+#include "esphome/core/time_conversion.h"
+
+#define IRAM_ATTR __attribute__((noinline, long_call, section(".time_critical")))
+#define PROGMEM
+
+// Forward decls from Arduino's <Arduino.h> for the inline wrappers below.
+// NOLINTBEGIN(google-runtime-int,readability-identifier-naming,readability-redundant-declaration)
+extern "C" void yield(void);
+extern "C" void delay(unsigned long ms);
+extern "C" unsigned long micros(void);
+extern "C" unsigned long millis(void);
+// NOLINTEND(google-runtime-int,readability-identifier-naming,readability-redundant-declaration)
+
+// Forward decl from <pico/time.h>.
+extern "C" uint64_t time_us_64(void);
+
+namespace esphome {
+
+/// Returns true when executing inside an interrupt handler.
+__attribute__((always_inline)) inline bool in_isr_context() {
+  uint32_t ipsr;
+  __asm__ volatile("mrs %0, ipsr" : "=r"(ipsr));
+  return ipsr != 0;
+}
+
+__attribute__((always_inline)) inline void yield() { ::yield(); }
+__attribute__((always_inline)) inline void delay(uint32_t ms) { ::delay(ms); }
+__attribute__((always_inline)) inline uint32_t micros() { return static_cast<uint32_t>(::micros()); }
+__attribute__((always_inline)) inline uint32_t millis() { return micros_to_millis(::time_us_64()); }
+__attribute__((always_inline)) inline uint64_t millis_64() { return micros_to_millis<uint64_t>(::time_us_64()); }
+
+}  // namespace esphome
+
+#endif  // USE_RP2040

esphome/core/hal/hal_zephyr.h

@@ -0,0 +1,24 @@
+#pragma once
+
+#ifdef USE_ZEPHYR
+
+#include <cstdint>
+
+#define IRAM_ATTR
+#define PROGMEM
+
+namespace esphome {
+
+/// Returns true when executing inside an interrupt handler.
+/// Zephyr/nRF52: not currently consulted — wake path is platform-specific.
+__attribute__((always_inline)) inline bool in_isr_context() { return false; }
+
+void yield();
+void delay(uint32_t ms);
+uint32_t micros();
+uint32_t millis();
+uint64_t millis_64();
+
+}  // namespace esphome
+
+#endif  // USE_ZEPHYR