esp32: add interrupt support (vector table, timer alarm, GPIO SetInterrupt)

src/device/esp/esp32.S

@@ -10,17 +10,64 @@
 .section .text.call_start_cpu0
 1:
     .long _stack_top
+.Lmain_addr:
+    .long main
+.Lrom_mmu_init:
+    .long 0x400095a4         // mmu_init(int cpu_no)
+.Lrom_cache_flash_mmu_set:
+    .long 0x400095e0         // cache_flash_mmu_set(cpu, pid, vaddr, paddr, pgsz, pgcnt)
+.Lrom_Cache_Read_Enable:
+    .long 0x40009a84         // Cache_Read_Enable(int cpu_no)
+.Lrom_Cache_Read_Disable:
+    .long 0x40009ab8         // Cache_Read_Disable(int cpu_no)
+.Lrom_Cache_Flush:
+    .long 0x40009a14         // Cache_Flush(int cpu_no)
+.Lrodata_start:
+    .long _rodata_start
+.Lrodata_end:
+    .long _rodata_end
+.Ltext_start:
+    .long _text_start
+.Ltext_end:
+    .long _text_end
+.Ldport_pro_cache_ctrl1:
+    .long 0x3FF00044         // DPORT_PRO_CACHE_CTRL1_REG
+.Ldrom_paddr_ptr:
+    .long _drom_flash_addr   // pointer to builder-patched DROM flash offset
+.Ldrom_vaddr:
+    .long 0x3F400000         // DROM virtual base address
+.Lirom_vaddr:
+    .long 0x400D0000         // IROM virtual base address
+.Lmmu_table_base:
+    .long 0x3FF10000         // PRO CPU Flash MMU table
+.Lrtc_wdt_protect:
+    .long 0x3FF480A4         // RTC_CNTL_WDTWPROTECT_REG
+.Lrtc_wdt_key:
+    .long 0x50D83AA1         // WDT write-protect key
+.Lrtc_wdt_config0:
+    .long 0x3FF4808C         // RTC_CNTL_WDTCONFIG0_REG
+.Ltimg0_wdt_protect:
+    .long 0x3FF5F064         // TIMG0_WDTWPROTECT_REG
+.Ltimg0_wdt_config0:
+    .long 0x3FF5F048         // TIMG0_WDTCONFIG0_REG
+.Lvector_table:
+    .long _vector_table
+
 .global call_start_cpu0
 call_start_cpu0:
     // We need to set the stack pointer to a different value. This is somewhat
     // complicated in the Xtensa architecture. The code below is a modified
     // version of the following code:
     // https://github.com/espressif/esp-idf/blob/c77c4ccf/components/xtensa/include/xt_instr_macros.h#L47
 
-    // Disable WOE.
+    // Disable WOE (bit 18 of PS).
+    // Avoid large movi constants to prevent auto-generated .literal section
+    // entries, which cause l32r offset miscalculation in LLVM 22 / lld.
     rsr.ps a2
-    movi a3, ~(PS_WOE_MASK)
-    and a2, a2, a3
+    movi a3, 1
+    slli a3, a3, 18         // a3 = PS_WOE_MASK (0x40000)
+    and  a3, a2, a3         // a3 = a2 & WOE_MASK (isolate WOE bit)
+    xor  a2, a2, a3         // clear WOE bit
     wsr.ps a2
     rsync
 
@@ -37,7 +84,8 @@ call_start_cpu0:
 
     // Re-enable WOE.
     rsr.ps a2
-    movi a3, PS_WOE
+    movi a3, 1
+    slli a3, a3, 18         // a3 = PS_WOE (0x40000)
     or a2, a2, a3
     wsr.ps a2
     rsync
@@ -47,11 +95,187 @@ call_start_cpu0:
     wsr.cpenable a2
     rsync
 
-    // Jump to the runtime start function written in Go.
-    call4 main
+    // Disable the RTC and TIMG0 watchdogs before configuring the flash cache.
+    // The ROM bootloader leaves them running; a fault during cache setup would
+    // otherwise reset the chip. The Go runtime re-disables them once it starts.
+    l32r a2, .Lrtc_wdt_protect
+    l32r a3, .Lrtc_wdt_key
+    s32i a3, a2, 0          // unlock WDT write-protect
+    memw
+    l32r a2, .Lrtc_wdt_config0
+    movi a3, 0
+    s32i a3, a2, 0          // disable WDT (write 0 to config0)
+    memw
+
+    // Disable TG0 WDT (Timer Group 0 Main Watchdog).
+    // TIMG0_WDTWPROTECT_REG = 0x3FF5F064, TIMG0_WDTCONFIG0_REG = 0x3FF5F048
+    l32r a2, .Ltimg0_wdt_protect
+    l32r a3, .Lrtc_wdt_key  // same unlock key 0x50D83AA1
+    s32i a3, a2, 0
+    memw
+    l32r a2, .Ltimg0_wdt_config0
+    movi a3, 0
+    s32i a3, a2, 0
+    memw
+
+    // Set VECBASE to our vector table. Must happen before any callx4 so that
+    // register-window overflow exceptions route to our handlers.
+    l32r a2, .Lvector_table
+    wsr.vecbase a2
+    rsync
+
+    // Clear PS.EXCM so window overflow exceptions work properly.
+    rsr.ps a2
+    movi a3, ~0x1F
+    and a2, a2, a3
+    movi a3, 0x20          // PS.UM = 1
+    or a2, a2, a3
+    wsr.ps a2
+    rsync
+
+    // ---- Configure flash cache and MMU ----
+    movi a6, 0
+    mov  a5, a1
+    l32r a4, .Lrom_Cache_Read_Disable
+    callx4 a4
+
+    movi a6, 0
+    mov  a5, a1
+    l32r a4, .Lrom_Cache_Flush
+    callx4 a4
+
+    movi a6, 0
+    mov  a5, a1
+    l32r a4, .Lrom_mmu_init
+    callx4 a4
+
+    l32r a2, .Lrodata_end
+    l32r a3, .Lrodata_start
+    sub  a2, a2, a3
+    beqz a2, .Lskip_drom
+    addi a2, a2, -1
+    srli a2, a2, 16
+    addi a2, a2, 1
+
+    movi a6, 0
+    movi a7, 0
+    l32r a8, .Ldrom_vaddr
+    l32r a9, .Ldrom_paddr_ptr
+    l32i a9, a9, 0
+    movi a10, 64
+    mov  a11, a2
+    mov  a5, a1
+    l32r a4, .Lrom_cache_flash_mmu_set
+    callx4 a4
+.Lskip_drom:
 
+    l32r a2, .Ltext_end
+    l32r a3, .Ltext_start
+    sub  a2, a2, a3
+    beqz a2, .Lskip_irom
+    addi a2, a2, -1
+    srli a2, a2, 16
+    addi a2, a2, 1
+
+    l32r a9, .Lrodata_end
+    l32r a3, .Lrodata_start
+    sub  a9, a9, a3
+    l32r a3, .Ldrom_paddr_ptr
+    l32i a3, a3, 0
+    beqz a9, .Lirom_paddr_ready
+    addi a9, a9, -1
+    srli a9, a9, 16
+    addi a9, a9, 1
+    slli a9, a9, 16
+    add  a3, a3, a9
+.Lirom_paddr_ready:
+
+    movi a6, 0
+    movi a7, 0
+    l32r a8, .Lirom_vaddr
+    mov  a9, a3
+    movi a10, 64
+    mov  a11, a2
+    mov  a5, a1
+    l32r a4, .Lrom_cache_flash_mmu_set
+    callx4 a4
+.Lskip_irom:
+
+    l32r a2, .Ldport_pro_cache_ctrl1
+    l32i a3, a2, 0
+    movi a4, ~0x11
+    and  a3, a3, a4
+    s32i a3, a2, 0
+    memw
+
+    movi a6, 0
+    mov  a5, a1
+    l32r a4, .Lrom_Cache_Read_Enable
+    callx4 a4
+
+    isync
+
+    // ---- Jump to main (in IROM/flash, now accessible) ----
+    mov  a5, a1
+    l32r a4, .Lmain_addr
+    callx4 a4
+
+    // If main returns, loop forever.
+1:  j 1b
+
+// -----------------------------------------------------------------------
+// tinygo_scanCurrentStack — Spill all Xtensa register windows to the
+// stack, then call tinygo_scanstack(sp) so the conservative GC can
+// discover live heap pointers that are currently in physical registers.
+//
+// On RISC-V / ARM the equivalent function pushes callee-saved registers
+// before the call.  On Xtensa windowed ABI the same effect is achieved
+// by forcing hardware window-overflow for every occupied pane: each
+// overflow saves the four registers in that pane to the stack frame
+// pointed to by the pane's a1 (sp).  After all panes are flushed, a
+// scan from the current sp to stackTop covers every live value.
+//
+// Without this spill the conservative GC misses heap pointers held only
+// in physical registers, frees live objects, and later crashes jumping
+// through a freed/garbage function pointer (e.g. a goroutine trampoline).
+// -----------------------------------------------------------------------
 .section .text.tinygo_scanCurrentStack
+
 .global tinygo_scanCurrentStack
 tinygo_scanCurrentStack:
-    // TODO: save callee saved registers on the stack
-    j tinygo_scanstack
+    entry a1, 48
+
+    // Disable interrupts while flushing register windows.
+    rsr   a4, PS
+    s32i  a4, a1, 0         // save PS for later restore
+    rsil  a4, 3             // XCHAL_EXCM_LEVEL
+
+    // Flush all register windows using recursive call4.
+    // For NAREG=64 (16 panes), 15 recursive levels cover all panes
+    // except the current one (which is kept active).
+    movi a6, 15
+    call4 .Lscan_spill
+
+    // Restore interrupts.
+    l32i  a4, a1, 0
+    wsr.ps a4
+    rsync
+
+    // Pass current sp to tinygo_scanstack.
+    // call4 maps caller's a5→callee's a1 (stack ptr for callee's entry)
+    // and caller's a6→callee's a2 (first argument = sp).
+    mov  a5, a1             // callee's a1 = valid stack pointer
+    mov  a6, a1             // callee's a2 = sp argument
+    call4 tinygo_scanstack
+
+    retw
+
+    .balign 4
+.Lscan_spill:
+    entry a1, 16
+    beqz  a2, .Lscan_spill_done
+    addi  a2, a2, -1
+    mov   a6, a2
+    call4 .Lscan_spill
+.Lscan_spill_done:
+    retw