Add Risc-V crt0 data_cpy skip

Also shrink Arm one, and standardise PICO_NO_FLASH skipped code between them

src/rp2_common/pico_crt0/crt0.S

@@ -554,29 +554,24 @@ data_cpy_check:
     // First check start of copy (r2)
     mov r7, r2
 check_r7:
-    ldr r5, =(SRAM_BASE >> 16)
-    cmp r5, r7, lsr #16
+    ldr r5, =SRAM_BASE
+    cmp r5, r7
     bgt check_xip_sram
-    ldr r5, =(SRAM4_BASE >> 16)
-    cmp r5, r7, lsr #16
+    ldr r5, =SRAM4_BASE
+    cmp r5, r7
     bgt check_sram0
-    b check_sram1
-check_xip_sram:
-    mov r5, #(1 << 2) // POWMAN_POWER_DOMAIN_XIP_CACHE
-    and r5, r6
-    cmp r5, #(1 << 2)
-    beq skip
-    b noskip
-check_sram0:
-    mov r5, #(1 << 1) // POWMAN_POWER_DOMAIN_SRAM_BANK0
-    and r5, r6
-    cmp r5, #(1 << 1)
-    beq skip
-    b noskip
 check_sram1:
     mov r5, #(1 << 0) // POWMAN_POWER_DOMAIN_SRAM_BANK1
-    and r5, r6
-    cmp r5, #(1 << 0)
+    b do_check
+check_xip_sram:
+    mov r5, #(1 << 2) // POWMAN_POWER_DOMAIN_XIP_CACHE
+    b do_check
+check_sram0:
+    mov r5, #(1 << 1) // POWMAN_POWER_DOMAIN_SRAM_BANK0
+    b do_check
+do_check:
+    and r7, r5, r6
+    cmp r5, r7
     beq skip
     b noskip
 skip:

src/rp2_common/pico_crt0/crt0_riscv.S

@@ -12,6 +12,10 @@
 #include "boot/picobin.h"
 #include "pico/bootrom_constants.h"
 
+#if !PICO_RP2040
+#include "hardware/regs/powman.h"
+#endif
+
 #ifdef NDEBUG
 #ifndef COLLAPSE_IRQS
 #define COLLAPSE_IRQS
@@ -341,6 +345,11 @@ _call_xip_setup:
     // In a NO_FLASH binary, don't perform .data etc copy, since it's loaded
     // in-place by the SRAM load. Still need to clear .bss
 #if !PICO_NO_FLASH
+#if LIB_PICO_LOW_POWER && !PICO_RP2040
+    // Load previous power state into r6
+    li a6, POWMAN_BASE+POWMAN_SCRATCH6_OFFSET
+    lw a6, 0(a6)
+#endif
     la a4, data_cpy_table
 
     // assume there is at least one entry
@@ -350,6 +359,10 @@ _call_xip_setup:
     lw a2, 4(a4)
     lw a3, 8(a4)
     addi a4, a4, 12
+#if LIB_PICO_LOW_POWER && !PICO_RP2040
+    jal data_cpy_check
+    bnez a0, 1b
+#endif
     jal data_cpy
     j 1b
 2:
@@ -379,6 +392,8 @@ platform_entry: // symbol for stack traces
     ebreak
     j 1b
 
+
+#if !PICO_NO_FLASH
 data_cpy_loop:
     lw a0, (a1)
     sw a0, (a2)
@@ -388,6 +403,45 @@ data_cpy:
     bltu a2, a3, data_cpy_loop
     ret
 
+#if LIB_PICO_LOW_POWER && !PICO_RP2040
+data_cpy_check:
+    // uses a5 and a7 as scratch
+    // uses the powman_power_state in a6
+    // returns in a0
+    li a0, 0
+    // First check start of copy (a2)
+    mv a7, a2
+check_a7:
+    li a5, SRAM_BASE
+    bgt a5, a7, check_xip_sram
+    li a5, SRAM4_BASE
+    bgt a5, a7, check_sram0
+check_sram1:
+    li a5, 1 << 0 // POWMAN_POWER_DOMAIN_SRAM_BANK1
+    j do_check
+check_xip_sram:
+    li a5, 1 << 2 // POWMAN_POWER_DOMAIN_XIP_CACHE
+    j do_check
+check_sram0:
+    li a5, 1 << 1 // POWMAN_POWER_DOMAIN_SRAM_BANK0
+    j do_check
+do_check:
+    and a7, a5, a6
+    beq a5, a7, skip
+    j noskip
+skip:
+    beqz a0, check_end
+    ret
+check_end:
+    addi a0, a0, 1
+    mv a7, a3
+    j check_a7
+noskip:
+    li a0, 0
+    ret
+#endif
+#endif
+
 .align 2
 data_cpy_table:
 #if PICO_RP2350 && PICO_EMBED_XIP_SETUP && !PICO_NO_FLASH