diff --git a/pio/spi/pio_spi.h b/pio/spi/pio_spi.h
index dfa929d8b..f2ce42d8a 100644
--- a/pio/spi/pio_spi.h
+++ b/pio/spi/pio_spi.h
@@ -13,6 +13,7 @@ typedef struct pio_spi_inst {
     PIO pio;
     uint sm;
     uint cs_pin;
+    uint offset;
 } pio_spi_inst_t;
 
 void pio_spi_write8_blocking(const pio_spi_inst_t *spi, const uint8_t *src, size_t len);
diff --git a/pio/spi/spi.pio b/pio/spi/spi.pio
index 4033a74ab..be3e9eee7 100644
--- a/pio/spi/spi.pio
+++ b/pio/spi/spi.pio
@@ -55,8 +55,8 @@ static inline void pio_spi_init(PIO pio, uint sm, uint prog_offs, uint n_bits,
     sm_config_set_clkdiv(&c, clkdiv);
 
     // MOSI, SCK output are low, MISO is input
-    pio_sm_set_pins_with_mask(pio, sm, 0, (1u << pin_sck) | (1u << pin_mosi));
-    pio_sm_set_pindirs_with_mask(pio, sm, (1u << pin_sck) | (1u << pin_mosi), (1u << pin_sck) | (1u << pin_mosi) | (1u << pin_miso));
+    pio_sm_set_pins_with_mask64(pio, sm, 0, (1ull << pin_sck) | (1ull << pin_mosi));
+    pio_sm_set_pindirs_with_mask64(pio, sm, (1ull << pin_sck) | (1ull << pin_mosi), (1ull << pin_sck) | (1ull << pin_mosi) | (1ull << pin_miso));
     pio_gpio_init(pio, pin_mosi);
     pio_gpio_init(pio, pin_miso);
     pio_gpio_init(pio, pin_sck);
@@ -65,7 +65,7 @@ static inline void pio_spi_init(PIO pio, uint sm, uint prog_offs, uint n_bits,
     // and this is a cheesy way to get CPOL=1
     gpio_set_outover(pin_sck, cpol ? GPIO_OVERRIDE_INVERT : GPIO_OVERRIDE_NORMAL);
     // SPI is synchronous, so bypass input synchroniser to reduce input delay.
-    hw_set_bits(&pio->input_sync_bypass, 1u << pin_miso);
+    hw_set_bits(&pio->input_sync_bypass, (pin_miso >= 32) ? (1u << (pin_miso - 16)) : (1u << pin_miso));
 
     pio_sm_init(pio, sm, prog_offs, &c);
     pio_sm_set_enabled(pio, sm, true);
@@ -150,14 +150,14 @@ static inline void pio_spi_cs_init(PIO pio, uint sm, uint prog_offs, uint n_bits
     sm_config_set_in_shift(&c, false, true, n_bits);
     sm_config_set_clkdiv(&c, clkdiv);
 
-    pio_sm_set_pins_with_mask(pio, sm, (2u << pin_sck), (3u << pin_sck) | (1u << pin_mosi));
-    pio_sm_set_pindirs_with_mask(pio, sm, (3u << pin_sck) | (1u << pin_mosi), (3u << pin_sck) | (1u << pin_mosi) | (1u << pin_miso));
+    pio_sm_set_pins_with_mask64(pio, sm, (2ull << pin_sck), (3ull << pin_sck) | (1ull << pin_mosi));
+    pio_sm_set_pindirs_with_mask64(pio, sm, (3ull << pin_sck) | (1ull << pin_mosi), (3ull << pin_sck) | (1ull << pin_mosi) | (1ull << pin_miso));
     pio_gpio_init(pio, pin_mosi);
     pio_gpio_init(pio, pin_miso);
     pio_gpio_init(pio, pin_sck);
     pio_gpio_init(pio, pin_sck + 1);
     gpio_set_outover(pin_sck, cpol ? GPIO_OVERRIDE_INVERT : GPIO_OVERRIDE_NORMAL);
-    hw_set_bits(&pio->input_sync_bypass, 1u << pin_miso);
+    hw_set_bits(&pio->input_sync_bypass, (pin_miso >= 32) ? (1u << (pin_miso - 16)) : (1u << pin_miso));
 
     uint entry_point = prog_offs + (cpha ? spi_cpha1_cs_offset_entry_point : spi_cpha0_cs_offset_entry_point);
     pio_sm_init(pio, sm, entry_point, &c);
diff --git a/pio/spi/spi_loopback.c b/pio/spi/spi_loopback.c
index ac5897f9c..014fa9243 100644
--- a/pio/spi/spi_loopback.c
+++ b/pio/spi/spi_loopback.c
@@ -49,19 +49,20 @@ void test(const pio_spi_inst_t *spi) {
 int main() {
     stdio_init_all();
 
-    pio_spi_inst_t spi = {
-            .pio = pio0,
-            .sm = 0
-    };
+    pio_spi_inst_t spi[2];
     float clkdiv = 31.25f;  // 1 MHz @ 125 clk_sys
-    uint cpha0_prog_offs = pio_add_program(spi.pio, &spi_cpha0_program);
-    uint cpha1_prog_offs = pio_add_program(spi.pio, &spi_cpha1_program);
+
+    // pio_claim_free_sm_and_add_program_for_gpio_range finds a free pio and state machine for a program and sets the gpio base correctly
+    const uint pin_base = MIN(PIN_SCK, MIN(PIN_MOSI, PIN_MISO));
+    const uint pin_count = MAX(PIN_SCK, MAX(PIN_MOSI, PIN_MISO)) - pin_base + 1;
+    hard_assert(pio_claim_free_sm_and_add_program_for_gpio_range(&spi_cpha0_program, &spi[0].pio, &spi[0].sm, &spi[0].offset, pin_base, pin_count, true));
+    hard_assert(pio_claim_free_sm_and_add_program_for_gpio_range(&spi_cpha1_program, &spi[1].pio, &spi[1].sm, &spi[1].offset, pin_base, pin_count, true));
 
     for (int cpha = 0; cpha <= 1; ++cpha) {
         for (int cpol = 0; cpol <= 1; ++cpol) {
             printf("CPHA = %d, CPOL = %d\n", cpha, cpol);
-            pio_spi_init(spi.pio, spi.sm,
-                         cpha ? cpha1_prog_offs : cpha0_prog_offs,
+            pio_spi_init(spi[cpha].pio, spi[cpha].sm,
+                         spi[cpha].offset,
                          8,       // 8 bits per SPI frame
                          clkdiv,
                          cpha,
@@ -70,7 +71,7 @@ int main() {
                          PIN_MOSI,
                          PIN_MISO
             );
-            test(&spi);
+            test(&spi[cpha]);
             sleep_ms(10);
         }
     }