chore: checkpoint before Python removal

vendor/psm/src/arch/aarch64_armasm.asm

@@ -0,0 +1,38 @@
+    AREA |.text|, CODE, READONLY
+
+    GLOBAL |rust_psm_stack_direction|
+    ALIGN 4
+|rust_psm_stack_direction| PROC
+    orr w0, wzr, #2
+    ret
+    ENDP
+
+
+    GLOBAL |rust_psm_stack_pointer|
+    ALIGN 4
+|rust_psm_stack_pointer| PROC
+    mov x0, sp
+    ret
+    ENDP
+
+
+    GLOBAL |rust_psm_replace_stack|
+    ALIGN 4
+|rust_psm_replace_stack| PROC
+    mov sp, x2
+    br x1
+    ENDP
+
+    GLOBAL |rust_psm_on_stack|
+    ALIGN 4
+|rust_psm_on_stack| PROC
+    stp x29, x30, [sp, #-16]!
+    mov x29, sp
+    mov sp, x3
+    blr x2
+    mov sp, x29
+    ldp x29, x30, [sp], #16
+    ret
+    ENDP
+
+    END

vendor/psm/src/arch/aarch_aapcs64.s

@@ -0,0 +1,89 @@
+#include "psm.h"
+#include "gnu_stack_note.s"
+
+.text
+
+#if defined(CFG_TARGET_OS_darwin) || defined(CFG_TARGET_OS_macos) || defined(CFG_TARGET_OS_ios)
+
+#define GLOBL(fnname) .globl _##fnname
+#define TYPE(fnname)
+#define FUNCTION(fnname) _##fnname
+#define END_FUNCTION(fnname)
+
+#elif defined(CFG_TARGET_OS_windows)
+
+#define GLOBL(fnname) .globl fnname
+#define TYPE(fnname)
+#define FUNCTION(fnname) fnname
+#define END_FUNCTION(fnname)
+
+#else
+
+#define GLOBL(fnname) .globl fnname
+#define TYPE(fnname) .type fnname,@function
+#define FUNCTION(fnname) fnname
+#define END_FUNCTION(fnname) .size fnname,.-fnname
+
+#endif
+
+
+GLOBL(rust_psm_stack_direction)
+.p2align 2
+TYPE(rust_psm_stack_direction)
+FUNCTION(rust_psm_stack_direction):
+/* extern "C" fn() -> u8 */
+.cfi_startproc
+    orr w0, wzr, #STACK_DIRECTION_DESCENDING
+    ret
+END_FUNCTION(rust_psm_stack_direction)
+.cfi_endproc
+
+
+GLOBL(rust_psm_stack_pointer)
+.p2align 2
+TYPE(rust_psm_stack_pointer)
+FUNCTION(rust_psm_stack_pointer):
+/* extern "C" fn() -> *mut u8 */
+.cfi_startproc
+    mov x0, sp
+    ret
+END_FUNCTION(rust_psm_stack_pointer)
+.cfi_endproc
+
+
+GLOBL(rust_psm_replace_stack)
+.p2align 2
+TYPE(rust_psm_replace_stack)
+FUNCTION(rust_psm_replace_stack):
+/* extern "C" fn(r0: usize, r1: extern "C" fn(usize), r2: *mut u8) */
+.cfi_startproc
+/* All we gotta do is set the stack pointer to %rdx & tail-call the callback in %rsi */
+    mov sp, x2
+    br x1
+END_FUNCTION(rust_psm_replace_stack)
+.cfi_endproc
+
+
+GLOBL(rust_psm_on_stack)
+.p2align 2
+TYPE(rust_psm_on_stack)
+FUNCTION(rust_psm_on_stack):
+/* extern "C" fn(r0: usize, r1: usize, r2: extern "C" fn(usize, usize), r3: *mut u8) */
+.cfi_startproc
+    stp x29, x30, [sp, #-16]!
+    .cfi_def_cfa sp, 16
+    mov x29, sp
+    .cfi_def_cfa x29, 16
+    .cfi_offset x29, -16
+    .cfi_offset x30, -8
+    mov sp, x3
+    blr x2
+    mov sp, x29
+    .cfi_def_cfa sp, 16
+    ldp x29, x30, [sp], #16
+    .cfi_def_cfa sp, 0
+    .cfi_restore x29
+    .cfi_restore x30
+    ret
+END_FUNCTION(rust_psm_on_stack)
+.cfi_endproc

vendor/psm/src/arch/arm64ec_armasm.asm

@@ -0,0 +1,38 @@
+    AREA |.text|, CODE, READONLY
+
+    GLOBAL |#rust_psm_stack_direction|
+    ALIGN 4
+|#rust_psm_stack_direction| PROC
+    orr w0, wzr, #2
+    ret
+    ENDP
+
+
+    GLOBAL |#rust_psm_stack_pointer|
+    ALIGN 4
+|#rust_psm_stack_pointer| PROC
+    mov x0, sp
+    ret
+    ENDP
+
+
+    GLOBAL |#rust_psm_replace_stack|
+    ALIGN 4
+|#rust_psm_replace_stack| PROC
+    mov sp, x2
+    br x1
+    ENDP
+
+    GLOBAL |#rust_psm_on_stack|
+    ALIGN 4
+|#rust_psm_on_stack| PROC
+    stp x29, x30, [sp, #-16]!
+    mov x29, sp
+    mov sp, x3
+    blr x2
+    mov sp, x29
+    ldp x29, x30, [sp], #16
+    ret
+    ENDP
+
+    END

vendor/psm/src/arch/arm_aapcs.s

@@ -0,0 +1,107 @@
+#include "psm.h"
+#include "gnu_stack_note.s"
+
+.text
+.syntax unified
+
+#if defined(CFG_TARGET_OS_darwin) || defined(CFG_TARGET_OS_macos) || defined(CFG_TARGET_OS_ios)
+
+#define GLOBL(fnname) .globl _##fnname
+#define THUMBTYPE(fnname) .thumb_func _##fnname
+#define FUNCTION(fnname) _##fnname
+#define THUMBFN .code 16
+#define SIZE(fnname,endlabel)
+#define FNSTART
+#define CANTUNWIND
+#define FNEND
+
+#else
+
+#define GLOBL(fnname) .globl fnname
+#define THUMBTYPE(fnname) .type fnname,%function
+#define FUNCTION(fnname) fnname
+#define THUMBFN .code 16
+#define SIZE(fnname,endlabel) .size fnname,endlabel-fnname
+#define FNSTART .fnstart
+#define CANTUNWIND .cantunwind
+#define FNEND .fnend
+
+#endif
+
+
+GLOBL(rust_psm_stack_direction)
+.p2align 2
+THUMBTYPE(rust_psm_stack_direction)
+THUMBFN
+FUNCTION(rust_psm_stack_direction):
+/* extern "C" fn() -> u8 */
+FNSTART
+.cfi_startproc
+    /* movs to support Thumb-1 */
+    movs r0, #STACK_DIRECTION_DESCENDING
+    bx lr
+.rust_psm_stack_direction_end:
+SIZE(rust_psm_stack_direction,.rust_psm_stack_direction_end)
+.cfi_endproc
+CANTUNWIND
+FNEND
+
+GLOBL(rust_psm_stack_pointer)
+.p2align 2
+THUMBTYPE(rust_psm_stack_pointer)
+THUMBFN
+FUNCTION(rust_psm_stack_pointer):
+/* extern "C" fn() -> *mut u8 */
+FNSTART
+.cfi_startproc
+    mov r0, sp
+    bx  lr
+.rust_psm_stack_pointer_end:
+SIZE(rust_psm_stack_pointer,.rust_psm_stack_pointer_end)
+.cfi_endproc
+CANTUNWIND
+FNEND
+
+
+GLOBL(rust_psm_replace_stack)
+.p2align 2
+THUMBTYPE(rust_psm_replace_stack)
+THUMBFN
+FUNCTION(rust_psm_replace_stack):
+/* extern "C" fn(r0: usize, r1: extern "C" fn(usize), r2: *mut u8) */
+FNSTART
+.cfi_startproc
+/* All we gotta do is set the stack pointer to %rdx & tail-call the callback in %rsi */
+    mov sp, r2
+    bx r1
+.rust_psm_replace_stack_end:
+SIZE(rust_psm_replace_stack,.rust_psm_replace_stack_end)
+.cfi_endproc
+CANTUNWIND
+FNEND
+
+
+GLOBL(rust_psm_on_stack)
+.p2align 2
+THUMBTYPE(rust_psm_on_stack)
+THUMBFN
+FUNCTION(rust_psm_on_stack):
+/* extern "C" fn(r0: usize, r1: usize, r2: extern "C" fn(usize, usize), r3: *mut u8) */
+FNSTART
+.cfi_startproc
+    push {r4, lr}
+    .cfi_def_cfa_offset 8
+    mov r4, sp
+    .cfi_def_cfa_register r4
+    .cfi_offset lr, -4
+    .cfi_offset r4, -8
+    mov sp, r3
+    blx r2
+    mov sp, r4
+    .cfi_restore sp
+    pop {r4, pc}
+.rust_psm_on_stack_end:
+SIZE(rust_psm_on_stack,.rust_psm_on_stack_end)
+.cfi_endproc
+CANTUNWIND
+FNEND

vendor/psm/src/arch/arm_armasm.asm

@@ -0,0 +1,39 @@
+    THUMB
+    AREA |.text|, CODE, READONLY
+
+
+    GLOBAL |rust_psm_stack_direction|
+    ALIGN 4
+|rust_psm_stack_direction| PROC
+    movs r0, #2
+    bx lr
+    ENDP
+
+    GLOBAL |rust_psm_stack_pointer|
+    ALIGN 4
+|rust_psm_stack_pointer| PROC
+    mov r0, sp
+    bx  lr
+    ENDP
+
+
+    GLOBAL |rust_psm_replace_stack|
+    ALIGN 4
+|rust_psm_replace_stack| PROC
+    mov sp, r2
+    bx r1
+    ENDP
+
+
+    GLOBAL |rust_psm_on_stack|
+    ALIGN 4
+|rust_psm_on_stack| PROC
+    push {r4, lr}
+    mov r4, sp
+    mov sp, r3
+    blx r2
+    mov sp, r4
+    pop {r4, pc}
+    ENDP
+
+    END

vendor/psm/src/arch/gnu_stack_note.s

@@ -0,0 +1,3 @@
+#if defined(__linux__) && defined(__ELF__)
+.section .note.GNU-stack,"",%progbits
+#endif

vendor/psm/src/arch/loongarch64.s

@@ -0,0 +1,64 @@
+#include "psm.h"
+#include "gnu_stack_note.s"
+
+.text
+.globl rust_psm_stack_direction
+.align 2
+.type rust_psm_stack_direction,@function
+rust_psm_stack_direction:
+/* extern "C" fn() -> u8 */
+.cfi_startproc
+    li.w $r4, STACK_DIRECTION_DESCENDING
+    jr $r1
+.rust_psm_stack_direction_end:
+.size       rust_psm_stack_direction,.rust_psm_stack_direction_end-rust_psm_stack_direction
+.cfi_endproc
+
+
+.globl rust_psm_stack_pointer
+.align 2
+.type rust_psm_stack_pointer,@function
+rust_psm_stack_pointer:
+/* extern "C" fn() -> *mut u8 */
+.cfi_startproc
+    move $r4, $r3
+    jr $r1
+.rust_psm_stack_pointer_end:
+.size       rust_psm_stack_pointer,.rust_psm_stack_pointer_end-rust_psm_stack_pointer
+.cfi_endproc
+
+
+.globl rust_psm_replace_stack
+.align 2
+.type rust_psm_replace_stack,@function
+rust_psm_replace_stack:
+/* extern "C" fn(r4: usize, r5: extern "C" fn(usize), r6: *mut u8) */
+.cfi_startproc
+    move $r3, $r6
+    jr $r5
+.rust_psm_replace_stack_end:
+.size       rust_psm_replace_stack,.rust_psm_replace_stack_end-rust_psm_replace_stack
+.cfi_endproc
+
+
+.globl rust_psm_on_stack
+.align 2
+.type rust_psm_on_stack,@function
+rust_psm_on_stack:
+/* extern "C" fn(r4: usize, r5: usize, r6: extern "C" fn(usize, usize), r7: *mut u8) */
+.cfi_startproc
+    st.d $r1, $r7, -8
+    st.d $r3, $r7, -16
+    addi.d $r3, $r7, -16
+    .cfi_def_cfa 3, 16
+    .cfi_offset 1, -8
+    .cfi_offset 3, -16
+    jirl $r1, $r6, 0
+    ld.d $r1, $r3, 8
+    .cfi_restore 1
+    ld.d $r3, $r3, 0
+    .cfi_restore 3
+    jr $r1
+.rust_psm_on_stack_end:
+.size       rust_psm_on_stack,.rust_psm_on_stack_end-rust_psm_on_stack
+.cfi_endproc

vendor/psm/src/arch/mips64_eabi.s

@@ -0,0 +1,88 @@
+/*
+Not only MIPS has 20 different ABIs... nobody tells anybody what specific variant of which ABI is
+used where.
+
+This is an "EABI" implementation based on the following page:
+
+http://www.cygwin.com/ml/binutils/2003-06/msg00436.html
+*/
+
+#include "psm.h"
+#include "gnu_stack_note.s"
+
+.set noreorder /* we’ll manage the delay slots on our own, thanks! */
+
+.text
+.globl rust_psm_stack_direction
+.p2align 3
+.type rust_psm_stack_direction,@function
+.ent rust_psm_stack_direction
+/* extern "C" fn() -> u8 */
+rust_psm_stack_direction:
+.cfi_startproc
+    jr $31
+    addiu $2, $zero, STACK_DIRECTION_DESCENDING
+.end rust_psm_stack_direction
+.rust_psm_stack_direction_end:
+.size       rust_psm_stack_direction,.rust_psm_stack_direction_end-rust_psm_stack_direction
+.cfi_endproc
+
+
+.globl rust_psm_stack_pointer
+.p2align 3
+.type rust_psm_stack_pointer,@function
+.ent rust_psm_stack_pointer
+/* extern "C" fn() -> *mut u8 */
+rust_psm_stack_pointer:
+.cfi_startproc
+    jr $31
+    move $2, $29
+.end rust_psm_stack_pointer
+.rust_psm_stack_pointer_end:
+.size       rust_psm_stack_pointer,.rust_psm_stack_pointer_end-rust_psm_stack_pointer
+.cfi_endproc
+
+
+
+.globl rust_psm_replace_stack
+.p2align 3
+.type rust_psm_replace_stack,@function
+.ent rust_psm_replace_stack
+/* extern "C" fn(r4: usize, r5: extern "C" fn(usize), r6: *mut u8) */
+rust_psm_replace_stack:
+.cfi_startproc
+    move $25, $5
+    jr $5
+    move $29, $6
+.end rust_psm_replace_stack
+.rust_psm_replace_stack_end:
+.size       rust_psm_replace_stack,.rust_psm_on_stack_end-rust_psm_on_stack
+.cfi_endproc
+
+
+.globl rust_psm_on_stack
+.p2align 3
+.type rust_psm_on_stack,@function
+.ent rust_psm_on_stack
+/* extern "C" fn(r4: usize, r5: usize, r6: extern "C" fn(usize), r7: *mut u8) */
+rust_psm_on_stack:
+.cfi_startproc
+    sd $29, -8($7)
+    sd $31, -16($7)
+    .cfi_def_cfa 7, 0
+    .cfi_offset 31, -16
+    .cfi_offset 29, -8
+    move $25, $6
+    jalr $31, $6
+    daddiu $29, $7, -16
+    .cfi_def_cfa 29, 16
+    ld $31, 0($29)
+    .cfi_restore 31
+    ld $29, 8($29)
+    .cfi_restore 29
+    jr $31
+    nop
+.end rust_psm_on_stack
+.rust_psm_on_stack_end:
+.size       rust_psm_on_stack,.rust_psm_on_stack_end-rust_psm_on_stack
+.cfi_endproc

vendor/psm/src/arch/mips_eabi.s

@@ -0,0 +1,89 @@
+/*
+Not only MIPS has 20 different ABIs... nobody tells anybody what specific variant of which ABI is
+used where.
+
+This is an "EABI" implementation based on the following page:
+
+http://www.cygwin.com/ml/binutils/2003-06/msg00436.html
+*/
+
+#include "psm.h"
+#include "gnu_stack_note.s"
+
+.set noreorder /* we’ll manage the delay slots on our own, thanks! */
+
+.text
+.abicalls
+.globl rust_psm_stack_direction
+.p2align 2
+.type rust_psm_stack_direction,@function
+.ent rust_psm_stack_direction
+/* extern "C" fn() -> u8 */
+rust_psm_stack_direction:
+.cfi_startproc
+    jr $31
+    addiu $2, $zero, STACK_DIRECTION_DESCENDING
+.end rust_psm_stack_direction
+.rust_psm_stack_direction_end:
+.size       rust_psm_stack_direction,.rust_psm_stack_direction_end-rust_psm_stack_direction
+.cfi_endproc
+
+
+.globl rust_psm_stack_pointer
+.p2align 2
+.type rust_psm_stack_pointer,@function
+.ent rust_psm_stack_pointer
+/* extern "C" fn() -> *mut u8 */
+rust_psm_stack_pointer:
+.cfi_startproc
+    jr $31
+    move $2, $29
+.end rust_psm_stack_pointer
+.rust_psm_stack_pointer_end:
+.size       rust_psm_stack_pointer,.rust_psm_stack_pointer_end-rust_psm_stack_pointer
+.cfi_endproc
+
+
+
+.globl rust_psm_replace_stack
+.p2align 2
+.type rust_psm_replace_stack,@function
+.ent rust_psm_replace_stack
+/* extern "C" fn(r4: usize, r5: extern "C" fn(usize), r6: *mut u8) */
+rust_psm_replace_stack:
+.cfi_startproc
+    move $25, $5
+    jr $5
+    move $29, $6
+.end rust_psm_replace_stack
+.rust_psm_replace_stack_end:
+.size       rust_psm_replace_stack,.rust_psm_on_stack_end-rust_psm_on_stack
+.cfi_endproc
+
+
+.globl rust_psm_on_stack
+.p2align 2
+.type rust_psm_on_stack,@function
+.ent rust_psm_on_stack
+/* extern "C" fn(r4: usize, r5: usize, r6: extern "C" fn(usize), r7: *mut u8) */
+rust_psm_on_stack:
+.cfi_startproc
+    sw $29, -4($7)
+    sw $31, -8($7)
+    .cfi_def_cfa 7, 0
+    .cfi_offset 31, -8
+    .cfi_offset 29, -4
+    move $25, $6
+    jalr $31, $6
+    addiu $29, $7, -8
+    .cfi_def_cfa 29, 8
+    lw $31, 0($29)
+    .cfi_restore 31
+    lw $29, 4($29)
+    .cfi_restore 29
+    jr $31
+    nop
+.end rust_psm_on_stack
+.rust_psm_on_stack_end:
+.size       rust_psm_on_stack,.rust_psm_on_stack_end-rust_psm_on_stack
+.cfi_endproc

vendor/psm/src/arch/powerpc32.s

@@ -0,0 +1,77 @@
+#include "psm.h"
+#include "gnu_stack_note.s"
+/* FIXME: this probably does not cover all ABIs? Tested with sysv only, possibly works for AIX as
+   well?
+*/
+
+.text
+.globl rust_psm_stack_direction
+.p2align 2
+.type rust_psm_stack_direction,@function
+rust_psm_stack_direction:
+/* extern "C" fn() -> u8 */
+.cfi_startproc
+    li 3, STACK_DIRECTION_DESCENDING
+    blr
+.rust_psm_stack_direction_end:
+.size       rust_psm_stack_direction,.rust_psm_stack_direction_end-rust_psm_stack_direction
+.cfi_endproc
+
+
+.globl rust_psm_stack_pointer
+.p2align 2
+.type rust_psm_stack_pointer,@function
+rust_psm_stack_pointer:
+/* extern "C" fn() -> *mut u8 */
+.cfi_startproc
+    mr 3, 1
+    blr
+.rust_psm_stack_pointer_end:
+.size       rust_psm_stack_pointer,.rust_psm_stack_pointer_end-rust_psm_stack_pointer
+.cfi_endproc
+
+
+.globl rust_psm_replace_stack
+.p2align 2
+.type rust_psm_replace_stack,@function
+rust_psm_replace_stack:
+/* extern "C" fn(3: usize, 4: extern "C" fn(usize), 5: *mut u8) */
+.cfi_startproc
+/* NOTE: perhaps add a debug-assertion for stack alignment? */
+    addi 5, 5, -16
+    mr 1, 5
+    mtctr 4
+    bctr
+.rust_psm_replace_stack_end:
+.size       rust_psm_replace_stack,.rust_psm_replace_stack_end-rust_psm_replace_stack
+.cfi_endproc
+
+
+.globl rust_psm_on_stack
+.p2align 2
+.type rust_psm_on_stack,@function
+rust_psm_on_stack:
+/* extern "C" fn(3: usize, 4: usize, 5: extern "C" fn(usize, usize), 6: *mut u8) */
+.cfi_startproc
+    mflr 0
+    stw 0, -24(6)
+    sub 6, 6, 1
+    addi 6, 6, -32
+    stwux 1, 1, 6
+    .cfi_def_cfa r1, 32
+    .cfi_offset r1, -32
+    .cfi_offset lr, -24
+    mtctr 5
+    bctrl
+    lwz 0, 8(1)
+    mtlr 0
+    .cfi_restore lr
+    /* FIXME: after this instruction backtrace breaks until control returns to the caller
+       That being said compiler-generated code has the same issue, so I guess that is fine for now?
+    */
+    lwz 1, 0(1)
+    .cfi_restore r1
+    blr
+.rust_psm_on_stack_end:
+.size       rust_psm_on_stack,.rust_psm_on_stack_end-rust_psm_on_stack
+.cfi_endproc

vendor/psm/src/arch/powerpc64.s

@@ -0,0 +1,91 @@
+/* Implementation of the AIX-like PowerPC ABI. Seems to be used by the big-endian PowerPC targets.
+   The following references were used during the implementation of this code:
+
+   https://www.ibm.com/support/knowledgecenter/en/ssw_aix_72/com.ibm.aix.alangref/idalangref_rntime_stack.htm
+   https://www.ibm.com/support/knowledgecenter/en/ssw_aix_72/com.ibm.aix.alangref/idalangref_reg_use_conv.htm
+   https://www.ibm.com/developerworks/library/l-powasm4/index.html
+*/
+
+#include "psm.h"
+#include "gnu_stack_note.s"
+
+.text
+.globl rust_psm_stack_direction
+.p2align 2
+.type rust_psm_stack_direction,@function
+rust_psm_stack_direction:
+/* extern "C" fn() -> u8 */
+.cfi_startproc
+    li 3, STACK_DIRECTION_DESCENDING
+    blr
+.rust_psm_stack_direction_end:
+.size       rust_psm_stack_direction,.rust_psm_stack_direction_end-rust_psm_stack_direction
+.cfi_endproc
+
+
+.globl rust_psm_stack_pointer
+.p2align 2
+.type rust_psm_stack_pointer,@function
+rust_psm_stack_pointer:
+/* extern "C" fn() -> *mut u8 */
+.cfi_startproc
+    mr 3, 1
+    blr
+.rust_psm_stack_pointer_end:
+.size       rust_psm_stack_pointer,.rust_psm_stack_pointer_end-rust_psm_stack_pointer
+.cfi_endproc
+
+
+.globl rust_psm_replace_stack
+.p2align 2
+.type rust_psm_replace_stack,@function
+rust_psm_replace_stack:
+/* extern "C" fn(3: usize, 4: extern "C" fn(usize), 5: *mut u8) */
+.cfi_startproc
+    ld 2, 8(4)
+    ld 4, 0(4)
+    /* do not allocate the whole 112-byte sized frame, we know wont be used */
+    addi 5, 5, -48
+    mr 1, 5
+    mtctr 4
+    bctr
+.rust_psm_replace_stack_end:
+.size       rust_psm_replace_stack,.rust_psm_replace_stack_end-rust_psm_replace_stack
+.cfi_endproc
+
+
+.globl rust_psm_on_stack
+.p2align 2
+.type rust_psm_on_stack,@function
+rust_psm_on_stack:
+/* extern "C" fn(3: usize, 4: usize, 5: extern "C" fn(usize, usize), 6: *mut u8) */
+.cfi_startproc
+    mflr 0
+    std 2, -72(6)
+    std 0, -8(6)
+    sub 6, 6, 1
+    addi 6, 6, -112
+    stdux 1, 1, 6
+    .cfi_def_cfa r1, 112
+    .cfi_offset r1, -112
+    .cfi_offset r2, -72
+    .cfi_offset lr, -8
+    /* load the function pointer from TOC and make the call */
+    ld 2, 8(5)
+    ld 5, 0(5)
+    mtctr 5
+    bctrl
+    ld 2, 40(1)
+    .cfi_restore r2
+    ld 0, 104(1)
+    mtlr 0
+    .cfi_restore lr
+    /* FIXME: after this instruction backtrace breaks until control returns to the caller.
+       That being said compiler-generated code has the same issue, so I guess that is fine for now?
+    */
+    ld 1, 0(1)
+    .cfi_restore r1
+    blr
+.rust_psm_on_stack_end:
+.size       rust_psm_on_stack,.rust_psm_on_stack_end-rust_psm_on_stack
+.cfi_endproc

vendor/psm/src/arch/powerpc64_aix.s

@@ -0,0 +1,137 @@
+.csect .text[PR],2
+.file "powerpc64_aix.s"
+
+.globl  rust_psm_stack_direction[DS]
+.globl  .rust_psm_stack_direction
+.align  4
+.csect rust_psm_stack_direction[DS],3
+.vbyte 8, .rust_psm_stack_direction
+.vbyte 8, TOC[TC0]
+.vbyte 8, 0
+.csect .text[PR],2
+.rust_psm_stack_direction:
+# extern "C" fn() -> u8
+  li 3, 2
+  blr
+L..rust_psm_stack_direction_end:
+# Following bytes form the traceback table on AIX.
+# For specification, see https://www.ibm.com/docs/en/aix/7.2?topic=processor-traceback-tables.
+# For implementation, see https://github.com/llvm/llvm-project/blob/main/llvm/lib/Target/PowerPC/PPCAsmPrinter.cpp,
+# `PPCAIXAsmPrinter::emitTracebackTable`.
+.vbyte 4, 0x00000000 # Traceback table begin, for unwinder to search the table.
+.byte 0x00 # Version = 0
+.byte 0x09 # Language = CPlusPlus, since rust is using C++-like LSDA.
+.byte 0x20 # -IsGlobaLinkage, -IsOutOfLineEpilogOrPrologue
+           # +HasTraceBackTableOffset, -IsInternalProcedure
+           # -HasControlledStorage, -IsTOCless
+           # -IsFloatingPointPresent
+           # -IsFloatingPointOperationLogOrAbortEnabled
+.byte 0x40 # -IsInterruptHandler, +IsFunctionNamePresent, -IsAllocaUsed
+           # OnConditionDirective = 0, -IsCRSaved, -IsLRSaved
+.byte 0x80 # +IsBackChainStored, -IsFixup, NumOfFPRsSaved = 0
+.byte 0x00 # -HasExtensionTable, -HasVectorInfo, NumOfGPRsSaved = 0
+.byte 0x00 # NumberOfFixedParms = 0
+.byte 0x01 # NumberOfFPParms = 0, +HasParmsOnStack
+.vbyte 4, L..rust_psm_stack_direction_end-.rust_psm_stack_direction #Function size
+.vbyte 2, 0x0018 # Function name len = 24
+.byte "rust_psm_stack_direction" # Function Name
+
+.globl rust_psm_stack_pointer[DS]
+.globl .rust_psm_stack_pointer
+.align 4
+.csect rust_psm_stack_pointer[DS],3
+.vbyte 8, .rust_psm_stack_pointer
+.vbyte 8, TOC[TC0]
+.vbyte 8, 0
+.csect .text[PR],2
+.rust_psm_stack_pointer:
+# extern "C" fn() -> *mut u8
+  mr 3, 1
+  blr
+L..rust_psm_stack_pointer_end:
+.vbyte 4, 0x00000000
+.byte 0x00
+.byte 0x09
+.byte 0x20
+.byte 0x40
+.byte 0x80
+.byte 0x00
+.byte 0x00
+.byte 0x01
+.vbyte 4, L..rust_psm_stack_pointer_end-.rust_psm_stack_pointer
+.vbyte 2, 0x0016
+.byte "rust_psm_stack_pointer"
+
+.globl rust_psm_replace_stack[DS]
+.globl .rust_psm_replace_stack
+.align 4
+.csect rust_psm_replace_stack[DS],3
+.vbyte 8, .rust_psm_replace_stack
+.vbyte 8, TOC[TC0]
+.vbyte 8, 0
+.csect .text[PR],2
+.rust_psm_replace_stack:
+# extern "C" fn(3: usize, 4: extern "C" fn(usize), 5: *mut u8)
+  # Load the function pointer and toc pointer from TOC and make the call.
+  ld 2, 8(4)
+  ld 4, 0(4)
+  addi 5, 5, -48
+  mr 1, 5
+  mtctr 4
+  bctr
+L..rust_psm_replace_stack_end:
+.vbyte 4, 0x00000000
+.byte 0x00
+.byte 0x09
+.byte 0x20
+.byte 0x40
+.byte 0x80
+.byte 0x00
+.byte 0x03
+.byte 0x01
+.vbyte 4, 0x00000000 # Parameter type = i, i, i
+.vbyte 4, L..rust_psm_replace_stack_end-.rust_psm_replace_stack
+.vbyte 2, 0x0016
+.byte "rust_psm_replace_stack"
+
+.globl rust_psm_on_stack[DS]
+.globl .rust_psm_on_stack
+.align 4
+.csect rust_psm_on_stack[DS],3
+.vbyte 8, .rust_psm_on_stack
+.vbyte 8, TOC[TC0]
+.vbyte 8, 0
+.csect .text[PR],2
+.rust_psm_on_stack:
+# extern "C" fn(3: usize, 4: usize, 5: extern "C" fn(usize, usize), 6: *mut u8)
+  mflr 0
+  std 2, -72(6)
+  std 0, -8(6)
+  sub 6, 6, 1
+  addi 6, 6, -112
+  stdux 1, 1, 6
+  ld 2, 8(5)
+  ld 5, 0(5)
+  mtctr 5
+  bctrl
+  ld 2, 40(1)
+  ld 0, 104(1)
+  mtlr 0
+  ld 1, 0(1)
+  blr
+L..rust_psm_on_stack_end:
+.vbyte 4, 0x00000000
+.byte 0x00
+.byte 0x09
+.byte 0x20
+.byte 0x41
+.byte 0x80
+.byte 0x00
+.byte 0x04
+.byte 0x01
+.vbyte 4, 0x00000000 # Parameter type = i, i, i, i
+.vbyte 4, L..rust_psm_on_stack_end-.rust_psm_on_stack
+.vbyte 2, 0x0011
+.byte "rust_psm_on_stack"
+
+.toc

vendor/psm/src/arch/powerpc64_openpower.s

@@ -0,0 +1,87 @@
+/* Implementation of stack swtiching routines for OpenPOWER 64-bit ELF ABI
+   The specification can be found at
+   http://openpowerfoundation.org/wp-content/uploads/resources/leabi/content/ch_preface.html
+
+   This ABI is usually used by the ppc64le targets.
+*/
+
+#include "psm.h"
+#include "gnu_stack_note.s"
+
+.text
+.abiversion 2
+
+
+.globl rust_psm_stack_direction
+.p2align 4
+.type rust_psm_stack_direction,@function
+rust_psm_stack_direction:
+/* extern "C" fn() -> u8 */
+.cfi_startproc
+    li 3, STACK_DIRECTION_DESCENDING
+    blr
+.rust_psm_stack_direction_end:
+.size       rust_psm_stack_direction,.rust_psm_stack_direction_end-rust_psm_stack_direction
+.cfi_endproc
+
+
+.globl rust_psm_stack_pointer
+.p2align 4
+.type rust_psm_stack_pointer,@function
+rust_psm_stack_pointer:
+/* extern "C" fn() -> *mut u8 */
+.cfi_startproc
+    mr 3, 1
+    blr
+.rust_psm_stack_pointer_end:
+.size       rust_psm_stack_pointer,.rust_psm_stack_pointer_end-rust_psm_stack_pointer
+.cfi_endproc
+
+
+.globl rust_psm_replace_stack
+.p2align 4
+.type rust_psm_replace_stack,@function
+rust_psm_replace_stack:
+/* extern "C" fn(3: usize, 4: extern "C" fn(usize), 5: *mut u8) */
+.cfi_startproc
+    addi 5, 5, -32
+    mtctr 4
+    mr 12, 4
+    mr 1, 5
+    bctr
+.rust_psm_replace_stack_end:
+.size       rust_psm_replace_stack,.rust_psm_replace_stack_end-rust_psm_replace_stack
+.cfi_endproc
+
+
+
+.globl rust_psm_on_stack
+.p2align 4
+.type rust_psm_on_stack,@function
+rust_psm_on_stack:
+/* extern "C" fn(3: usize, 4: usize, 5: extern "C" fn(usize, usize), 6: *mut u8) */
+.cfi_startproc
+    mflr 0
+    std 0, -8(6)
+    std 2, -24(6)
+    sub 6, 6, 1
+    addi 6, 6, -48
+    stdux 1, 1, 6
+    .cfi_def_cfa r1, 48
+    .cfi_offset r1, -48
+    .cfi_offset r2, -24
+    .cfi_offset lr, -8
+    mr 12, 5
+    mtctr 5
+    bctrl
+    ld 2, 24(1)
+    .cfi_restore r2
+    ld 0, 40(1)
+    mtlr 0
+    .cfi_restore lr
+    /* FIXME: after this instructin backtrace breaks until control returns to the caller */
+    ld 1, 0(1)
+    blr
+.rust_psm_on_stack_end:
+.size       rust_psm_on_stack,.rust_psm_on_stack_end-rust_psm_on_stack
+.cfi_endproc

vendor/psm/src/arch/psm.h

@@ -0,0 +1,10 @@
+#define STACK_DIRECTION_ASCENDING 1
+#define STACK_DIRECTION_DESCENDING 2
+
+
+/*
+Various defines for values produced by `-DCFG_TARGET_*` flags. Only needs to be mutually unique
+*/
+#define darwin 1
+#define macos 2
+#define ios 3

vendor/psm/src/arch/riscv.s

@@ -0,0 +1,64 @@
+#include "psm.h"
+#include "gnu_stack_note.s"
+
+.text
+.globl rust_psm_stack_direction
+.p2align 2
+.type rust_psm_stack_direction,@function
+rust_psm_stack_direction:
+/* extern "C" fn() -> u8 */
+.cfi_startproc
+    li x10, STACK_DIRECTION_DESCENDING
+    jr x1
+.rust_psm_stack_direction_end:
+.size       rust_psm_stack_direction,.rust_psm_stack_direction_end-rust_psm_stack_direction
+.cfi_endproc
+
+
+.globl rust_psm_stack_pointer
+.p2align 2
+.type rust_psm_stack_pointer,@function
+rust_psm_stack_pointer:
+/* extern "C" fn() -> *mut u8 */
+.cfi_startproc
+    add x10, x2, x0
+    jr x1
+.rust_psm_stack_pointer_end:
+.size       rust_psm_stack_pointer,.rust_psm_stack_pointer_end-rust_psm_stack_pointer
+.cfi_endproc
+
+
+.globl rust_psm_replace_stack
+.p2align 2
+.type rust_psm_replace_stack,@function
+rust_psm_replace_stack:
+/* extern "C" fn(x10: usize, x11: extern "C" fn(usize), x12: *mut u8) */
+.cfi_startproc
+    add x2, x12, x0
+    jr x11
+.rust_psm_replace_stack_end:
+.size       rust_psm_replace_stack,.rust_psm_replace_stack_end-rust_psm_replace_stack
+.cfi_endproc
+
+
+.globl rust_psm_on_stack
+.p2align 2
+.type rust_psm_on_stack,@function
+rust_psm_on_stack:
+/* extern "C" fn(x10: usize, x11: usize, x12: extern "C" fn(usize, usize), x13: *mut u8) */
+.cfi_startproc
+    sw x1, -12(x13)
+    sw x2, -16(x13)
+    addi x2, x13, -16
+    .cfi_def_cfa x2, 16
+    .cfi_offset x1, -12
+    .cfi_offset x2, -16
+    jalr x1, x12, 0
+    lw x1, 4(x2)
+    .cfi_restore x1
+    lw x2, 0(x2)
+    .cfi_restore x2
+    jr x1
+.rust_psm_on_stack_end:
+.size       rust_psm_on_stack,.rust_psm_on_stack_end-rust_psm_on_stack
+.cfi_endproc

vendor/psm/src/arch/riscv64.s

@@ -0,0 +1,64 @@
+#include "psm.h"
+#include "gnu_stack_note.s"
+
+.text
+.globl rust_psm_stack_direction
+.p2align 2
+.type rust_psm_stack_direction,@function
+rust_psm_stack_direction:
+/* extern "C" fn() -> u8 */
+.cfi_startproc
+    li x10, STACK_DIRECTION_DESCENDING
+    jr x1
+.rust_psm_stack_direction_end:
+.size       rust_psm_stack_direction,.rust_psm_stack_direction_end-rust_psm_stack_direction
+.cfi_endproc
+
+
+.globl rust_psm_stack_pointer
+.p2align 2
+.type rust_psm_stack_pointer,@function
+rust_psm_stack_pointer:
+/* extern "C" fn() -> *mut u8 */
+.cfi_startproc
+    add x10, x2, x0
+    jr x1
+.rust_psm_stack_pointer_end:
+.size       rust_psm_stack_pointer,.rust_psm_stack_pointer_end-rust_psm_stack_pointer
+.cfi_endproc
+
+
+.globl rust_psm_replace_stack
+.p2align 2
+.type rust_psm_replace_stack,@function
+rust_psm_replace_stack:
+/* extern "C" fn(x10: usize, x11: extern "C" fn(usize), x12: *mut u8) */
+.cfi_startproc
+    add x2, x12, x0
+    jr x11
+.rust_psm_replace_stack_end:
+.size       rust_psm_replace_stack,.rust_psm_replace_stack_end-rust_psm_replace_stack
+.cfi_endproc
+
+
+.globl rust_psm_on_stack
+.p2align 2
+.type rust_psm_on_stack,@function
+rust_psm_on_stack:
+/* extern "C" fn(x10: usize, x11: usize, x12: extern "C" fn(usize, usize), x13: *mut u8) */
+.cfi_startproc
+    sd x1, -8(x13)
+    sd x2, -16(x13)
+    addi x2, x13, -16
+    .cfi_def_cfa x2, 16
+    .cfi_offset x1, -8
+    .cfi_offset x2, -16
+    jalr x1, x12, 0
+    ld x1, 8(x2)
+    .cfi_restore x1
+    ld x2, 0(x2)
+    .cfi_restore x2
+    jr x1
+.rust_psm_on_stack_end:
+.size       rust_psm_on_stack,.rust_psm_on_stack_end-rust_psm_on_stack
+.cfi_endproc

vendor/psm/src/arch/sparc64.s

@@ -0,0 +1,68 @@
+#include "psm.h"
+#include "gnu_stack_note.s"
+
+.text
+.globl rust_psm_stack_direction
+.p2align 2
+.type rust_psm_stack_direction,@function
+rust_psm_stack_direction:
+/* extern "C" fn() -> u8 */
+.cfi_startproc
+    jmpl %o7 + 8, %g0
+    mov STACK_DIRECTION_DESCENDING, %o0
+.rust_psm_stack_direction_end:
+.size       rust_psm_stack_direction,.rust_psm_stack_direction_end-rust_psm_stack_direction
+.cfi_endproc
+
+
+.globl rust_psm_stack_pointer
+.p2align 2
+.type rust_psm_stack_pointer,@function
+rust_psm_stack_pointer:
+/* extern "C" fn() -> *mut u8 */
+.cfi_startproc
+    jmpl %o7 + 8, %g0
+    mov %o6, %o0
+.rust_psm_stack_pointer_end:
+.size       rust_psm_stack_pointer,.rust_psm_stack_pointer_end-rust_psm_stack_pointer
+.cfi_endproc
+
+
+.globl rust_psm_replace_stack
+.p2align 2
+.type rust_psm_replace_stack,@function
+rust_psm_replace_stack:
+/* extern "C" fn(%i0: usize, %i1: extern "C" fn(usize), %i2: *mut u8) */
+.cfi_startproc
+    .cfi_def_cfa 0, 0
+    .cfi_return_column 0
+    jmpl %o1, %g0
+    /* WEIRD: Why is the LSB set for the %sp and %fp on SPARC?? */
+    add %o2, -0x7ff, %o6
+.rust_psm_replace_stack_end:
+.size       rust_psm_replace_stack,.rust_psm_replace_stack_end-rust_psm_replace_stack
+.cfi_endproc
+
+
+.globl rust_psm_on_stack
+.p2align 2
+.type rust_psm_on_stack,@function
+rust_psm_on_stack:
+/* extern "C" fn(%i0: usize, %i1: usize, %i2: extern "C" fn(usize, usize), %i3: *mut u8) */
+.cfi_startproc
+    /* The fact that locals and saved register windows are offset by 2kB is
+       very nasty property of SPARC architecture and ABI. In this case it forces us to slice off
+       2kB of the stack space outright for no good reason other than adapting to a botched design.
+    */
+    save %o3, -0x87f, %o6
+    .cfi_def_cfa_register %fp
+    .cfi_window_save
+    .cfi_register %r15, %r31
+    mov %i1, %o1
+    jmpl %i2, %o7
+    mov %i0, %o0
+    ret
+    restore
+.rust_psm_on_stack_end:
+.size       rust_psm_on_stack,.rust_psm_on_stack_end-rust_psm_on_stack
+.cfi_endproc

vendor/psm/src/arch/sparc_sysv.s

@@ -0,0 +1,66 @@
+#include "psm.h"
+#include "gnu_stack_note.s"
+
+/* FIXME: this ABI has definitely not been verified at all */
+
+.text
+.globl rust_psm_stack_direction
+.p2align 2
+.type rust_psm_stack_direction,@function
+rust_psm_stack_direction:
+/* extern "C" fn() -> u8 */
+.cfi_startproc
+    jmpl %o7 + 8, %g0
+    mov STACK_DIRECTION_DESCENDING, %o0
+.rust_psm_stack_direction_end:
+.size       rust_psm_stack_direction,.rust_psm_stack_direction_end-rust_psm_stack_direction
+.cfi_endproc
+
+
+.globl rust_psm_stack_pointer
+.p2align 2
+.type rust_psm_stack_pointer,@function
+rust_psm_stack_pointer:
+/* extern "C" fn() -> *mut u8 */
+.cfi_startproc
+    jmpl %o7 + 8, %g0
+    mov %o6, %o0
+.rust_psm_stack_pointer_end:
+.size       rust_psm_stack_pointer,.rust_psm_stack_pointer_end-rust_psm_stack_pointer
+.cfi_endproc
+
+
+.globl rust_psm_replace_stack
+.p2align 2
+.type rust_psm_replace_stack,@function
+rust_psm_replace_stack:
+/* extern "C" fn(%i0: usize, %i1: extern "C" fn(usize), %i2: *mut u8) */
+.cfi_startproc
+    .cfi_def_cfa 0, 0
+    .cfi_return_column 0
+    jmpl %o1, %g0
+    /* WEIRD: Why is the LSB set for the %sp and %fp on SPARC?? */
+    add %o2, -0x3ff, %o6
+.rust_psm_replace_stack_end:
+.size       rust_psm_replace_stack,.rust_psm_replace_stack_end-rust_psm_replace_stack
+.cfi_endproc
+
+
+.globl rust_psm_on_stack
+.p2align 2
+.type rust_psm_on_stack,@function
+rust_psm_on_stack:
+/* extern "C" fn(%i0: usize, %i1: usize, %i2: extern "C" fn(usize, usize), %i3: *mut u8) */
+.cfi_startproc
+    save %o3, -0x43f, %o6
+    .cfi_def_cfa_register %fp
+    .cfi_window_save
+    .cfi_register %r15, %r31
+    mov %i1, %o1
+    jmpl %i2, %o7
+    mov %i0, %o0
+    ret
+    restore
+.rust_psm_on_stack_end:
+.size       rust_psm_on_stack,.rust_psm_on_stack_end-rust_psm_on_stack
+.cfi_endproc

vendor/psm/src/arch/wasm32.s

@@ -0,0 +1,61 @@
+#include "psm.h"
+#include "gnu_stack_note.s"
+
+# Note that this function is not compiled when this package is uploaded to
+# crates.io, this source is only here as a reference for how the corresponding
+# wasm32.o was generated. This file can be compiled with:
+#
+#    cpp psm/src/arch/wasm32.s | llvm-mc -o psm/src/arch/wasm32.o --arch=wasm32 -filetype=obj
+#
+# where you'll want to ensure that `llvm-mc` is from a relatively recent
+# version of LLVM.
+
+.globaltype __stack_pointer, i32
+
+.globl rust_psm_stack_direction
+.type rust_psm_stack_direction,@function
+rust_psm_stack_direction:
+.functype rust_psm_stack_direction () -> (i32)
+    i32.const STACK_DIRECTION_DESCENDING
+    end_function
+
+.globl rust_psm_stack_pointer
+.type rust_psm_stack_pointer,@function
+rust_psm_stack_pointer:
+.functype rust_psm_stack_pointer () -> (i32)
+    global.get __stack_pointer
+    end_function
+
+.globl rust_psm_on_stack
+.type rust_psm_on_stack,@function
+rust_psm_on_stack:
+.functype rust_psm_on_stack (i32, i32, i32, i32) -> ()
+    # get our new stack argument, then save the old stack
+    # pointer into that local
+    local.get 3
+    global.get __stack_pointer
+    local.set 3
+    global.set __stack_pointer
+
+    # Call our indirect function specified
+    local.get 0
+    local.get 1
+    local.get 2
+    call_indirect (i32, i32) -> ()
+
+    # restore the stack pointer before returning
+    local.get 3
+    global.set __stack_pointer
+    end_function
+
+.globl rust_psm_replace_stack
+.type rust_psm_replace_stack,@function
+rust_psm_replace_stack:
+.functype rust_psm_replace_stack (i32, i32, i32) -> ()
+    local.get 2
+    global.set __stack_pointer
+    local.get 0
+    local.get 1
+    call_indirect (i32) -> ()
+    unreachable
+    end_function

vendor/psm/src/arch/x86.s

@@ -0,0 +1,92 @@
+#include "psm.h"
+#include "gnu_stack_note.s"
+/* NOTE: fastcall calling convention used on all x86 targets */
+
+.text
+
+#if defined(CFG_TARGET_OS_darwin) || defined(CFG_TARGET_OS_macos) || defined(CFG_TARGET_OS_ios)
+
+#define GLOBL(fnname) .globl _##fnname
+#define TYPE(fnname)
+#define FUNCTION(fnname) _##fnname
+#define SIZE(fnname,endlabel)
+
+#else
+
+#define GLOBL(fnname) .globl fnname
+#define TYPE(fnname) .type fnname,@function
+#define FUNCTION(fnname) fnname
+#define SIZE(fnname,endlabel) .size fnname,endlabel-fnname
+
+#endif
+
+
+GLOBL(rust_psm_stack_direction)
+.p2align 4
+TYPE(rust_psm_stack_direction)
+FUNCTION(rust_psm_stack_direction):
+/* extern "fastcall" fn() -> u8 (%al) */
+.cfi_startproc
+    movb $STACK_DIRECTION_DESCENDING, %al # always descending on x86_64
+    retl
+.rust_psm_stack_direction_end:
+SIZE(rust_psm_stack_direction,.rust_psm_stack_direction_end)
+.cfi_endproc
+
+
+GLOBL(rust_psm_stack_pointer)
+.p2align 4
+TYPE(rust_psm_stack_pointer)
+FUNCTION(rust_psm_stack_pointer):
+/* extern "fastcall" fn() -> *mut u8 (%rax) */
+.cfi_startproc
+    leal 4(%esp), %eax
+    retl
+.rust_psm_stack_pointer_end:
+SIZE(rust_psm_stack_pointer,.rust_psm_stack_pointer_end)
+.cfi_endproc
+
+
+GLOBL(rust_psm_replace_stack)
+.p2align 4
+TYPE(rust_psm_replace_stack)
+FUNCTION(rust_psm_replace_stack):
+/* extern "fastcall" fn(%ecx: usize, %edx: extern "fastcall" fn(usize), 4(%esp): *mut u8) */
+.cfi_startproc
+/*
+   All we gotta do is set the stack pointer to 4(%esp) & tail-call the callback in %edx
+
+   Note, that the callee expects the stack to be offset by 4 bytes (normally, a return address
+   would be store there) off the required stack alignment on entry. To offset the stack in such a
+   way we use the `calll` instruction, however it would also be possible to to use plain `jmpl` but
+   would require to adjust the stack manually, which cannot be easily done, because the stack
+   pointer argument is already stored in memory.
+ */
+    movl 4(%esp), %esp
+    calll *%edx
+    ud2
+.rust_psm_replace_stack_end:
+SIZE(rust_psm_replace_stack,.rust_psm_replace_stack_end)
+.cfi_endproc
+
+
+GLOBL(rust_psm_on_stack)
+.p2align 4
+TYPE(rust_psm_on_stack)
+FUNCTION(rust_psm_on_stack):
+/* extern "fastcall" fn(%ecx: usize, %edx: usize, 4(%esp): extern "fastcall" fn(usize, usize), 8(%esp): *mut u8) */
+.cfi_startproc
+    pushl %ebp
+    .cfi_def_cfa %esp, 8
+    .cfi_offset %ebp, -8
+    movl  %esp, %ebp
+    .cfi_def_cfa_register %ebp
+    movl  12(%ebp), %esp
+    calll *8(%ebp)
+    movl  %ebp, %esp
+    popl  %ebp
+    .cfi_def_cfa %esp, 4
+    retl  $8
+.rust_psm_on_stack_end:
+SIZE(rust_psm_on_stack,.rust_psm_on_stack_end)
+.cfi_endproc

vendor/psm/src/arch/x86_64.s

@@ -0,0 +1,86 @@
+#include "psm.h"
+#include "gnu_stack_note.s"
+/* NOTE: sysv64 calling convention is used on all x86_64 targets, including Windows! */
+
+.text
+
+#if defined(CFG_TARGET_OS_darwin) || defined(CFG_TARGET_OS_macos) || defined(CFG_TARGET_OS_ios)
+
+#define GLOBL(fnname) .globl _##fnname
+#define TYPE(fnname)
+#define FUNCTION(fnname) _##fnname
+#define END_FUNCTION(fnname)
+
+#else
+
+#define GLOBL(fnname) .globl fnname
+#define TYPE(fnname) .type fnname,@function
+#define FUNCTION(fnname) fnname
+#define END_FUNCTION(fnname) .size fnname,.-fnname
+
+#endif
+
+
+GLOBL(rust_psm_stack_direction)
+.p2align 4
+TYPE(rust_psm_stack_direction)
+FUNCTION(rust_psm_stack_direction):
+/* extern "sysv64" fn() -> u8 (%al) */
+.cfi_startproc
+    movb $STACK_DIRECTION_DESCENDING, %al # always descending on x86_64
+    retq
+END_FUNCTION(rust_psm_stack_direction)
+.cfi_endproc
+
+
+GLOBL(rust_psm_stack_pointer)
+.p2align 4
+TYPE(rust_psm_stack_pointer)
+FUNCTION(rust_psm_stack_pointer):
+/* extern "sysv64" fn() -> *mut u8 (%rax) */
+.cfi_startproc
+    leaq 8(%rsp), %rax
+    retq
+.rust_psm_stack_pointer_end:
+END_FUNCTION(rust_psm_stack_pointer)
+.cfi_endproc
+
+
+GLOBL(rust_psm_replace_stack)
+.p2align 4
+TYPE(rust_psm_replace_stack)
+FUNCTION(rust_psm_replace_stack):
+/* extern "sysv64" fn(%rdi: usize, %rsi: extern "sysv64" fn(usize), %rdx: *mut u8) */
+.cfi_startproc
+/*
+    All we gotta do is set the stack pointer to %rdx & tail-call the callback in %rsi.
+
+    8-byte offset necessary to account for the "return" pointer that would otherwise be placed onto
+    stack with a regular call
+*/
+    leaq -8(%rdx), %rsp
+    jmpq *%rsi
+.rust_psm_replace_stack_end:
+END_FUNCTION(rust_psm_replace_stack)
+.cfi_endproc
+
+
+GLOBL(rust_psm_on_stack)
+.p2align 4
+TYPE(rust_psm_on_stack)
+FUNCTION(rust_psm_on_stack):
+/* extern "sysv64" fn(%rdi: usize, %rsi: usize, %rdx: extern "sysv64" fn(usize, usize), %rcx: *mut u8) */
+.cfi_startproc
+    pushq %rbp
+    .cfi_def_cfa %rsp, 16
+    .cfi_offset %rbp, -16
+    movq  %rsp, %rbp
+    .cfi_def_cfa_register %rbp
+    movq  %rcx, %rsp
+    callq *%rdx
+    movq  %rbp, %rsp
+    popq  %rbp
+    .cfi_def_cfa %rsp, 8
+    retq
+END_FUNCTION(rust_psm_on_stack)
+.cfi_endproc

vendor/psm/src/arch/x86_64_msvc.asm

@@ -0,0 +1,61 @@
+PUBLIC rust_psm_stack_direction
+PUBLIC rust_psm_stack_pointer
+PUBLIC rust_psm_replace_stack
+PUBLIC rust_psm_on_stack
+
+_TEXT SEGMENT
+
+; extern "sysv64" fn() -> u8 (%al)
+rust_psm_stack_direction PROC
+    mov al, 2
+    ret
+rust_psm_stack_direction ENDP
+
+; extern "sysv64" fn() -> *mut u8 (%rax)
+rust_psm_stack_pointer PROC
+    lea rax, [rsp + 8]
+    ret
+rust_psm_stack_pointer ENDP
+
+; extern "sysv64" fn(%rdi: usize, %rsi: extern "sysv64" fn(usize), %rdx: *mut u8, %rcx: *mut u8)
+rust_psm_replace_stack PROC
+    mov gs:[08h], rdx
+    mov gs:[10h], rcx
+    lea rsp, [rdx - 8]
+    jmp rsi
+rust_psm_replace_stack ENDP
+
+; extern "sysv64" fn(%rdi: usize, %rsi: usize,
+;                    %rdx: extern "sysv64" fn(usize, usize), %rcx: *mut u8, %r8: *mut u8)
+;
+; NB: on Windows for SEH to work at all, the pointers in TIB, thread information block, need to be
+; fixed up. Otherwise, it seems that exception mechanism on Windows will not bother looking for
+; exception handlers at *all* if they happen to fall outside the are specified in TIB.
+;
+; This necessitates an API difference from the usual 4-argument signature used elsewhere.
+;
+; FIXME: this needs a catch-all exception handler that aborts in case somebody unwinds into here.
+rust_psm_on_stack PROC FRAME
+    push rbp
+    .pushreg rbp
+    mov rbp, rsp
+    .setframe rbp, 0
+    .endprolog
+
+    push gs:[08h]
+    mov gs:[08h], rcx
+    push gs:[10h]
+    mov gs:[10h], r8
+    mov rsp, rcx
+    call rdx
+    lea rsp, [rbp - 010h]
+    pop gs:[10h]
+    pop gs:[08h]
+
+    pop rbp
+    ret
+rust_psm_on_stack ENDP
+
+_TEXT ENDS
+
+END

vendor/psm/src/arch/x86_64_windows_gnu.s

@@ -0,0 +1,97 @@
+#include "gnu_stack_note.s"
+
+.text
+
+.def rust_psm_stack_direction
+.scl 2
+.type 32
+.endef
+.globl rust_psm_stack_direction
+.p2align 4
+rust_psm_stack_direction:
+/* extern "sysv64" fn() -> u8 (%al) */
+.cfi_startproc
+    movb $2, %al # always descending on x86_64
+    retq
+.cfi_endproc
+
+.def rust_psm_stack_pointer
+.scl 2
+.type 32
+.endef
+.globl rust_psm_stack_pointer
+.p2align 4
+rust_psm_stack_pointer:
+/* extern "sysv64" fn() -> *mut u8 (%rax) */
+.cfi_startproc
+    leaq 8(%rsp), %rax
+    retq
+.cfi_endproc
+
+
+.def rust_psm_replace_stack
+.scl 2
+.type 32
+.endef
+.globl rust_psm_replace_stack
+.p2align 4
+rust_psm_replace_stack:
+/* extern "sysv64" fn(%rdi: usize, %rsi: extern "sysv64" fn(usize), %rdx: *mut u8, %rcx: *mut u8) */
+.cfi_startproc
+/*
+    All we gotta do is set the stack pointer to %rdx & tail-call the callback in %rsi.
+
+    8-byte offset necessary to account for the "return" pointer that would otherwise be placed onto
+    stack with a regular call
+*/
+    movq %gs:0x08, %rdx
+    movq %gs:0x10, %rcx
+    leaq -8(%rdx), %rsp
+    jmpq *%rsi
+.cfi_endproc
+
+
+.def rust_psm_on_stack
+.scl 2
+.type 32
+.endef
+.globl rust_psm_on_stack
+.p2align 4
+rust_psm_on_stack:
+/*
+extern "sysv64" fn(%rdi: usize, %rsi: usize,
+                   %rdx: extern "sysv64" fn(usize, usize), %rcx: *mut u8, %r8: *mut u8)
+
+NB: on Windows for SEH to work at all, the pointers in TIB, thread information block, need to be
+fixed up. Otherwise, it seems that exception mechanism on Windows will not bother looking for
+exception handlers at *all* if they happen to fall outside the are specified in TIB.
+
+This necessitates an API difference from the usual 4-argument signature used elsewhere.
+
+FIXME: this needs a catch-all exception handler that aborts in case somebody unwinds into here.
+*/
+.cfi_startproc
+    pushq %rbp
+    .cfi_def_cfa %rsp, 16
+    .cfi_offset %rbp, -16
+    pushq %gs:0x08
+    .cfi_def_cfa %rsp, 24
+    pushq %gs:0x10
+    .cfi_def_cfa %rsp, 32
+
+    movq  %rsp, %rbp
+    .cfi_def_cfa_register %rbp
+    movq %rcx, %gs:0x08
+    movq %r8, %gs:0x10
+    movq  %rcx, %rsp
+    callq *%rdx
+    movq  %rbp, %rsp
+
+    popq %gs:0x10
+    .cfi_def_cfa %rsp, 24
+    popq %gs:0x08
+    .cfi_def_cfa %rsp, 16
+    popq  %rbp
+    .cfi_def_cfa %rsp, 8
+    retq
+.cfi_endproc

vendor/psm/src/arch/x86_msvc.asm

@@ -0,0 +1,70 @@
+; FIXME: this is weird, this works locally but not on appveyor?!??!
+
+.386
+.model flat
+
+ASSUME FS:NOTHING
+
+; WTF: PUBLIC conflicts with "SYSCALL" but "SYSCALL" is the only way to stop MASM from manging the
+; symbol names?
+;
+; PUBLIC @rust_psm_stack_direction@0
+; PUBLIC @rust_psm_stack_pointer@0
+; PUBLIC @rust_psm_replace_stack@12
+; PUBLIC @rust_psm_on_stack@16
+
+_TEXT SEGMENT
+
+; extern "fastcall" fn() -> u8 (%al)
+@rust_psm_stack_direction@0 PROC SYSCALL
+    mov al, 2
+    ret
+@rust_psm_stack_direction@0 ENDP
+
+; extern "fastcall" fn() -> *mut u8 (%rax)
+@rust_psm_stack_pointer@0 PROC SYSCALL
+    lea eax, [esp + 4]
+    ret
+@rust_psm_stack_pointer@0 ENDP
+
+; extern "fastcall" fn(%ecx: usize, %edx: extern "fastcall" fn(usize),
+;                      4(%esp): *mut u8, 8(%esp): *mut u8)
+@rust_psm_replace_stack@16 PROC SYSCALL
+    mov eax, dword ptr [esp + 8]
+    mov fs:[08h], eax
+    mov esp, dword ptr [esp + 4]
+    mov fs:[04h], esp
+    jmp edx
+@rust_psm_replace_stack@16 ENDP
+
+; extern "fastcall" fn(%ecx: usize, %edx: usize, 4(%esp): extern "fastcall" fn(usize, usize),
+;                      8(%esp): *mut u8, 12(%esp): *mut u8)
+;
+; NB: on Windows for SEH to work at all, the pointers in TIB, thread information block, need to be
+; fixed up. Otherwise, it seems that exception mechanism on Windows will not bother looking for
+; exception handlers at *all* if they happen to fall outside the are specified in TIB.
+;
+; This necessitates an API difference from the usual 4-argument signature used elsewhere.
+@rust_psm_on_stack@20 PROC SYSCALL
+    push ebp
+    mov ebp, esp
+
+    push fs:[0E0Ch]
+    push fs:[08h]
+    mov eax, dword ptr [ebp + 4 + 12]
+    mov dword ptr fs:[08h], eax
+    mov dword ptr fs:[0E0Ch], eax
+    push fs:[04h]
+    mov esp, dword ptr [ebp + 4 + 8]
+    mov dword ptr fs:[04h], esp
+    call dword ptr [ebp + 4 + 4]
+
+    lea esp, [ebp - 12]
+    pop fs:[04h]
+    pop fs:[08h]
+    pop fs:[0E0Ch]
+    pop ebp
+    ret 12
+@rust_psm_on_stack@20 ENDP
+
+END

vendor/psm/src/arch/x86_windows_gnu.s

@@ -0,0 +1,94 @@
+/* FIXME: this works locally but not on appveyor??!? */
+/* NOTE: fastcall calling convention used on all x86 targets */
+.text
+
+.def @rust_psm_stack_direction@0
+.scl 2
+.type 32
+.endef
+.globl @rust_psm_stack_direction@0
+.p2align 4
+@rust_psm_stack_direction@0:
+/* extern "fastcall" fn() -> u8 (%al) */
+.cfi_startproc
+    movb $2, %al # always descending on x86_64
+    retl
+.cfi_endproc
+
+
+.def @rust_psm_stack_pointer@0
+.scl 2
+.type 32
+.endef
+.globl @rust_psm_stack_pointer@0
+.p2align 4
+@rust_psm_stack_pointer@0:
+/* extern "fastcall" fn() -> *mut u8 (%rax) */
+.cfi_startproc
+    leal 4(%esp), %eax
+    retl
+.cfi_endproc
+
+
+.def @rust_psm_replace_stack@16
+.scl 2
+.type 32
+.endef
+.globl @rust_psm_replace_stack@16
+.p2align 4
+@rust_psm_replace_stack@16:
+/* extern "fastcall" fn(%ecx: usize, %edx: extern "fastcall" fn(usize), 4(%esp): *mut u8) */
+.cfi_startproc
+/*
+   All we gotta do is set the stack pointer to 4(%esp) & tail-call the callback in %edx
+
+   Note, that the callee expects the stack to be offset by 4 bytes (normally, a return address
+   would be store there) off the required stack alignment on entry. To offset the stack in such a
+   way we use the `calll` instruction, however it would also be possible to to use plain `jmpl` but
+   would require to adjust the stack manually, which cannot be easily done, because the stack
+   pointer argument is already stored in memory.
+ */
+    movl 8(%esp), %eax
+    mov %eax, %fs:0x08
+    movl 4(%esp), %esp
+    mov %esp, %fs:0x04
+    calll *%edx
+    ud2
+.cfi_endproc
+
+
+.def @rust_psm_on_stack@16
+.scl 2
+.type 32
+.endef
+.globl @rust_psm_on_stack@16
+.p2align 4
+@rust_psm_on_stack@16:
+/* extern "fastcall" fn(%ecx: usize, %edx: usize, 4(%esp): extern "fastcall" fn(usize, usize), 8(%esp): *mut u8) */
+.cfi_startproc
+    pushl %ebp
+    .cfi_def_cfa %esp, 8
+    .cfi_offset %ebp, -8
+    pushl %fs:0x04
+    .cfi_def_cfa %esp, 12
+    pushl %fs:0x08
+    .cfi_def_cfa %esp, 16
+    movl  %esp, %ebp
+    .cfi_def_cfa_register %ebp
+
+    movl 24(%ebp), %eax
+    movl %eax, %fs:0x08
+    movl  20(%ebp), %esp
+    movl %esp, %fs:0x04
+    calll *16(%ebp)
+
+    movl  %ebp, %esp
+    popl %fs:0x08
+    .cfi_def_cfa %esp, 12
+    popl %fs:0x04
+    .cfi_def_cfa %esp, 8
+    popl  %ebp
+    .cfi_def_cfa %esp, 4
+    retl  $12
+.cfi_endproc
+

vendor/psm/src/arch/zseries_linux.s

@@ -0,0 +1,76 @@
+/* Implementation of stack swtiching routines for zSeries LINUX ABI.
+
+   This ABI is used by the s390x-unknown-linux-gnu target.
+
+   Documents used:
+
+   * LINUX for zSeries: ELF Application Binary Interface Supplement (1st ed., 2001) (LNUX-1107-01)
+   * z/Architecture: Principles of Operation (4th ed., 2004) (SA22-7832-03)
+*/
+
+#include "psm.h"
+#include "gnu_stack_note.s"
+
+.text
+
+
+.globl rust_psm_stack_direction
+.p2align 4
+.type rust_psm_stack_direction,@function
+rust_psm_stack_direction:
+/* extern "C" fn() -> u8 */
+.cfi_startproc
+    lghi %r2, STACK_DIRECTION_DESCENDING
+    br %r14
+.rust_psm_stack_direction_end:
+.size       rust_psm_stack_direction,.rust_psm_stack_direction_end-rust_psm_stack_direction
+.cfi_endproc
+
+
+.globl rust_psm_stack_pointer
+.p2align 4
+.type rust_psm_stack_pointer,@function
+rust_psm_stack_pointer:
+/* extern "C" fn() -> *mut u8 */
+.cfi_startproc
+    la %r2, 0(%r15)
+    br %r14
+.rust_psm_stack_pointer_end:
+.size       rust_psm_stack_pointer,.rust_psm_stack_pointer_end-rust_psm_stack_pointer
+.cfi_endproc
+
+
+.globl rust_psm_replace_stack
+.p2align 4
+.type rust_psm_replace_stack,@function
+rust_psm_replace_stack:
+/* extern "C" fn(r2: usize, r3: extern "C" fn(usize), r4: *mut u8) */
+.cfi_startproc
+    /* FIXME: backtrace does not terminate cleanly for some reason */
+    lay %r15, -160(%r4)
+    /* FIXME: this is `basr` instead of `br` purely to remove the backtrace link to the caller */
+    basr %r14, %r3
+.rust_psm_replace_stack_end:
+.size       rust_psm_replace_stack,.rust_psm_replace_stack_end-rust_psm_replace_stack
+.cfi_endproc
+
+
+.globl rust_psm_on_stack
+.p2align 4
+.type rust_psm_on_stack,@function
+rust_psm_on_stack:
+/* extern "C" fn(r2: usize, r3: usize, r4: extern "C" fn(usize, usize), r5: *mut u8) */
+.cfi_startproc
+    stmg %r14, %r15, -16(%r5)
+    lay %r15, -176(%r5)
+    .cfi_def_cfa %r15, 176
+    .cfi_offset %r14, -16
+    .cfi_offset %r15, -8
+    basr %r14, %r4
+    lmg %r14, %r15, 160(%r15)
+    .cfi_restore %r14
+    .cfi_restore %r15
+    br %r14
+.rust_psm_on_stack_end:
+.size       rust_psm_on_stack,.rust_psm_on_stack_end-rust_psm_on_stack
+.cfi_endproc