Mercurial > pub > dyncall > dyncall
annotate dyncall/dyncall_call_arm32_arm_armhf.S @ 407:7364f285cac8
- use UAL syntax for some arm instructions for armhf targets, as clang's integrated as disallows pre-UAL syntax
| author | Tassilo Philipp |
|---|---|
| date | Sun, 03 Oct 2021 10:41:46 +0200 |
| parents | 351bb41d3bb1 |
| children |
| rev | line source |
|---|---|
| 0 | 1 /* |
| 2 | |
| 3 Package: dyncall | |
| 4 Library: dyncall | |
| 5 File: dyncall/dyncall_call_arm32_arm_armhf.S | |
| 6 Description: Call Kernel for ARM 32-bit ARM Architecture - Hard Float | |
| 7 License: | |
| 8 | |
| 281 | 9 Copyright (c) 2007-2018 Daniel Adler <dadler@uni-goettingen.de>, |
| 0 | 10 Tassilo Philipp <tphilipp@potion-studios.com> |
| 11 | |
| 12 Permission to use, copy, modify, and distribute this software for any | |
| 13 purpose with or without fee is hereby granted, provided that the above | |
| 14 copyright notice and this permission notice appear in all copies. | |
| 15 | |
| 16 THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES | |
| 17 WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF | |
| 18 MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR | |
| 19 ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES | |
| 20 WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN | |
| 21 ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF | |
| 22 OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE. | |
| 23 | |
| 24 */ | |
| 25 | |
| 26 | |
| 27 | |
| 28 #include "../portasm/portasm-arm.S" | |
| 29 | |
| 30 /* ============================================================================ | |
| 31 DynCall Call Kernel for ARM 32-bit ARM Architecture Hard-Float | |
| 32 ---------------------------------------------------------------------------- | |
| 33 C Interface: | |
| 34 dcCall_arm32_armhf (DCpointer target, DCpointer argv, DCsize size, DCfloat* regdata); | |
| 35 | |
| 36 This Call Kernel was tested on Raspberry Pi/Raspbian (Debian) | |
| 37 */ | |
| 38 | |
| 39 .text | |
| 40 .code 32 /* ARM mode */ | |
| 41 .arch armv6 | |
| 42 .fpu vfp | |
| 43 | |
| 36 | 44 /* |
| 45 1st arg / r0 = funptr | |
| 46 2st arg / r1 = ptr to int args | |
| 47 3st arg / r2 = size | |
| 48 4st arg / r3 = ptr to float args | |
| 49 */ | |
| 0 | 50 GLOBAL_C(dcCall_arm32_armhf) |
| 51 ENTRY_C(dcCall_arm32_armhf) | |
| 52 | |
| 53 /* Prolog. This function never needs to spill inside its prolog, so just store the permanent registers. */ | |
|
148
12729fd52ab7
- performance improvement and more correct handling of arm32 armhf calls, not copying bogus data just to reserve spill area space
cslag
parents:
146
diff
changeset
|
54 mov r12 , r13 /* Stack ptr (r13) -> temporary (r12). */ |
|
12729fd52ab7
- performance improvement and more correct handling of arm32 armhf calls, not copying bogus data just to reserve spill area space
cslag
parents:
146
diff
changeset
|
55 stmdb r13!, {r4-r5, r11, r12, r14} /* Permanent registers and stack pointer (now in r12), etc... -> save area on stack (except counter). */ |
| 0 | 56 |
|
148
12729fd52ab7
- performance improvement and more correct handling of arm32 armhf calls, not copying bogus data just to reserve spill area space
cslag
parents:
146
diff
changeset
|
57 mov r11, r12 /* Set frame ptr. */ |
|
12729fd52ab7
- performance improvement and more correct handling of arm32 armhf calls, not copying bogus data just to reserve spill area space
cslag
parents:
146
diff
changeset
|
58 mov r4, r0 /* r4 = fptr */ |
|
12729fd52ab7
- performance improvement and more correct handling of arm32 armhf calls, not copying bogus data just to reserve spill area space
cslag
parents:
146
diff
changeset
|
59 add r5, r1, #16 /* r5 = stack args (after intreg ones) */ |
|
12729fd52ab7
- performance improvement and more correct handling of arm32 armhf calls, not copying bogus data just to reserve spill area space
cslag
parents:
146
diff
changeset
|
60 |
| 0 | 61 /* Load 16 single-precision registers (= 8 double-precision registers). */ |
|
407
7364f285cac8
- use UAL syntax for some arm instructions for armhf targets, as clang's integrated as disallows pre-UAL syntax
Tassilo Philipp
parents:
406
diff
changeset
|
62 vldmia r3, {d0-d7} |
| 0 | 63 |
|
148
12729fd52ab7
- performance improvement and more correct handling of arm32 armhf calls, not copying bogus data just to reserve spill area space
cslag
parents:
146
diff
changeset
|
64 /* prep stack parameter area (includes room for spill area, callee spills if needed) */ |
|
12729fd52ab7
- performance improvement and more correct handling of arm32 armhf calls, not copying bogus data just to reserve spill area space
cslag
parents:
146
diff
changeset
|
65 sub r13, r13, r2 |
|
12729fd52ab7
- performance improvement and more correct handling of arm32 armhf calls, not copying bogus data just to reserve spill area space
cslag
parents:
146
diff
changeset
|
66 and r13, r13, #-8 /* align 8-byte. */ |
| 0 | 67 |
|
148
12729fd52ab7
- performance improvement and more correct handling of arm32 armhf calls, not copying bogus data just to reserve spill area space
cslag
parents:
146
diff
changeset
|
68 sub r2, r2, #16 /* loop counters for stack params to copy */ |
|
12729fd52ab7
- performance improvement and more correct handling of arm32 armhf calls, not copying bogus data just to reserve spill area space
cslag
parents:
146
diff
changeset
|
69 mov r3, #0 |
| 0 | 70 armhf_pushArgs: |
|
148
12729fd52ab7
- performance improvement and more correct handling of arm32 armhf calls, not copying bogus data just to reserve spill area space
cslag
parents:
146
diff
changeset
|
71 cmp r2, r3 |
|
12729fd52ab7
- performance improvement and more correct handling of arm32 armhf calls, not copying bogus data just to reserve spill area space
cslag
parents:
146
diff
changeset
|
72 ble armhf_call |
|
12729fd52ab7
- performance improvement and more correct handling of arm32 armhf calls, not copying bogus data just to reserve spill area space
cslag
parents:
146
diff
changeset
|
73 ldr r0, [r5, +r3] /* load word into r0 ... */ |
|
12729fd52ab7
- performance improvement and more correct handling of arm32 armhf calls, not copying bogus data just to reserve spill area space
cslag
parents:
146
diff
changeset
|
74 str r0, [r13, +r3] /* ... then push onto stack */ |
|
12729fd52ab7
- performance improvement and more correct handling of arm32 armhf calls, not copying bogus data just to reserve spill area space
cslag
parents:
146
diff
changeset
|
75 add r3, r3, #4 |
|
12729fd52ab7
- performance improvement and more correct handling of arm32 armhf calls, not copying bogus data just to reserve spill area space
cslag
parents:
146
diff
changeset
|
76 b armhf_pushArgs |
| 0 | 77 |
| 78 armhf_call: | |
|
148
12729fd52ab7
- performance improvement and more correct handling of arm32 armhf calls, not copying bogus data just to reserve spill area space
cslag
parents:
146
diff
changeset
|
79 ldmia r1, {r0-r3} /* Load first 4 arguments for new call into r0-r3. */ |
|
406
351bb41d3bb1
- removed %-prefixes for register names from arm assembly files (was wrong to begin with, gas accepted them but the clang integrated assembler does not)
Tassilo Philipp
parents:
281
diff
changeset
|
80 /* 'blx r4' workaround for ARMv4t: */ |
|
148
12729fd52ab7
- performance improvement and more correct handling of arm32 armhf calls, not copying bogus data just to reserve spill area space
cslag
parents:
146
diff
changeset
|
81 mov r14, r15 /* Branch return address(r15) -> link register (r14) -- r15 always points to address of current + 2 instructions (= Epilog code). */ |
|
12729fd52ab7
- performance improvement and more correct handling of arm32 armhf calls, not copying bogus data just to reserve spill area space
cslag
parents:
146
diff
changeset
|
82 bx r4 /* Call (ARM/THUMB), available for ARMv4t. */ |
| 0 | 83 |
| 84 /* Epilog. */ | |
|
148
12729fd52ab7
- performance improvement and more correct handling of arm32 armhf calls, not copying bogus data just to reserve spill area space
cslag
parents:
146
diff
changeset
|
85 ldmdb r11, {r4-r5, r11, r13, r15} /* Restore permanent registers (ignore temporary (r12), restore stack ptr and program counter).@@@db not needed since we rewrite r13? */ |
| 0 | 86 |