Crash calling back into a C function on a Mac M1

Code Generation AArch64
1

Hi all,
I’m trying to call back into a C function from some code JITd by LLVM, and it’s crashing with EXC_BAD_ACCESS.
This is on a Mac M1. LLVM version is 15.0.7.

The C function being called:

static void testExternalFuncVec2(float* ret, float x)
{
	ret[0] = x;
	ret[1] = x*x;
}

And the disassembly of the C function: (from Xcode)

winter`Winter::testExternalFuncVec2:
    0x10010e474 <+0>:  sub    sp, sp, #0x10
    0x10010e478 <+4>:  str    x0, [sp, #0x8]
    0x10010e47c <+8>:  str    s0, [sp, #0x4]
    0x10010e480 <+12>: ldr    s0, [sp, #0x4]
    0x10010e484 <+16>: ldr    x8, [sp, #0x8]
->  0x10010e488 <+20>: str    s0, [x8]
    0x10010e48c <+24>: ldr    s0, [sp, #0x4]
    0x10010e490 <+28>: ldr    s1, [sp, #0x4]
    0x10010e494 <+32>: fmul   s0, s0, s1
    0x10010e498 <+36>: ldr    x8, [sp, #0x8]
    0x10010e49c <+40>: str    s0, [x8, #0x4]
    0x10010e4a0 <+44>: add    sp, sp, #0x10
    0x10010e4a4 <+48>: ret

The disassembled LLVM JIT’d code, that calls the C function:

	.text
	.file	"WinterModule"
	.globl	main_float_
	.p2align	2
	.type	main_float_,@function
main_float_:
	sub	sp, sp, #48
	stp	d9, d8, [sp, #16]
	str	x30, [sp, #32]
	fmov	s8, s0
	add	x8, sp, #8
	bl	testExternalFuncVec2_float_
	ldr	s9, [sp, #8]
	add	x8, sp, #40
	fmov	s0, s8
	bl	testExternalFuncVec2_float_
	ldr	s0, [sp, #40]
	fadd	s0, s9, s0
	ldr	x30, [sp, #32]
	ldp	d9, d8, [sp, #16]
	add	sp, sp, #48
	ret
.Lfunc_end0:
	.size	main_float_, .Lfunc_end0-main_float_

	.section	".note.GNU-stack","",@progbits

And the corresponding LLVM IR:

; ModuleID = 'WinterModule'
source_filename = "WinterModule"
target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"

%vec2 = type { <2 x float> }

; Function Attrs: nounwind
declare void @testExternalFuncVec2_float_(ptr noalias sret(%vec2), float) local_unnamed_addr #0

; Function Attrs: nofree nosync nounwind readnone
define float @main_float_(float %x) local_unnamed_addr #1 {
entry:
  %res1 = alloca %vec2, align 8
  %res = alloca %vec2, align 8
  call void @testExternalFuncVec2_float_(ptr nonnull %res, float %x)
  %0 = load float, ptr %res, align 8
  call void @testExternalFuncVec2_float_(ptr nonnull %res1, float %x)
  %1 = load float, ptr %res1, align 8
  %2 = fadd nnan ninf nsz float %0, %1
  ret float %2
}

attributes #0 = { nounwind "warn-stack-size"="0" }
attributes #1 = { nofree nosync nounwind readnone "warn-stack-size"="0" }

Is this an issue of LLVM not using the correct calling convention on M1 Mac?
This code works fine on Windows, Linux and Mac x64.

2

Why have you marked it sret when the C function takes it as a normal pointer? As you can see here:

the ABI is to pass the sret (struct return) pointer via x8, not as if it were a normal argument, and this is true of both Apple’s and the standard Arm PCS.

3

Hi,
Removing SRET seems to solve the crash, thanks.
I was using it because conceptually the function result is returned via the first argument.

4

For that IR to be accurate you’d need to have C of the form:

struct { float f[2]; } testExternalFuncVec2(float x)
{
    ret[0] = x;
    ret[1] = x*x;
}

(or some kind of vector type)