parameter passing pointer and __m256

parameter passing pointer and __m256.cpp

void test_parameterPassing_pointer( double *d1p, 
                                    __m256 *m256_1p, __m128 *m128_1p, 
                                    __m256 m256_1) {

	double mul1 = (*d1p) + (*d1p);
	
	double *m256_dp = ((double*)m256_1p);
	double *m128_dp = ((double*)m128_1p);
	
	double mul2 = m256_dp[0] * mul1 - m128_dp[1];
	double div1 = ((double*)&m256_1)[2] / mul2;
	
	printf("mul2=%f\n", div1);
}


In OSX/Linux x64 and using gcc, all parameters are passed using registers.
Pointer parameters are treated like normal parameters and hence passed using regular %rdi, %rsi, %rdx, %rcx, %r8, %r9
__m256, which is 32bytes in size, is passed through %ymm register, as can be seen below:

_test_parameterPassing_pointer:         ## @test_parameterPassing_pointer
	.cfi_startproc
## BB#0:
	pushq	%rbp
Ltmp24:
	.cfi_def_cfa_offset 16
Ltmp25:
	.cfi_offset %rbp, -16
	movq	%rsp, %rbp
Ltmp26:
	.cfi_def_cfa_register %rbp
	vmovsd	(%rdi), %xmm1
	vaddsd	%xmm1, %xmm1, %xmm1
	vmulsd	(%rsi), %xmm1, %xmm1
	vsubsd	8(%rdx), %xmm1, %xmm1
	vextractf128	$1, %ymm0, %xmm0	# <<---- our __m256 param
	//xmm0 == ymm0[128:255]
	
	vpshufd	$1, %xmm0, %xmm2        ## xmm2 = xmm0[1,0,0,0]
	
	vinsertps	$16, %xmm2, %xmm0, %xmm0 ## xmm0 = xmm0[0],xmm2[0],xmm0[2,3]
	//$16 == 0001 0000
	//source 00 = 0, %xmm2[31:0]
	//dest	 01 = 1, %xmm0[63:32]
	//mask 0000 
	//xmm0 == xmm0[127:96], xmm0[95:64], xmm2[31:0], xmm0[31:0]
	
	vdivsd	%xmm1, %xmm0, %xmm0
	leaq	L_.str5(%rip), %rdi
	movb	$1, %al
	popq	%rbp
	vzeroupper
	jmp	_printf                 ## TAILCALL
	



In Windows x64, the standard __fastcall applies, which means only up to 4 parameters are passed using registers.
Our last param, m256_1 above, is passed using reference, not value (because it is more than 64 bits long)
test_parameterPassing_pointer	PROC 
; parameter 1(d1p): rcx
; parameter 2(m256_1p): rdx
; parameter 3(m128_1p): r8
; parameter 4(m256_1): [r9]
.B3.1::                         ; Preds .B3.0

;;; void test_parameterPassing_pointer(double *d1p, __m256 *m256_1p, __m128 *m128_1p, __m256 m256_1) {

$LN290:
  00000 48 83 ec 28      sub rsp, 40                            ; main.c:148.98
$LN291:

;;; 
;;; 	double mul1 = (*d1p) + (*d1p);

  00004 f2 0f 10 01      movsd xmm0, QWORD PTR [rcx]            ; main.c:150.18
$LN292:

;;; 	double *m256_dp = ((double*)m256_1p);
;;; 	double *m128_dp = ((double*)m128_1p);
;;; 	double mul2 = m256_dp[0] * mul1 - m128_dp[1];
;;; 	double div1 = ((double*)&m256_1)[2] / mul2;
;;; 	printf("mul2=%f\n", div1);

  00008 48 8d 0d 00 00 
        00 00            lea rcx, QWORD PTR [??_C@_08A@mul2?$DN?$CFf?6?$AA@] ; main.c:155.2
$LN293:
  0000f f2 41 0f 10 49 
        10               movsd xmm1, QWORD PTR [16+r9]          ; main.c:155.2 <<----- our __m256 param
$LN294:
  00015 f2 0f 58 c0      addsd xmm0, xmm0                       ; main.c:155.2
$LN295:
  00019 f2 0f 59 02      mulsd xmm0, QWORD PTR [rdx]            ; main.c:155.2
$LN296:
  0001d 4c 89 4c 24 48   mov QWORD PTR [72+rsp], r9             ; main.c:148.98  
$LN297:
  00022 f2 41 0f 5c 40 
        08               subsd xmm0, QWORD PTR [8+r8]           ; main.c:155.2
$LN298:
  00028 f2 0f 5e c8      divsd xmm1, xmm0                       ; main.c:155.2
$LN299:
  0002c 66 48 0f 7e ca   movd rdx, xmm1                         ; main.c:155.2
$LN300:
  00031 ff 15 fc ff ff 
        ff               call QWORD PTR [__imp_printf]          ; main.c:155.2
$LN301:
                                ; LOE rbx rbp rsi rdi r12 r13 r14 r15 xmm6 xmm7 xmm8 xmm9 xmm10 xmm11 xmm12 xmm13 xmm14 xmm15
.B3.2::                         ; Preds .B3.1

;;; }