sum_sse4_amd64.s

//+build !noasm !appengine
// AUTO-GENERATED BY C2GOASM -- DO NOT EDIT

TEXT ·_sum_float64_sse4(SB), $0-24

	MOVQ buf+0(FP), DI
	MOVQ len+8(FP), SI
	MOVQ res+16(FP), DX

	WORD $0xf685             // test    esi, esi
	JLE  LBB0_1
	WORD $0xf089             // mov    eax, esi
	LONG $0xc0570f66         // xorpd    xmm0, xmm0
	WORD $0xfe83; BYTE $0x03 // cmp    esi, 3
	JBE  LBB0_3
	WORD $0xe683; BYTE $0x03 // and    esi, 3
	WORD $0x8949; BYTE $0xc0 // mov    r8, rax
	WORD $0x2949; BYTE $0xf0 // sub    r8, rsi
	JE   LBB0_3
	LONG $0x104f8d48         // lea    rcx, [rdi + 16]
	LONG $0xc9570f66         // xorpd    xmm1, xmm1
	WORD $0x894d; BYTE $0xc1 // mov    r9, r8
	LONG $0xc0570f66         // xorpd    xmm0, xmm0

LBB0_8:
	LONG $0x51100f66; BYTE $0xf0 // movupd    xmm2, oword [rcx - 16]
	LONG $0x19100f66             // movupd    xmm3, oword [rcx]
	LONG $0xca580f66             // addpd    xmm1, xmm2
	LONG $0xc3580f66             // addpd    xmm0, xmm3
	LONG $0x20c18348             // add    rcx, 32
	LONG $0xfcc18349             // add    r9, -4
	JNE  LBB0_8
	LONG $0xc1580f66             // addpd    xmm0, xmm1
	LONG $0xc07c0f66             // haddpd    xmm0, xmm0
	WORD $0xf685                 // test    esi, esi
	JNE  LBB0_4
	JMP  LBB0_10

LBB0_3:
	WORD $0x3145; BYTE $0xc0 // xor    r8d, r8d

LBB0_4:
	LONG $0xc70c8d4a         // lea    rcx, [rdi + 8*r8]
	WORD $0x294c; BYTE $0xc0 // sub    rax, r8

LBB0_5:
	LONG $0x01580ff2         // addsd    xmm0, qword [rcx]
	LONG $0x08c18348         // add    rcx, 8
	WORD $0xff48; BYTE $0xc8 // dec    rax
	JNE  LBB0_5
	JMP  LBB0_10

LBB0_1:
	LONG $0xc0570f66 // xorpd    xmm0, xmm0

LBB0_10:
	LONG $0x02110ff2 // movsd    qword [rdx], xmm0
	RET