MULQ SI
	ADDQ (CX), AX
	ADCQ $0x00, DX
	ADDQ DI, AX
	ADCQ $0x00, DX
	MOVQ DX, DI
	MOVQ AX, (CX)

	// Iteration 1
	MOVQ 8(BX), AX
	MULQ SI
	ADDQ 8(CX), AX
	ADCQ $0x00, DX
	ADDQ DI, AX
	ADCQ $0x00, DX
	MOVQ DX, DI
	MOVQ AX, 8(CX)

	// Iteration 2
	MOVQ 16(BX), AX
	MULQ SI
	ADDQ 16(CX), AX
	ADCQ $0x00, DX
	ADDQ DI, AX
	ADCQ $0x00, DX
	MOVQ DX, DI
	MOVQ AX, 16(CX)

	// Iteration 3

	ADDQ AX, acc3
	ADCQ DX, acc4
	ADCQ $0, acc5
	// x * y[1]
	MOVQ (8*1)(y_ptr), t0

	MOVQ (8*0)(x_ptr), AX
	MULQ t0
	ADDQ AX, acc1
	ADCQ $0, DX
	MOVQ DX, t1

	MOVQ (8*1)(x_ptr), AX
	MULQ t0
	ADDQ t1, acc2
	ADCQ $0, DX
	ADDQ AX, acc2
	ADCQ $0, DX
	MOVQ DX, t1

	MOVQ (8*2)(x_ptr), AX
	MULQ t0
	ADDQ t1, acc3
	ADCQ $0, DX
	ADDQ AX, acc3
	ADCQ $0, DX
	MOVQ DX, t1

	ROUND2(AX, BX, CX, DX, BP, 30)
	ROUND2(BP, AX, BX, CX, DX, 31)
	ROUND2(DX, BP, AX, BX, CX, 32)
	ROUND2(CX, DX, BP, AX, BX, 33)
	ROUND2(BX, CX, DX, BP, AX, 34)
	ROUND2(AX, BX, CX, DX, BP, 35)
	ROUND2(BP, AX, BX, CX, DX, 36)
	ROUND2(DX, BP, AX, BX, CX, 37)
	ROUND2(CX, DX, BP, AX, BX, 38)
	ROUND2(BX, CX, DX, BP, AX, 39)

	ROUND3(AX, BX, CX, DX, BP, 40)
	ROUND3(BP, AX, BX, CX, DX, 41)

	MOVL	g(CX), BX	// BX = g
	MOVL	g_m(BX), DX	// DX = curm

	// set g to gcrash
	LEAL	runtime·gcrash(SB), BX // g = &gcrash
	MOVL	DX, g_m(BX)            // g.m = curm
	MOVL	BX, m_g0(DX)           // curm.g0 = g
	get_tls(CX)
	MOVL	BX, g(CX)

	// switch to crashstack
	MOVL	(g_stack+stack_hi)(BX), DX
	SUBL	$(4*8), DX
	MOVL	DX, SP

	// call target function
	MOVL	AX, DX
	MOVL	0(AX), AX
	CALL	AX

TEXT runtime·mmap_trampoline(SB),NOSPLIT,$0
	PUSHL	BP
	MOVL	SP, BP
	SUBL	$32, SP
	MOVL	40(SP), DX		// pointer to args
	MOVL	0(DX), AX
	MOVL	4(DX), BX
	MOVL	8(DX), CX
	MOVL	AX, 0(SP)		// arg 1 - addr
	MOVL	BX, 4(SP)		// arg 2 - len
	MOVL	CX, 8(SP)		// arg 3 - prot
	MOVL	12(DX), AX
	MOVL	16(DX), BX
	MOVL	20(DX), CX
	MOVL	AX, 12(SP)		// arg 4 - flags
	MOVL	BX, 16(SP)		// arg 5 - fid
	MOVL	$0, 20(SP)		// pad

	MOVQ	((index-2)*8)(BP), AX; \
	MOVQ	AX, CX; \
	RORQ	$19, AX; \
	MOVQ	CX, DX; \
	RORQ	$61, CX; \
	SHRQ	$6, DX; \
	MOVQ	((index-15)*8)(BP), BX; \
	XORQ	CX, AX; \
	MOVQ	BX, CX; \
	XORQ	DX, AX; \
	RORQ	$1, BX; \
	MOVQ	CX, DX; \
	SHRQ	$7, DX; \
	RORQ	$8, CX; \
	ADDQ	((index-7)*8)(BP), AX; \
	XORQ	CX, BX; \
	XORQ	DX, BX; \
	ADDQ	((index-16)*8)(BP), BX; \
	ADDQ	BX, AX; \
	MOVQ	AX, ((index)*8)(BP)

	VPMADD52HUQ 15(DX)(BX*2), Y31, K7, Y17             // 62e28527b58c5a0f000000
	VPMADD52HUQ Y28, Y8, K7, Y17                       // 6282bd2fb5cc
	VPMADD52HUQ Y13, Y8, K7, Y17                       // 62c2bd2fb5cd
	VPMADD52HUQ Y7, Y8, K7, Y17                        // 62e2bd2fb5cf
	VPMADD52HUQ (R8), Y8, K7, Y17                      // 62c2bd2fb508
	VPMADD52HUQ 15(DX)(BX*2), Y8, K7, Y17              // 62e2bd2fb58c5a0f000000

	// so this function is reentrant.
	MOVQ	m_vdsoPC(BX), CX
	MOVQ	m_vdsoSP(BX), DX
	MOVQ	CX, 0(SP)
	MOVQ	DX, 8(SP)

	LEAQ	ret+0(FP), DX
	MOVQ	-8(DX), CX
	MOVQ	CX, m_vdsoPC(BX)
	MOVQ	DX, m_vdsoSP(BX)

	CMPQ	R14, m_curg(BX)	// Only switch if on curg.
	JNE	noswitch

	MOVQ	m_g0(BX), DX
	MOVQ	(g_sched+gobuf_sp)(DX), SP	// Set SP to g0 stack

noswitch:
	SUBQ	$16, SP		// Space for results

  // Softmax gradient function.
  // p = softmax(x) maps from [batch, n] to [batch, m]
  // dp/dx = [dp0/dx0   ... dp0/dxn-1  ]
  //         [  ...           ...      ]
  //         [dpm-1/dx0 ... dpm-1/dxn-1]
  // dL/dx = dp/dx * dL/dy
  //
  // Using alternative formula:
  // dL/dx = dL/dy * y - sum(dL/dy * y) * y
  //    = (dL/dy - sum(dL/dy * y)) * y
  auto y = op.output(0);

	MOVQ	0(BX), DI		// arg 1 addr
	MOVQ	8(BX), SI		// arg 2 len
	MOVL	16(BX), DX		// arg 3 prot
	MOVL	20(BX), CX		// arg 4 flags
	MOVL	24(BX), R8		// arg 5 fid
	MOVL	28(BX), R9		// arg 6 offset
	CALL	libc_mmap(SB)
	XORL	DX, DX
	CMPQ	AX, $-1
	JNE	ok
	CALL	libc_error(SB)
	MOVLQSX	(AX), DX		// errno
	XORL	AX, AX
ok:
	MOVQ	AX, 32(BX)
	MOVQ	DX, 40(BX)
	RET

TEXT runtime·munmap_trampoline(SB),NOSPLIT,$0