x64 from https://stackoverflow.com/questions/77116639/is-there-any-faster-way-to-do-this-optimization-or-just-straight-up-doing-it-a

gistfile1.txt

# Main

L0000	push	rdi
L0001	push	rsi
L0002	push	rbp
L0003	push	rbx
L0004	sub	rsp, 0x58
L0008	vzeroupper	
L000b	xor	eax, eax
L000d	mov	[rsp+0x50], rax
L0012	mov	[rsp+0x48], rax
L0017	call	UserQuery.CreateSampleData()
L001c	mov	[rsp+0x30], rax
L0021	mov	rcx, 0x7ffe4a5824e8
L002b	call	0x00007ffea9bd09c0
L0030	mov	rdi, rax
L0033	xor	ecx, ecx
L0035	mov	[rdi+8], rcx
L0039	mov	byte ptr [rdi+0x18], 0
L003d	mov	[rdi+0x10], rcx
L0041	mov	[rsp+0x20], rdi
L0046	cmp	byte ptr [rdi+0x18], 0
L004a	jne	L017f
L0050	lea	rcx, [rsp+0x50]
L0055	mov	rax, 0x7fff5a0061f0
L005f	call	rax
L0061	mov	rcx, [rsp+0x50]
L0066	mov	rdi, [rsp+0x20]
L006b	mov	[rdi+0x10], rcx
L006f	mov	byte ptr [rdi+0x18], 1
L0073	cmp	dword ptr [0x7ffea9f12e74], 0
L007a	jne	L0189
L0080	mov	rsi, [rsp+0x30]
L0085	mov	rdi, [rsp+0x20]
L008a	mov	[rsp+0x38], rdi
L008f	mov	rcx, rsi
L0092	call	UserQuery.Flatten(Char[,])
L0097	mov	[rsp+0x28], rax
L009c	mov	rcx, rdi
L009f	call	qword ptr [0x7ffe4a591318]
L00a5	mov	rbp, [rdi+8]
L00a9	cmp	byte ptr [rdi+0x18], 0
L00ad	je	L0175
L00b3	lea	rcx, [rsp+0x48]
L00b8	mov	rax, 0x7fff5a0061f0
L00c2	call	rax
L00c4	mov	rcx, [rsp+0x48]
L00c9	mov	rdi, [rsp+0x38]
L00ce	sub	rcx, [rdi+0x10]
L00d2	add	rbp, rcx
L00d5	cmp	dword ptr [0x7ffea9f12e74], 0
L00dc	jne	L0198
L00e2	mov	rbx, [rsp+0x28]
L00e7	vxorps	xmm0, xmm0, xmm0
L00eb	vcvtsi2sd	xmm0, xmm0, rbp
L00f0	vcvttsd2si	rcx, xmm0
L00f5	vxorps	xmm0, xmm0, xmm0
L00f9	vcvtsi2sd	xmm0, xmm0, rcx
L00fe	vdivsd	xmm0, xmm0, [UserQuery.Main()]
L0106	vucomisd	xmm0, [UserQuery.Main()]
L010e	ja	L01a7
L0114	vmovsd	xmm1, [UserQuery.Main()]
L011c	vucomisd	xmm1, xmm0
L0120	ja	L01ba
L0126	vmovsd	[rsp+0x40], xmm0
L012c	mov	rcx, 0x7ffe4a0c77b8
L0136	call	0x00007ffea9bd09c0
L013b	vmovsd	xmm0, [rsp+0x40]
L0141	vmovsd	[rax+8], xmm0
L0146	mov	rdx, rax
L0149	mov	rcx, 0x1d6d9012578
L0153	mov	rcx, [rcx]
L0156	call	qword ptr [0x7ffe4a7f1a08]
L015c	call	qword ptr [0x7ffe4a7f18b8]
L0162	mov	rcx, rbx
L0165	call	qword ptr [0x7ffe4a7f19f0]
L016b	nop	
L016c	add	rsp, 0x58
L0170	pop	rbx
L0171	pop	rbp
L0172	pop	rsi
L0173	pop	rdi
L0174	ret	
L0175	mov	rbx, [rsp+0x28]
L017a	jmp	L00e7
L017f	mov	rsi, [rsp+0x30]
L0184	jmp	L0085
L0189	call	0x00007ffea9bcd900
L018e	mov	rsi, [rsp+0x30]
L0193	jmp	L0085
L0198	call	0x00007ffea9bcd900
L019d	mov	rbx, [rsp+0x28]
L01a2	jmp	L00e7
L01a7	vmovsd	xmm0, [UserQuery.Main()]
L01af	vmovsd	[rsp+0x40], xmm0
L01b5	jmp	L012c
L01ba	vmovsd	xmm0, [UserQuery.Main()]
L01c2	vmovsd	[rsp+0x40], xmm0
L01c8	jmp	L012c


# Flatten

L0000	push	r15
L0002	push	r14
L0004	push	rdi
L0005	push	rsi
L0006	push	rbp
L0007	push	rbx
L0008	sub	rsp, 0x28
L000c	mov	rsi, rcx
L000f	test	rsi, rsi
L0012	je	L00d3
L0018	mov	edi, [rsi+0x10]
L001b	mov	ebx, [rsi+0x14]
L001e	mov	rcx, 0x7ffe4a24a4c0
L0028	call	0x00007ffea9bd09c0
L002d	mov	rbp, rax
L0030	mov	edx, edi
L0032	imul	edx, ebx
L0035	mov	rcx, rbp
L0038	mov	r8d, 0x7fffffff
L003e	call	qword ptr [0x7ffe4a2530f0]
L0044	xor	r14d, r14d
L0047	test	edi, edi
L0049	jle	short L00b9
L004b	xor	r15d, r15d
L004e	test	ebx, ebx
L0050	jle	short L00b1
L0052	mov	ecx, r14d
L0055	sub	ecx, [rsi+0x18]
L0058	cmp	ecx, [rsi+0x10]
L005b	jae	short L00ce
L005d	mov	edx, [rsi+0x14]
L0060	imul	ecx, edx
L0063	mov	r8d, r15d
L0066	sub	r8d, [rsi+0x1c]
L006a	cmp	r8d, edx
L006d	jae	short L00ce
L006f	add	ecx, r8d
L0072	movzx	edx, word ptr [rsi+rcx*2+0x20]
L0077	mov	ecx, [rbp+0x18]
L007a	mov	r8d, ecx
L007d	mov	rax, [rbp+8]
L0081	mov	r9d, [rax+8]
L0085	cmp	r9d, r8d
L0088	jbe	short L009a
L008a	mov	r8d, r8d
L008d	mov	[rax+r8*2+0x10], dx
L0093	inc	ecx
L0095	mov	[rbp+0x18], ecx
L0098	jmp	short L00a9
L009a	mov	rcx, rbp
L009d	mov	r8d, 1
L00a3	call	qword ptr [0x7ffe4a253258]
L00a9	inc	r15d
L00ac	cmp	r15d, ebx
L00af	jl	short L0052
L00b1	inc	r14d
L00b4	cmp	r14d, edi
L00b7	jl	short L004b
L00b9	mov	rcx, rbp
L00bc	add	rsp, 0x28
L00c0	pop	rbx
L00c1	pop	rbp
L00c2	pop	rsi
L00c3	pop	rdi
L00c4	pop	r14
L00c6	pop	r15
L00c8	jmp	qword ptr [0x7ffe4a24a528]
L00ce	call	0x00007ffea9cf9b30
L00d3	mov	rcx, 0x7ffe4a1fe830
L00dd	call	0x00007ffea9bd09c0
L00e2	mov	rsi, rax
L00e5	mov	ecx, 0x65
L00ea	mov	rdx, 0x7ffe4aacc060
L00f4	call	0x00007ffea9b52360
L00f9	mov	rdx, rax
L00fc	mov	rcx, rsi
L00ff	call	qword ptr [0x7ffe4a211348]
L0105	mov	rcx, rsi
L0108	call	0x00007ffea9b28570
L010d	int3	

# CreateSampleData

L0000	push	rdi
L0001	push	rsi
L0002	push	rbp
L0003	push	rbx
L0004	sub	rsp, 0x48
L0008	vzeroupper	
L000b	vxorps	xmm4, xmm4, xmm4
L000f	vmovdqa	[rsp+0x20], xmm4
L0015	vmovdqa	[rsp+0x30], xmm4
L001b	xor	eax, eax
L001d	mov	[rsp+0x40], rax
L0022	mov	rcx, 0x7ffe4a7c1ca0
L002c	call	0x00007ffea9bd09c0
L0031	mov	rsi, rax
L0034	mov	rcx, 0x7ffe4a8d8650
L003e	call	0x00007ffea9bd09c0
L0043	mov	rdi, rax
L0046	vxorps	xmm0, xmm0, xmm0
L004a	vmovdqu	[rsp+0x20], xmm0
L0050	lea	rcx, [rsp+0x20]
L0055	mov	edx, 0x34d0058
L005a	call	qword ptr [0x7ffe4a8e7000]
L0060	lea	rbx, [rdi+8]
L0064	mov	rdx, [rsp+0x20]
L0069	mov	rcx, rbx
L006c	call	0x00007ffea9bd0440
L0071	mov	edx, [rsp+0x28]
L0075	mov	[rbx+8], edx
L0078	mov	edx, [rsp+0x2c]
L007c	mov	[rbx+0xc], edx
L007f	mov	rdx, rdi
L0082	lea	rcx, [rsi+8]
L0086	call	0x00007ffe49f10010
L008b	mov	dword ptr [rsp+0x40], 0x100
L0093	mov	dword ptr [rsp+0x44], 0x100
L009b	lea	r8, [rsp+0x40]
L00a0	mov	rcx, 0x7ffe4a7c1e08
L00aa	mov	edx, 2
L00af	call	0x00007ffea9bb6180
L00b4	mov	rdi, rax
L00b7	xor	ebx, ebx
L00b9	xor	ebp, ebp
L00bb	mov	rcx, [rsi+8]
L00bf	mov	edx, 0x4c
L00c4	mov	rax, [rcx]
L00c7	mov	rax, [rax+0x40]
L00cb	call	qword ptr [rax+0x30]
L00ce	add	eax, 0x32
L00d1	mov	edx, ebx
L00d3	sub	edx, [rdi+0x18]
L00d6	cmp	edx, [rdi+0x10]
L00d9	jae	short L0115
L00db	mov	ecx, [rdi+0x14]
L00de	imul	edx, ecx
L00e1	mov	r8d, ebp
L00e4	sub	r8d, [rdi+0x1c]
L00e8	cmp	r8d, ecx
L00eb	jae	short L0115
L00ed	add	edx, r8d
L00f0	mov	[rdi+rdx*2+0x20], ax
L00f5	inc	ebp
L00f7	cmp	ebp, 0x100
L00fd	jl	short L00bb
L00ff	inc	ebx
L0101	cmp	ebx, 0x100
L0107	jl	short L00b9
L0109	mov	rax, rdi
L010c	add	rsp, 0x48
L0110	pop	rbx
L0111	pop	rbp
L0112	pop	rsi
L0113	pop	rdi
L0114	ret	
L0115	call	0x00007ffea9cf9b30
L011a	int3