1 files changed, 211 insertions, 0 deletions
diff --git a/vere/ext/gmp/gen/x86_64-macos/mpn/popcount.s b/vere/ext/gmp/gen/x86_64-macos/mpn/popcount.s
new file mode 100644
index 0000000..c7695bb
--- /dev/null
+++ b/vere/ext/gmp/gen/x86_64-macos/mpn/popcount.s
@@ -0,0 +1,211 @@
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+	.text
+	.align	5, 0x90
+	.globl	___gmpn_popcount
+	
+	
+___gmpn_popcount:
+
+	
+
+	mov	%esi, %r8d
+	and	$7, %r8d
+
+	.byte	0xf3,0x48,0x0f,0xb8,0x07	
+	xor	%ecx, %ecx
+
+	lea	Ltab(%rip), %r9
+
+	movslq	(%r9,%r8,4), %r8
+	add	%r9, %r8
+	jmp	*%r8
+
+
+L3:	.byte	0xf3,0x4c,0x0f,0xb8,0x57,0x08	
+	.byte	0xf3,0x4c,0x0f,0xb8,0x5f,0x10	
+	add	$24, %rdi
+	sub	$8, %rsi
+	jg	Le34
+	add	%r10, %rax
+	add	%r11, %rax
+Ls1:	
+	ret
+
+L1:	sub	$8, %rsi
+	jle	Ls1
+	.byte	0xf3,0x4c,0x0f,0xb8,0x47,0x08	
+	.byte	0xf3,0x4c,0x0f,0xb8,0x4f,0x10	
+	add	$8, %rdi
+	jmp	Le12
+
+L7:	.byte	0xf3,0x4c,0x0f,0xb8,0x57,0x08	
+	.byte	0xf3,0x4c,0x0f,0xb8,0x5f,0x10	
+	add	$-8, %rdi
+	jmp	Le07
+
+L0:	.byte	0xf3,0x48,0x0f,0xb8,0x4f,0x08	
+	.byte	0xf3,0x4c,0x0f,0xb8,0x57,0x10	
+	.byte	0xf3,0x4c,0x0f,0xb8,0x5f,0x18	
+	jmp	Le07
+
+L4:	.byte	0xf3,0x48,0x0f,0xb8,0x4f,0x08	
+	.byte	0xf3,0x4c,0x0f,0xb8,0x57,0x10	
+	.byte	0xf3,0x4c,0x0f,0xb8,0x5f,0x18	
+	add	$32, %rdi
+	sub	$8, %rsi
+	jle	Lx4
+
+	.align	4, 0x90
+Ltop:
+Le34:	.byte	0xf3,0x4c,0x0f,0xb8,0x07	
+	.byte	0xf3,0x4c,0x0f,0xb8,0x4f,0x08	
+	add	%r10, %rcx
+	add	%r11, %rax
+Le12:	.byte	0xf3,0x4c,0x0f,0xb8,0x57,0x10	
+	.byte	0xf3,0x4c,0x0f,0xb8,0x5f,0x18	
+	add	%r8, %rcx
+	add	%r9, %rax
+Le07:	.byte	0xf3,0x4c,0x0f,0xb8,0x47,0x20	
+	.byte	0xf3,0x4c,0x0f,0xb8,0x4f,0x28	
+	add	%r10, %rcx
+	add	%r11, %rax
+Le56:	.byte	0xf3,0x4c,0x0f,0xb8,0x57,0x30	
+	.byte	0xf3,0x4c,0x0f,0xb8,0x5f,0x38	
+	add	$64, %rdi
+	add	%r8, %rcx
+	add	%r9, %rax
+	sub	$8, %rsi
+	jg	Ltop
+
+Lx4:	add	%r10, %rcx
+	add	%r11, %rax
+Lx2:	add	%rcx, %rax
+
+	
+	ret
+
+L2:	.byte	0xf3,0x48,0x0f,0xb8,0x4f,0x08	
+	sub	$8, %rsi
+	jle	Lx2
+	.byte	0xf3,0x4c,0x0f,0xb8,0x47,0x10	
+	.byte	0xf3,0x4c,0x0f,0xb8,0x4f,0x18	
+	add	$16, %rdi
+	jmp	Le12
+
+L5:	.byte	0xf3,0x4c,0x0f,0xb8,0x47,0x08	
+	.byte	0xf3,0x4c,0x0f,0xb8,0x4f,0x10	
+	add	$-24, %rdi
+	jmp	Le56
+
+L6:	.byte	0xf3,0x48,0x0f,0xb8,0x4f,0x08	
+	.byte	0xf3,0x4c,0x0f,0xb8,0x47,0x10	
+	.byte	0xf3,0x4c,0x0f,0xb8,0x4f,0x18	
+	add	$-16, %rdi
+	jmp	Le56
+	
+	.text
+	.align	3, 0x90
+Ltab:	.set	L0_tmp, L0-Ltab
+	.long	L0_tmp
+
+	.set	L1_tmp, L1-Ltab
+	.long	L1_tmp
+
+	.set	L2_tmp, L2-Ltab
+	.long	L2_tmp
+
+	.set	L3_tmp, L3-Ltab
+	.long	L3_tmp
+
+	.set	L4_tmp, L4-Ltab
+	.long	L4_tmp
+
+	.set	L5_tmp, L5-Ltab
+	.long	L5_tmp
+
+	.set	L6_tmp, L6-Ltab
+	.long	L6_tmp
+
+	.set	L7_tmp, L7-Ltab
+	.long	L7_tmp
+