
########## ra_onesided_probe.clang.s ##########

===== copy_generic_bounded (simd=True, lines=121) =====
copy_generic_bounded:                   # @copy_generic_bounded
	.cfi_startproc
# %bb.0:
	cmpq	%rsi, %rdi
	je	.LBB0_1
# %bb.3:
	cmpq	%rcx, %rdx
	je	.LBB0_1
# %bb.4:
	movq	%rcx, %rax
	subq	%rdx, %rax
	addq	$-4, %rax
	shrq	$2, %rax
	movq	%rsi, %r8
	subq	%rdi, %r8
	addq	$-4, %r8
	shrq	$2, %r8
	cmpq	%r8, %rax
	cmovbq	%rax, %r8
	cmpq	$19, %r8
	jb	.LBB0_5
# %bb.6:
	movl	%esi, %eax
	subl	%edi, %eax
	movl	%ecx, %r9d
	subl	%edx, %r9d
	orl	%eax, %r9d
	testb	$3, %r9b
	setne	%al
	movq	%rdx, %r9
	subq	%rdi, %r9
	cmpq	$32, %r9
	setb	%r9b
	orb	%al, %r9b
	je	.LBB0_7
.LBB0_5:
	movq	%rdi, %r9
	movq	%rdx, %rax
.LBB0_10:
	addq	$4, %r9
	.p2align	4
.LBB0_11:                               # =>This Inner Loop Header: Depth=1
	movl	-4(%r9), %edx
	movl	%edx, (%rax)
	addq	$4, %rax
	cmpq	%rsi, %r9
	je	.LBB0_2
# %bb.12:                               #   in Loop: Header=BB0_11 Depth=1
	addq	$4, %r9
	cmpq	%rcx, %rax
	jne	.LBB0_11
.LBB0_2:
	retq
.LBB0_1:
	movq	%rdx, %rax
	retq
.LBB0_7:
	incq	%r8
	movq	%r8, %r10
	andq	$-8, %r10
	leaq	(%rdi,%r10,4), %r9
	leaq	(%rdx,%r10,4), %rax
	xorl	%r11d, %r11d
	.p2align	4
.LBB0_8:                                # =>This Inner Loop Header: Depth=1
	movups	(%rdi,%r11,4), %xmm0
	movups	16(%rdi,%r11,4), %xmm1
	movups	%xmm0, (%rdx,%r11,4)
	movups	%xmm1, 16(%rdx,%r11,4)
	addq	$8, %r11
	cmpq	%r11, %r10
	jne	.LBB0_8
# %bb.9:
	cmpq	%r10, %r8
	jne	.LBB0_10
	jmp	.LBB0_2
.Lfunc_end0:
	.size	copy_generic_bounded, .Lfunc_end0-copy_generic_bounded
	.cfi_endproc
                                        # -- End function
	.globl	copy_both_ra                    # -- Begin function copy_both_ra
	.p2align	4
	.type	copy_both_ra,@function
copy_both_ra:                           # @copy_both_ra
	.cfi_startproc
# %bb.0:
	subq	%rdi, %rsi
	sarq	$2, %rsi
	subq	%rdx, %rcx
	sarq	$2, %rcx

===== copy_both_ra (simd=True, lines=121) =====
copy_both_ra:                           # @copy_both_ra
	.cfi_startproc
# %bb.0:
	subq	%rdi, %rsi
	sarq	$2, %rsi
	subq	%rdx, %rcx
	sarq	$2, %rcx
	cmpq	%rcx, %rsi
	cmovlq	%rsi, %rcx
	testq	%rcx, %rcx
	je	.LBB1_1
# %bb.3:
	leaq	(,%rcx,4), %r8
	addq	$-4, %r8
	cmpq	$28, %r8
	setb	%al
	movq	%rdx, %rsi
	subq	%rdi, %rsi
	cmpq	$32, %rsi
	setb	%sil
	orb	%al, %sil
	je	.LBB1_5
# %bb.4:
	movq	%rdi, %rsi
	movq	%rdx, %rax
	jmp	.LBB1_8
.LBB1_1:
	movq	%rdx, %rax
	retq
.LBB1_5:
	shrq	$2, %r8
	incq	%r8
	movq	%r8, %r9
	andq	$-8, %r9
	leaq	(%rdi,%r9,4), %rsi
	leaq	(%rdx,%r9,4), %rax
	xorl	%r10d, %r10d
	.p2align	4
.LBB1_6:                                # =>This Inner Loop Header: Depth=1
	movups	(%rdi,%r10,4), %xmm0
	movups	16(%rdi,%r10,4), %xmm1
	movups	%xmm0, (%rdx,%r10,4)
	movups	%xmm1, 16(%rdx,%r10,4)
	addq	$8, %r10
	cmpq	%r10, %r9
	jne	.LBB1_6
# %bb.7:
	cmpq	%r9, %r8
	je	.LBB1_2
.LBB1_8:
	leaq	(%rdi,%rcx,4), %rcx
	.p2align	4
.LBB1_9:                                # =>This Inner Loop Header: Depth=1
	movl	(%rsi), %edx
	movl	%edx, (%rax)
	addq	$4, %rsi
	addq	$4, %rax
	cmpq	%rcx, %rsi
	jne	.LBB1_9
.LBB1_2:
	retq
.Lfunc_end1:
	.size	copy_both_ra, .Lfunc_end1-copy_both_ra
	.cfi_endproc
                                        # -- End function
	.globl	copy_ra_dst_only                # -- Begin function copy_ra_dst_only
	.p2align	4
	.type	copy_ra_dst_only,@function
copy_ra_dst_only:                       # @copy_ra_dst_only
	.cfi_startproc
# %bb.0:
	subq	%rdx, %rcx
	setne	%al
	cmpq	%rsi, %rdi
	setne	%r8b
	andb	%al, %r8b
	cmpb	$1, %r8b
	jne	.LBB2_1
# %bb.3:
	sarq	$2, %rcx
	movq	%rsi, %r9
	subq	%rdi, %r9
	addq	$-4, %r9
	shrq	$2, %r9
	leaq	-1(%rcx), %rax
	cmpq	%rax, %r9
	cmovaeq	%rax, %r9
	cmpq	$11, %r9
	jb	.LBB2_4
# %bb.5:

===== copy_ra_dst_only (simd=True, lines=121) =====
copy_ra_dst_only:                       # @copy_ra_dst_only
	.cfi_startproc
# %bb.0:
	subq	%rdx, %rcx
	setne	%al
	cmpq	%rsi, %rdi
	setne	%r8b
	andb	%al, %r8b
	cmpb	$1, %r8b
	jne	.LBB2_1
# %bb.3:
	sarq	$2, %rcx
	movq	%rsi, %r9
	subq	%rdi, %r9
	addq	$-4, %r9
	shrq	$2, %r9
	leaq	-1(%rcx), %rax
	cmpq	%rax, %r9
	cmovaeq	%rax, %r9
	cmpq	$11, %r9
	jb	.LBB2_4
# %bb.5:
	movl	%esi, %eax
	subl	%edi, %eax
	testb	$3, %al
	setne	%al
	movq	%rdx, %r8
	subq	%rdi, %r8
	cmpq	$32, %r8
	setb	%r8b
	orb	%al, %r8b
	je	.LBB2_6
.LBB2_4:
	movq	%rdx, %rax
	movq	%rdi, %r8
.LBB2_9:
	decq	%rcx
	addq	$4, %r8
	.p2align	4
.LBB2_10:                               # =>This Inner Loop Header: Depth=1
	movl	-4(%r8), %edx
	movl	%edx, (%rax)
	addq	$4, %rax
	addq	$-1, %rcx
	jae	.LBB2_2
# %bb.11:                               #   in Loop: Header=BB2_10 Depth=1
	cmpq	%rsi, %r8
	leaq	4(%r8), %r8
	jne	.LBB2_10
.LBB2_2:
	retq
.LBB2_1:
	movq	%rdx, %rax
	retq
.LBB2_6:
	incq	%r9
	movabsq	$9223372036854775800, %r10      # imm = 0x7FFFFFFFFFFFFFF8
	andq	%r9, %r10
	leaq	(%rdx,%r10,4), %rax
	subq	%r10, %rcx
	leaq	(%rdi,%r10,4), %r8
	xorl	%r11d, %r11d
	.p2align	4
.LBB2_7:                                # =>This Inner Loop Header: Depth=1
	movups	(%rdi,%r11,4), %xmm0
	movups	16(%rdi,%r11,4), %xmm1
	movups	%xmm0, (%rdx,%r11,4)
	movups	%xmm1, 16(%rdx,%r11,4)
	addq	$8, %r11
	cmpq	%r11, %r10
	jne	.LBB2_7
# %bb.8:
	cmpq	%r10, %r9
	jne	.LBB2_9
	jmp	.LBB2_2
.Lfunc_end2:
	.size	copy_ra_dst_only, .Lfunc_end2-copy_ra_dst_only
	.cfi_endproc
                                        # -- End function
	.globl	copy_ra_src_only_bounded        # -- Begin function copy_ra_src_only_bounded
	.p2align	4
	.type	copy_ra_src_only_bounded,@function
copy_ra_src_only_bounded:               # @copy_ra_src_only_bounded
	.cfi_startproc
# %bb.0:
	subq	%rdi, %rsi
	setne	%al
	cmpq	%rcx, %rdx
	setne	%r8b
	andb	%al, %r8b

===== copy_ra_src_only_bounded (simd=True, lines=121) =====
copy_ra_src_only_bounded:               # @copy_ra_src_only_bounded
	.cfi_startproc
# %bb.0:
	subq	%rdi, %rsi
	setne	%al
	cmpq	%rcx, %rdx
	setne	%r8b
	andb	%al, %r8b
	cmpb	$1, %r8b
	jne	.LBB3_1
# %bb.3:
	sarq	$2, %rsi
	movq	%rcx, %r9
	subq	%rdx, %r9
	addq	$-4, %r9
	shrq	$2, %r9
	leaq	-1(%rsi), %rax
	cmpq	%rax, %r9
	cmovaeq	%rax, %r9
	cmpq	$11, %r9
	jb	.LBB3_4
# %bb.5:
	movl	%ecx, %eax
	subl	%edx, %eax
	testb	$3, %al
	setne	%al
	movq	%rdx, %r8
	subq	%rdi, %r8
	cmpq	$32, %r8
	setb	%r8b
	orb	%al, %r8b
	je	.LBB3_6
.LBB3_4:
	movq	%rdi, %r8
	movq	%rdx, %rax
.LBB3_9:
	decq	%rsi
	.p2align	4
.LBB3_10:                               # =>This Inner Loop Header: Depth=1
	movl	(%r8), %edx
	movl	%edx, (%rax)
	addq	$4, %rax
	addq	$-1, %rsi
	jae	.LBB3_2
# %bb.11:                               #   in Loop: Header=BB3_10 Depth=1
	addq	$4, %r8
	cmpq	%rcx, %rax
	jne	.LBB3_10
.LBB3_2:
	retq
.LBB3_1:
	movq	%rdx, %rax
	retq
.LBB3_6:
	incq	%r9
	movabsq	$9223372036854775800, %r10      # imm = 0x7FFFFFFFFFFFFFF8
	andq	%r9, %r10
	leaq	(%rdi,%r10,4), %r8
	subq	%r10, %rsi
	leaq	(%rdx,%r10,4), %rax
	xorl	%r11d, %r11d
	.p2align	4
.LBB3_7:                                # =>This Inner Loop Header: Depth=1
	movups	(%rdi,%r11,4), %xmm0
	movups	16(%rdi,%r11,4), %xmm1
	movups	%xmm0, (%rdx,%r11,4)
	movups	%xmm1, 16(%rdx,%r11,4)
	addq	$8, %r11
	cmpq	%r11, %r10
	jne	.LBB3_7
# %bb.8:
	cmpq	%r10, %r9
	jne	.LBB3_9
	jmp	.LBB3_2
.Lfunc_end3:
	.size	copy_ra_src_only_bounded, .Lfunc_end3-copy_ra_src_only_bounded
	.cfi_endproc
                                        # -- End function
	.globl	copy_ra_src_unbounded           # -- Begin function copy_ra_src_unbounded
	.p2align	4
	.type	copy_ra_src_unbounded,@function
copy_ra_src_unbounded:                  # @copy_ra_src_unbounded
	.cfi_startproc
# %bb.0:
	cmpq	%rsi, %rdi
	je	.LBB4_1
# %bb.3:
	movq	%rsi, %r8
	subq	%rdi, %r8
	addq	$-4, %r8

===== copy_ra_src_unbounded (simd=True, lines=121) =====
copy_ra_src_unbounded:                  # @copy_ra_src_unbounded
	.cfi_startproc
# %bb.0:
	cmpq	%rsi, %rdi
	je	.LBB4_1
# %bb.3:
	movq	%rsi, %r8
	subq	%rdi, %r8
	addq	$-4, %r8
	cmpq	$28, %r8
	setb	%al
	movq	%rdx, %rcx
	subq	%rdi, %rcx
	cmpq	$32, %rcx
	setb	%cl
	orb	%al, %cl
	je	.LBB4_5
# %bb.4:
	movq	%rdx, %rax
	movq	%rdi, %rcx
	jmp	.LBB4_8
.LBB4_1:
	movq	%rdx, %rax
	retq
.LBB4_5:
	shrq	$2, %r8
	incq	%r8
	movq	%r8, %r9
	andq	$-8, %r9
	leaq	(%rdx,%r9,4), %rax
	leaq	(%rdi,%r9,4), %rcx
	xorl	%r10d, %r10d
	.p2align	4
.LBB4_6:                                # =>This Inner Loop Header: Depth=1
	movups	(%rdi,%r10,4), %xmm0
	movups	16(%rdi,%r10,4), %xmm1
	movups	%xmm0, (%rdx,%r10,4)
	movups	%xmm1, 16(%rdx,%r10,4)
	addq	$8, %r10
	cmpq	%r10, %r9
	jne	.LBB4_6
# %bb.7:
	cmpq	%r9, %r8
	je	.LBB4_2
	.p2align	4
.LBB4_8:                                # =>This Inner Loop Header: Depth=1
	movl	(%rcx), %edx
	movl	%edx, (%rax)
	addq	$4, %rcx
	addq	$4, %rax
	cmpq	%rsi, %rcx
	jne	.LBB4_8
.LBB4_2:
	retq
.Lfunc_end4:
	.size	copy_ra_src_unbounded, .Lfunc_end4-copy_ra_src_unbounded
	.cfi_endproc
                                        # -- End function
	.globl	copy_ra_src_counted             # -- Begin function copy_ra_src_counted
	.p2align	4
	.type	copy_ra_src_counted,@function
copy_ra_src_counted:                    # @copy_ra_src_counted
	.cfi_startproc
# %bb.0:
	subq	%rdi, %rsi
	je	.LBB5_1
# %bb.3:
	sarq	$2, %rsi
	cmpq	$8, %rsi
	setb	%al
	movq	%rdx, %rcx
	subq	%rdi, %rcx
	cmpq	$32, %rcx
	setb	%cl
	orb	%al, %cl
	je	.LBB5_5
# %bb.4:
	movq	%rsi, %rcx
	movq	%rdx, %rax
	movq	%rdi, %r8
	jmp	.LBB5_8
.LBB5_1:
	movq	%rdx, %rax
	retq
.LBB5_5:
	movq	%rsi, %r9
	andq	$-8, %r9
	movl	%esi, %ecx
	andl	$7, %ecx
	leaq	(%rdx,%r9,4), %rax

===== copy_ra_src_counted (simd=True, lines=121) =====
copy_ra_src_counted:                    # @copy_ra_src_counted
	.cfi_startproc
# %bb.0:
	subq	%rdi, %rsi
	je	.LBB5_1
# %bb.3:
	sarq	$2, %rsi
	cmpq	$8, %rsi
	setb	%al
	movq	%rdx, %rcx
	subq	%rdi, %rcx
	cmpq	$32, %rcx
	setb	%cl
	orb	%al, %cl
	je	.LBB5_5
# %bb.4:
	movq	%rsi, %rcx
	movq	%rdx, %rax
	movq	%rdi, %r8
	jmp	.LBB5_8
.LBB5_1:
	movq	%rdx, %rax
	retq
.LBB5_5:
	movq	%rsi, %r9
	andq	$-8, %r9
	movl	%esi, %ecx
	andl	$7, %ecx
	leaq	(%rdx,%r9,4), %rax
	leaq	(%rdi,%r9,4), %r8
	xorl	%r10d, %r10d
	.p2align	4
.LBB5_6:                                # =>This Inner Loop Header: Depth=1
	movups	(%rdi,%r10,4), %xmm0
	movups	16(%rdi,%r10,4), %xmm1
	movups	%xmm0, (%rdx,%r10,4)
	movups	%xmm1, 16(%rdx,%r10,4)
	addq	$8, %r10
	cmpq	%r10, %r9
	jne	.LBB5_6
# %bb.7:
	cmpq	%r9, %rsi
	je	.LBB5_2
.LBB5_8:
	leaq	-1(%rcx), %rdx
	movq	%rcx, %rsi
	andq	$7, %rsi
	je	.LBB5_12
# %bb.9:
	xorl	%edi, %edi
	.p2align	4
.LBB5_10:                               # =>This Inner Loop Header: Depth=1
	movl	(%r8), %r9d
	movl	%r9d, (%rax)
	addq	$4, %r8
	addq	$4, %rax
	incq	%rdi
	cmpq	%rdi, %rsi
	jne	.LBB5_10
# %bb.11:
	subq	%rdi, %rcx
.LBB5_12:
	cmpq	$7, %rdx
	jb	.LBB5_2
# %bb.13:
	xorl	%edx, %edx
	.p2align	4
.LBB5_14:                               # =>This Inner Loop Header: Depth=1
	movl	(%r8,%rdx,4), %esi
	movl	%esi, (%rax)
	movl	4(%r8,%rdx,4), %esi
	movl	%esi, 4(%rax)
	movl	8(%r8,%rdx,4), %esi
	movl	%esi, 8(%rax)
	movl	12(%r8,%rdx,4), %esi
	movl	%esi, 12(%rax)
	movl	16(%r8,%rdx,4), %esi
	movl	%esi, 16(%rax)
	movl	20(%r8,%rdx,4), %esi
	movl	%esi, 20(%rax)
	movl	24(%r8,%rdx,4), %esi
	movl	%esi, 24(%rax)
	movl	28(%r8,%rdx,4), %esi
	movl	%esi, 28(%rax)
	addq	$32, %rax
	addq	$8, %rdx
	cmpq	%rdx, %rcx
	jne	.LBB5_14
.LBB5_2:
	retq

===== copy_if_generic (simd=False, lines=121) =====
copy_if_generic:                        # @copy_if_generic
	.cfi_startproc
# %bb.0:
	movq	%rdx, %rax
	cmpq	%rcx, %rdx
	jne	.LBB6_1
	jmp	.LBB6_5
	.p2align	4
.LBB6_4:                                #   in Loop: Header=BB6_1 Depth=1
	addq	$4, %rdi
.LBB6_1:                                # =>This Inner Loop Header: Depth=1
	cmpq	%rsi, %rdi
	je	.LBB6_5
# %bb.2:                                #   in Loop: Header=BB6_1 Depth=1
	movl	(%rdi), %edx
	testb	$1, %dl
	je	.LBB6_4
# %bb.3:                                #   in Loop: Header=BB6_1 Depth=1
	movl	%edx, (%rax)
	addq	$4, %rax
	cmpq	%rcx, %rax
	jne	.LBB6_4
.LBB6_5:
	retq
.Lfunc_end6:
	.size	copy_if_generic, .Lfunc_end6-copy_if_generic
	.cfi_endproc
                                        # -- End function
	.globl	copy_if_both_ra                 # -- Begin function copy_if_both_ra
	.p2align	4
	.type	copy_if_both_ra,@function
copy_if_both_ra:                        # @copy_if_both_ra
	.cfi_startproc
# %bb.0:
	movq	%rdx, %rax
	cmpq	%rcx, %rdx
	je	.LBB7_43
# %bb.1:
	subq	%rdi, %rsi
	sarq	$2, %rsi
	cmpq	$16, %rsi
	setl	%dl
	movq	%rcx, %r8
	subq	%rax, %r8
	cmpq	$61, %r8
	setl	%r8b
	orb	%dl, %r8b
	jne	.LBB7_2
	.p2align	4
.LBB7_3:                                # =>This Inner Loop Header: Depth=1
	movl	(%rdi), %edx
	testb	$1, %dl
	jne	.LBB7_4
# %bb.5:                                #   in Loop: Header=BB7_3 Depth=1
	movl	4(%rdi), %edx
	testb	$1, %dl
	jne	.LBB7_6
.LBB7_7:                                #   in Loop: Header=BB7_3 Depth=1
	movl	8(%rdi), %edx
	testb	$1, %dl
	jne	.LBB7_8
.LBB7_9:                                #   in Loop: Header=BB7_3 Depth=1
	movl	12(%rdi), %edx
	testb	$1, %dl
	jne	.LBB7_10
.LBB7_11:                               #   in Loop: Header=BB7_3 Depth=1
	movl	16(%rdi), %edx
	testb	$1, %dl
	jne	.LBB7_12
.LBB7_13:                               #   in Loop: Header=BB7_3 Depth=1
	movl	20(%rdi), %edx
	testb	$1, %dl
	jne	.LBB7_14
.LBB7_15:                               #   in Loop: Header=BB7_3 Depth=1
	movl	24(%rdi), %edx
	testb	$1, %dl
	jne	.LBB7_16
.LBB7_17:                               #   in Loop: Header=BB7_3 Depth=1
	movl	28(%rdi), %edx
	testb	$1, %dl
	jne	.LBB7_18
.LBB7_19:                               #   in Loop: Header=BB7_3 Depth=1
	movl	32(%rdi), %edx
	testb	$1, %dl
	jne	.LBB7_20
.LBB7_21:                               #   in Loop: Header=BB7_3 Depth=1
	movl	36(%rdi), %edx
	testb	$1, %dl
	jne	.LBB7_22
.LBB7_23:                               #   in Loop: Header=BB7_3 Depth=1

===== copy_if_both_ra (simd=False, lines=121) =====
copy_if_both_ra:                        # @copy_if_both_ra
	.cfi_startproc
# %bb.0:
	movq	%rdx, %rax
	cmpq	%rcx, %rdx
	je	.LBB7_43
# %bb.1:
	subq	%rdi, %rsi
	sarq	$2, %rsi
	cmpq	$16, %rsi
	setl	%dl
	movq	%rcx, %r8
	subq	%rax, %r8
	cmpq	$61, %r8
	setl	%r8b
	orb	%dl, %r8b
	jne	.LBB7_2
	.p2align	4
.LBB7_3:                                # =>This Inner Loop Header: Depth=1
	movl	(%rdi), %edx
	testb	$1, %dl
	jne	.LBB7_4
# %bb.5:                                #   in Loop: Header=BB7_3 Depth=1
	movl	4(%rdi), %edx
	testb	$1, %dl
	jne	.LBB7_6
.LBB7_7:                                #   in Loop: Header=BB7_3 Depth=1
	movl	8(%rdi), %edx
	testb	$1, %dl
	jne	.LBB7_8
.LBB7_9:                                #   in Loop: Header=BB7_3 Depth=1
	movl	12(%rdi), %edx
	testb	$1, %dl
	jne	.LBB7_10
.LBB7_11:                               #   in Loop: Header=BB7_3 Depth=1
	movl	16(%rdi), %edx
	testb	$1, %dl
	jne	.LBB7_12
.LBB7_13:                               #   in Loop: Header=BB7_3 Depth=1
	movl	20(%rdi), %edx
	testb	$1, %dl
	jne	.LBB7_14
.LBB7_15:                               #   in Loop: Header=BB7_3 Depth=1
	movl	24(%rdi), %edx
	testb	$1, %dl
	jne	.LBB7_16
.LBB7_17:                               #   in Loop: Header=BB7_3 Depth=1
	movl	28(%rdi), %edx
	testb	$1, %dl
	jne	.LBB7_18
.LBB7_19:                               #   in Loop: Header=BB7_3 Depth=1
	movl	32(%rdi), %edx
	testb	$1, %dl
	jne	.LBB7_20
.LBB7_21:                               #   in Loop: Header=BB7_3 Depth=1
	movl	36(%rdi), %edx
	testb	$1, %dl
	jne	.LBB7_22
.LBB7_23:                               #   in Loop: Header=BB7_3 Depth=1
	movl	40(%rdi), %edx
	testb	$1, %dl
	jne	.LBB7_24
.LBB7_25:                               #   in Loop: Header=BB7_3 Depth=1
	movl	44(%rdi), %edx
	testb	$1, %dl
	jne	.LBB7_26
.LBB7_27:                               #   in Loop: Header=BB7_3 Depth=1
	movl	48(%rdi), %edx
	testb	$1, %dl
	jne	.LBB7_28
.LBB7_29:                               #   in Loop: Header=BB7_3 Depth=1
	movl	52(%rdi), %edx
	testb	$1, %dl
	jne	.LBB7_30
.LBB7_31:                               #   in Loop: Header=BB7_3 Depth=1
	movl	56(%rdi), %edx
	testb	$1, %dl
	jne	.LBB7_32
.LBB7_33:                               #   in Loop: Header=BB7_3 Depth=1
	movl	60(%rdi), %edx
	testb	$1, %dl
	jne	.LBB7_34
.LBB7_35:                               #   in Loop: Header=BB7_3 Depth=1
	addq	$64, %rdi
	leaq	-16(%rsi), %rdx
	cmpq	$32, %rsi
	jge	.LBB7_36
	jmp	.LBB7_37
	.p2align	4
.LBB7_4:                                #   in Loop: Header=BB7_3 Depth=1

===== copy_if_ra_dst_only (simd=False, lines=121) =====
copy_if_ra_dst_only:                    # @copy_if_ra_dst_only
	.cfi_startproc
# %bb.0:
	movq	%rdx, %rax
	cmpq	%rcx, %rdx
	jne	.LBB8_1
.LBB8_9:
	retq
	.p2align	4
.LBB8_8:                                #   in Loop: Header=BB8_1 Depth=1
	cmpq	%rsi, %rdi
	je	.LBB8_9
.LBB8_1:                                # =>This Loop Header: Depth=1
                                        #     Child Loop BB8_4 Depth 2
	movq	%rcx, %rdx
	subq	%rax, %rdx
	cmpq	$61, %rdx
	jl	.LBB8_10
# %bb.2:                                #   in Loop: Header=BB8_1 Depth=1
	cmpq	%rsi, %rdi
	je	.LBB8_10
# %bb.3:                                #   in Loop: Header=BB8_1 Depth=1
	movl	$60, %edx
	.p2align	4
.LBB8_4:                                #   Parent Loop BB8_1 Depth=1
                                        # =>  This Inner Loop Header: Depth=2
	movl	(%rdi), %r8d
	testb	$1, %r8b
	je	.LBB8_6
# %bb.5:                                #   in Loop: Header=BB8_4 Depth=2
	movl	%r8d, (%rax)
	addq	$4, %rax
.LBB8_6:                                #   in Loop: Header=BB8_4 Depth=2
	addq	$4, %rdi
	testq	%rdx, %rdx
	je	.LBB8_8
# %bb.7:                                #   in Loop: Header=BB8_4 Depth=2
	addq	$-4, %rdx
	cmpq	%rsi, %rdi
	jne	.LBB8_4
	jmp	.LBB8_8
.LBB8_10:
	cmpq	%rcx, %rax
	jne	.LBB8_11
	jmp	.LBB8_9
	.p2align	4
.LBB8_14:                               #   in Loop: Header=BB8_11 Depth=1
	addq	$4, %rdi
.LBB8_11:                               # =>This Inner Loop Header: Depth=1
	cmpq	%rsi, %rdi
	je	.LBB8_9
# %bb.12:                               #   in Loop: Header=BB8_11 Depth=1
	movl	(%rdi), %edx
	testb	$1, %dl
	je	.LBB8_14
# %bb.13:                               #   in Loop: Header=BB8_11 Depth=1
	movl	%edx, (%rax)
	addq	$4, %rax
	cmpq	%rcx, %rax
	jne	.LBB8_14
	jmp	.LBB8_9
.Lfunc_end8:
	.size	copy_if_ra_dst_only, .Lfunc_end8-copy_if_ra_dst_only
	.cfi_endproc
                                        # -- End function
	.globl	copy_if_ra_src_unbounded        # -- Begin function copy_if_ra_src_unbounded
	.p2align	4
	.type	copy_if_ra_src_unbounded,@function
copy_if_ra_src_unbounded:               # @copy_if_ra_src_unbounded
	.cfi_startproc
# %bb.0:
	movq	%rdx, %rax
	subq	%rdi, %rsi
	je	.LBB9_12
# %bb.1:
	movq	%rsi, %rcx
	sarq	$2, %rcx
	testb	$4, %sil
	je	.LBB9_5
# %bb.2:
	movl	(%rdi), %edx
	testb	$1, %dl
	je	.LBB9_4
# %bb.3:
	movl	%edx, (%rax)
	addq	$4, %rax
.LBB9_4:
	decq	%rcx
	addq	$4, %rdi
.LBB9_5:

===== copy_if_ra_src_unbounded (simd=False, lines=121) =====
copy_if_ra_src_unbounded:               # @copy_if_ra_src_unbounded
	.cfi_startproc
# %bb.0:
	movq	%rdx, %rax
	subq	%rdi, %rsi
	je	.LBB9_12
# %bb.1:
	movq	%rsi, %rcx
	sarq	$2, %rcx
	testb	$4, %sil
	je	.LBB9_5
# %bb.2:
	movl	(%rdi), %edx
	testb	$1, %dl
	je	.LBB9_4
# %bb.3:
	movl	%edx, (%rax)
	addq	$4, %rax
.LBB9_4:
	decq	%rcx
	addq	$4, %rdi
.LBB9_5:
	cmpq	$4, %rsi
	jne	.LBB9_6
.LBB9_12:
	retq
.LBB9_6:
	xorl	%edx, %edx
	jmp	.LBB9_7
	.p2align	4
.LBB9_11:                               #   in Loop: Header=BB9_7 Depth=1
	addq	$2, %rdx
	cmpq	%rdx, %rcx
	je	.LBB9_12
.LBB9_7:                                # =>This Inner Loop Header: Depth=1
	movl	(%rdi,%rdx,4), %esi
	testb	$1, %sil
	jne	.LBB9_8
# %bb.9:                                #   in Loop: Header=BB9_7 Depth=1
	movl	4(%rdi,%rdx,4), %esi
	testb	$1, %sil
	je	.LBB9_11
	jmp	.LBB9_10
	.p2align	4
.LBB9_8:                                #   in Loop: Header=BB9_7 Depth=1
	movl	%esi, (%rax)
	addq	$4, %rax
	movl	4(%rdi,%rdx,4), %esi
	testb	$1, %sil
	je	.LBB9_11
.LBB9_10:                               #   in Loop: Header=BB9_7 Depth=1
	movl	%esi, (%rax)
	addq	$4, %rax
	jmp	.LBB9_11
.Lfunc_end9:
	.size	copy_if_ra_src_unbounded, .Lfunc_end9-copy_if_ra_src_unbounded
	.cfi_endproc
                                        # -- End function
	.globl	equal_generic                   # -- Begin function equal_generic
	.p2align	4
	.type	equal_generic,@function
equal_generic:                          # @equal_generic
	.cfi_startproc
# %bb.0:
	cmpq	%rsi, %rdi
	sete	%al
	je	.LBB10_7
# %bb.1:
	cmpq	%rcx, %rdx
	je	.LBB10_7
# %bb.2:
	addq	$4, %rdx
	addq	$4, %rdi
	.p2align	4
.LBB10_3:                               # =>This Inner Loop Header: Depth=1
	movl	-4(%rdi), %eax
	cmpl	-4(%rdx), %eax
	jne	.LBB10_4
# %bb.5:                                #   in Loop: Header=BB10_3 Depth=1
	cmpq	%rsi, %rdi
	sete	%al
	je	.LBB10_7
# %bb.6:                                #   in Loop: Header=BB10_3 Depth=1
	addq	$4, %rdi
	cmpq	%rcx, %rdx
	leaq	4(%rdx), %rdx
	jne	.LBB10_3
.LBB10_7:
                                        # kill: def $al killed $al killed $eax
	retq

===== equal_generic (simd=False, lines=121) =====
equal_generic:                          # @equal_generic
	.cfi_startproc
# %bb.0:
	cmpq	%rsi, %rdi
	sete	%al
	je	.LBB10_7
# %bb.1:
	cmpq	%rcx, %rdx
	je	.LBB10_7
# %bb.2:
	addq	$4, %rdx
	addq	$4, %rdi
	.p2align	4
.LBB10_3:                               # =>This Inner Loop Header: Depth=1
	movl	-4(%rdi), %eax
	cmpl	-4(%rdx), %eax
	jne	.LBB10_4
# %bb.5:                                #   in Loop: Header=BB10_3 Depth=1
	cmpq	%rsi, %rdi
	sete	%al
	je	.LBB10_7
# %bb.6:                                #   in Loop: Header=BB10_3 Depth=1
	addq	$4, %rdi
	cmpq	%rcx, %rdx
	leaq	4(%rdx), %rdx
	jne	.LBB10_3
.LBB10_7:
                                        # kill: def $al killed $al killed $eax
	retq
.LBB10_4:
	xorl	%eax, %eax
                                        # kill: def $al killed $al killed $eax
	retq
.Lfunc_end10:
	.size	equal_generic, .Lfunc_end10-equal_generic
	.cfi_endproc
                                        # -- End function
	.globl	equal_both_ra                   # -- Begin function equal_both_ra
	.p2align	4
	.type	equal_both_ra,@function
equal_both_ra:                          # @equal_both_ra
	.cfi_startproc
# %bb.0:
	movq	%rsi, %rax
	subq	%rdi, %rax
	sarq	$2, %rax
	subq	%rdx, %rcx
	sarq	$2, %rcx
	cmpq	%rcx, %rax
	cmovlq	%rax, %rcx
	testq	%rcx, %rcx
	je	.LBB11_6
# %bb.1:
	leaq	(%rdi,%rcx,4), %rax
	shlq	$2, %rcx
	xorl	%r8d, %r8d
	.p2align	4
.LBB11_2:                               # =>This Inner Loop Header: Depth=1
	movl	(%rdi,%r8), %r9d
	cmpl	(%rdx,%r8), %r9d
	jne	.LBB11_3
# %bb.4:                                #   in Loop: Header=BB11_2 Depth=1
	addq	$4, %r8
	cmpq	%r8, %rcx
	jne	.LBB11_2
# %bb.5:
	movq	%rax, %rdi
.LBB11_6:
	cmpq	%rsi, %rdi
	sete	%al
                                        # kill: def $al killed $al killed $eax
	retq
.LBB11_3:
	xorl	%eax, %eax
                                        # kill: def $al killed $al killed $eax
	retq
.Lfunc_end11:
	.size	equal_both_ra, .Lfunc_end11-equal_both_ra
	.cfi_endproc
                                        # -- End function
	.globl	equal_ra_src_only               # -- Begin function equal_ra_src_only
	.p2align	4
	.type	equal_ra_src_only,@function
equal_ra_src_only:                      # @equal_ra_src_only
	.cfi_startproc
# %bb.0:
	subq	%rdi, %rsi
	sete	%al
	je	.LBB12_7
# %bb.1:

===== equal_both_ra (simd=False, lines=121) =====
equal_both_ra:                          # @equal_both_ra
	.cfi_startproc
# %bb.0:
	movq	%rsi, %rax
	subq	%rdi, %rax
	sarq	$2, %rax
	subq	%rdx, %rcx
	sarq	$2, %rcx
	cmpq	%rcx, %rax
	cmovlq	%rax, %rcx
	testq	%rcx, %rcx
	je	.LBB11_6
# %bb.1:
	leaq	(%rdi,%rcx,4), %rax
	shlq	$2, %rcx
	xorl	%r8d, %r8d
	.p2align	4
.LBB11_2:                               # =>This Inner Loop Header: Depth=1
	movl	(%rdi,%r8), %r9d
	cmpl	(%rdx,%r8), %r9d
	jne	.LBB11_3
# %bb.4:                                #   in Loop: Header=BB11_2 Depth=1
	addq	$4, %r8
	cmpq	%r8, %rcx
	jne	.LBB11_2
# %bb.5:
	movq	%rax, %rdi
.LBB11_6:
	cmpq	%rsi, %rdi
	sete	%al
                                        # kill: def $al killed $al killed $eax
	retq
.LBB11_3:
	xorl	%eax, %eax
                                        # kill: def $al killed $al killed $eax
	retq
.Lfunc_end11:
	.size	equal_both_ra, .Lfunc_end11-equal_both_ra
	.cfi_endproc
                                        # -- End function
	.globl	equal_ra_src_only               # -- Begin function equal_ra_src_only
	.p2align	4
	.type	equal_ra_src_only,@function
equal_ra_src_only:                      # @equal_ra_src_only
	.cfi_startproc
# %bb.0:
	subq	%rdi, %rsi
	sete	%al
	je	.LBB12_7
# %bb.1:
	cmpq	%rcx, %rdx
	je	.LBB12_7
# %bb.2:
	sarq	$2, %rsi
	decq	%rsi
	addq	$4, %rdx
	.p2align	4
.LBB12_3:                               # =>This Inner Loop Header: Depth=1
	movl	(%rdi), %eax
	cmpl	-4(%rdx), %eax
	jne	.LBB12_4
# %bb.5:                                #   in Loop: Header=BB12_3 Depth=1
	subq	$1, %rsi
	setb	%al
	jb	.LBB12_7
# %bb.6:                                #   in Loop: Header=BB12_3 Depth=1
	addq	$4, %rdi
	cmpq	%rcx, %rdx
	leaq	4(%rdx), %rdx
	jne	.LBB12_3
.LBB12_7:
                                        # kill: def $al killed $al killed $eax
	retq
.LBB12_4:
	xorl	%eax, %eax
                                        # kill: def $al killed $al killed $eax
	retq
.Lfunc_end12:
	.size	equal_ra_src_only, .Lfunc_end12-equal_ra_src_only
	.cfi_endproc
                                        # -- End function
	.globl	equal_ra_dst_only               # -- Begin function equal_ra_dst_only
	.p2align	4
	.type	equal_ra_dst_only,@function
equal_ra_dst_only:                      # @equal_ra_dst_only
	.cfi_startproc
# %bb.0:
	subq	%rdx, %rcx
	setne	%al
	cmpq	%rsi, %rdi

===== equal_ra_src_only (simd=False, lines=121) =====
equal_ra_src_only:                      # @equal_ra_src_only
	.cfi_startproc
# %bb.0:
	subq	%rdi, %rsi
	sete	%al
	je	.LBB12_7
# %bb.1:
	cmpq	%rcx, %rdx
	je	.LBB12_7
# %bb.2:
	sarq	$2, %rsi
	decq	%rsi
	addq	$4, %rdx
	.p2align	4
.LBB12_3:                               # =>This Inner Loop Header: Depth=1
	movl	(%rdi), %eax
	cmpl	-4(%rdx), %eax
	jne	.LBB12_4
# %bb.5:                                #   in Loop: Header=BB12_3 Depth=1
	subq	$1, %rsi
	setb	%al
	jb	.LBB12_7
# %bb.6:                                #   in Loop: Header=BB12_3 Depth=1
	addq	$4, %rdi
	cmpq	%rcx, %rdx
	leaq	4(%rdx), %rdx
	jne	.LBB12_3
.LBB12_7:
                                        # kill: def $al killed $al killed $eax
	retq
.LBB12_4:
	xorl	%eax, %eax
                                        # kill: def $al killed $al killed $eax
	retq
.Lfunc_end12:
	.size	equal_ra_src_only, .Lfunc_end12-equal_ra_src_only
	.cfi_endproc
                                        # -- End function
	.globl	equal_ra_dst_only               # -- Begin function equal_ra_dst_only
	.p2align	4
	.type	equal_ra_dst_only,@function
equal_ra_dst_only:                      # @equal_ra_dst_only
	.cfi_startproc
# %bb.0:
	subq	%rdx, %rcx
	setne	%al
	cmpq	%rsi, %rdi
	setne	%r8b
	andb	%al, %r8b
	cmpb	$1, %r8b
	jne	.LBB13_6
# %bb.1:
	sarq	$2, %rcx
	decq	%rcx
	.p2align	4
.LBB13_2:                               # =>This Inner Loop Header: Depth=1
	movl	(%rdi), %eax
	cmpl	(%rdx), %eax
	jne	.LBB13_3
# %bb.4:                                #   in Loop: Header=BB13_2 Depth=1
	addq	$4, %rdi
	addq	$-1, %rcx
	jae	.LBB13_6
# %bb.5:                                #   in Loop: Header=BB13_2 Depth=1
	addq	$4, %rdx
	cmpq	%rsi, %rdi
	jne	.LBB13_2
.LBB13_6:
	cmpq	%rsi, %rdi
	sete	%al
                                        # kill: def $al killed $al killed $eax
	retq
.LBB13_3:
	xorl	%eax, %eax
                                        # kill: def $al killed $al killed $eax
	retq
.Lfunc_end13:
	.size	equal_ra_dst_only, .Lfunc_end13-equal_ra_dst_only
	.cfi_endproc
                                        # -- End function
	.globl	merge_ra_dst_block              # -- Begin function merge_ra_dst_block
	.p2align	4
	.type	merge_ra_dst_block,@function
merge_ra_dst_block:                     # @merge_ra_dst_block
	.cfi_startproc
# %bb.0:
	movq	%r8, %rax
	subq	%r8, %r9
	cmpq	$125, %r9
	setge	%r8b

===== equal_ra_dst_only (simd=False, lines=121) =====
equal_ra_dst_only:                      # @equal_ra_dst_only
	.cfi_startproc
# %bb.0:
	subq	%rdx, %rcx
	setne	%al
	cmpq	%rsi, %rdi
	setne	%r8b
	andb	%al, %r8b
	cmpb	$1, %r8b
	jne	.LBB13_6
# %bb.1:
	sarq	$2, %rcx
	decq	%rcx
	.p2align	4
.LBB13_2:                               # =>This Inner Loop Header: Depth=1
	movl	(%rdi), %eax
	cmpl	(%rdx), %eax
	jne	.LBB13_3
# %bb.4:                                #   in Loop: Header=BB13_2 Depth=1
	addq	$4, %rdi
	addq	$-1, %rcx
	jae	.LBB13_6
# %bb.5:                                #   in Loop: Header=BB13_2 Depth=1
	addq	$4, %rdx
	cmpq	%rsi, %rdi
	jne	.LBB13_2
.LBB13_6:
	cmpq	%rsi, %rdi
	sete	%al
                                        # kill: def $al killed $al killed $eax
	retq
.LBB13_3:
	xorl	%eax, %eax
                                        # kill: def $al killed $al killed $eax
	retq
.Lfunc_end13:
	.size	equal_ra_dst_only, .Lfunc_end13-equal_ra_dst_only
	.cfi_endproc
                                        # -- End function
	.globl	merge_ra_dst_block              # -- Begin function merge_ra_dst_block
	.p2align	4
	.type	merge_ra_dst_block,@function
merge_ra_dst_block:                     # @merge_ra_dst_block
	.cfi_startproc
# %bb.0:
	movq	%r8, %rax
	subq	%r8, %r9
	cmpq	$125, %r9
	setge	%r8b
	cmpq	%rsi, %rdi
	setne	%r10b
	cmpq	%rcx, %rdx
	setne	%r11b
	andb	%r10b, %r11b
	andb	%r8b, %r11b
	cmpb	$1, %r11b
	jne	.LBB14_6
# %bb.1:
	pushq	%rbp
	.cfi_def_cfa_offset 16
	pushq	%r15
	.cfi_def_cfa_offset 24
	pushq	%r14
	.cfi_def_cfa_offset 32
	pushq	%rbx
	.cfi_def_cfa_offset 40
	.cfi_offset %rbx, -40
	.cfi_offset %r14, -32
	.cfi_offset %r15, -24
	.cfi_offset %rbp, -16
	addq	$-128, %r9
	.p2align	4
.LBB14_2:                               # =>This Inner Loop Header: Depth=1
	movl	(%rdx), %r8d
	movl	(%rdi), %r10d
	xorl	%r11d, %r11d
	xorl	%ebx, %ebx
	cmpl	%r10d, %r8d
	setge	%r11b
	setl	%bl
	cmovll	%r8d, %r10d
	leaq	(%rdi,%r11,4), %r8
	leaq	(%rdx,%rbx,4), %r14
	movl	%r10d, (%rax)
	movl	(%rdx,%rbx,4), %edx
	movl	(%rdi,%r11,4), %edi
	xorl	%r10d, %r10d
	xorl	%r11d, %r11d
	cmpl	%edi, %edx
	setge	%r10b


