asm_broadcast.s

	.text
	.intel_syntax noprefix
	.file	"broadcast_dispatch_0"
	.section	.text.broadcast_dispatch_0_generic_Dx8640x3200_f16,"ax",@progbits
	.p2align	4, 0x90
	.type	broadcast_dispatch_0_generic_Dx8640x3200_f16,@function
broadcast_dispatch_0_generic_Dx8640x3200_f16:
.Lfunc_begin0:
	.file	1 "-"
	.loc	1 1 0
	.cfi_startproc
	push	rbp
	.cfi_def_cfa_offset 16
	.cfi_offset rbp, -16
	mov	rbp, rsp
	.cfi_def_cfa_register rbp
.Ltmp0:
	push	r15
	push	r14
	push	r13
	push	r12
	push	rbx
	.cfi_offset rbx, -56
	.cfi_offset r12, -48
	.cfi_offset r13, -40
	.cfi_offset r14, -32
	.cfi_offset r15, -24
	.loc	1 4 3 prologue_end
	mov	rax, qword ptr [rsi + 24]
	mov	ebx, dword ptr [rax + 4]
	mov	ecx, dword ptr [rax]
	movzx	eax, word ptr [rdx + 8]
	shl	rbx, 32
	mov	r8d, eax
	shl	r8d, 6
	lea	rdi, [rbx + rcx]
	mov	qword ptr [rbp - 48], r8
	mov	qword ptr [rbp - 64], rdi
	cmp	r8, rdi
	jge	.LBB0_16
	.loc	1 0 3 is_stmt 0
	mov	r11d, dword ptr [rdx]
	mov	edx, dword ptr [rdx + 4]
	mov	rdi, qword ptr [rsi + 32]
	movzx	r10d, word ptr [rsi + 20]
	mov	r9d, dword ptr [rsi + 16]
	.loc	1 4 3
	or	rbx, rcx
	mov	r8d, dword ptr [rsi + 12]
	mov	r14d, 3538944000
	sub	rbx, qword ptr [rbp - 48]
	imul	rax, r14
	mov	rsi, qword ptr [rdi]
	mov	rdi, qword ptr [rdi + 8]
	imul	rcx, rdx, 409600
	imul	r14, r10
	shl	r10d, 6
	mov	qword ptr [rbp - 120], rdx
	mov	r13, r8
	mov	qword ptr [rbp - 160], r11
	shl	r13, 6
	shl	r8, 7
	mov	qword ptr [rbp - 104], r10
	mov	r10, rdx
	mov	rdx, r11
	shl	rdx, 7
	shl	r10, 6
	shl	r11, 6
	mov	qword ptr [rbp - 96], r10
	imul	r10, r9, 409600
	shl	r9, 6
	mov	qword ptr [rbp - 136], r11
	add	rdx, rcx
	mov	qword ptr [rbp - 112], r14
	mov	qword ptr [rbp - 144], r9
	add	rax, rdx
	lea	rax, [rdi + rax + 96000]
	lea	rcx, [rsi + rdx + 96000]
	mov	qword ptr [rbp - 152], r10
	mov	qword ptr [rbp - 56], rax
	mov	qword ptr [rbp - 88], rcx
	jmp	.LBB0_2
	.p2align	4, 0x90
.LBB0_15:
	.loc	1 0 3
	mov	rdx, qword ptr [rbp - 56]
	mov	rax, qword ptr [rbp - 48]
	mov	rcx, qword ptr [rbp - 104]
	mov	rbx, qword ptr [rbp - 128]
	.loc	1 4 3
	add	rdx, qword ptr [rbp - 112]
	add	rax, rcx
	sub	rbx, rcx
	mov	qword ptr [rbp - 56], rdx
	mov	qword ptr [rbp - 48], rax
	cmp	rax, qword ptr [rbp - 64]
	jge	.LBB0_16
.LBB0_2:
	cmp	rbx, 64
	mov	r11d, 64
	mov	eax, 1
	mov	qword ptr [rbp - 128], rbx
	cmovl	r11, rbx
	cmp	r11, 2
	cmovl	r11, rax
	cmp	dword ptr [rbp - 120], 134
	ja	.LBB0_15
	.loc	1 0 3
	mov	rcx, qword ptr [rbp - 64]
	mov	rax, qword ptr [rbp - 88]
	mov	rdx, qword ptr [rbp - 56]
	mov	r10, qword ptr [rbp - 96]
	.loc	1 4 3
	sub	rcx, qword ptr [rbp - 48]
	mov	qword ptr [rbp - 80], rax
	mov	qword ptr [rbp - 72], rdx
	jmp	.LBB0_4
	.p2align	4, 0x90
.LBB0_14:
	.loc	1 0 3
	mov	rax, qword ptr [rbp - 152]
	.loc	1 4 3
	add	r10, qword ptr [rbp - 144]
	add	qword ptr [rbp - 72], rax
	add	qword ptr [rbp - 80], rax
	cmp	r10, 8640
	jge	.LBB0_15
.LBB0_4:
	.loc	1 0 3
	cmp	dword ptr [rbp - 160], 49
	.loc	1 4 3
	ja	.LBB0_14
	.loc	1 0 3
	mov	rdi, qword ptr [rbp - 80]
	mov	rax, qword ptr [rbp - 72]
	mov	r15, qword ptr [rbp - 136]
	jmp	.LBB0_6
	.p2align	4, 0x90
.LBB0_13:
	.loc	1 4 3
	add	r15, r13
	add	rax, r8
	add	rdi, r8
	cmp	r15, 3200
	jge	.LBB0_14
.LBB0_6:
	.loc	1 0 3
	test	rcx, rcx
	.loc	1 4 3
	jle	.LBB0_13
	.loc	1 0 3
	mov	r9, rax
	xor	esi, esi
	.p2align	4, 0x90
.LBB0_8:
	mov	rdx, rdi
	mov	r12, r9
	xor	r14d, r14d
	.p2align	4, 0x90
.LBB0_9:
	mov	rbx, -16
	.p2align	4, 0x90
.LBB0_10:
	.loc	1 4 3
	vmovaps	ymm0, ymmword ptr [rdx + 2*rbx - 95968]
	vmovaps	ymm1, ymmword ptr [rdx + 2*rbx - 89568]
	vmovaps	ymm2, ymmword ptr [rdx + 2*rbx - 83168]
	vmovaps	ymm3, ymmword ptr [rdx + 2*rbx - 76768]
	vmovaps	ymm4, ymmword ptr [rdx + 2*rbx - 70368]
	vmovaps	ymm5, ymmword ptr [rdx + 2*rbx - 63968]
	vmovaps	ymm6, ymmword ptr [rdx + 2*rbx - 57568]
	vmovaps	ymm7, ymmword ptr [rdx + 2*rbx - 51168]
	vmovaps	ymm8, ymmword ptr [rdx + 2*rbx - 44768]
	vmovaps	ymm9, ymmword ptr [rdx + 2*rbx - 38368]
	vmovaps	ymm10, ymmword ptr [rdx + 2*rbx - 31968]
	vmovaps	ymm11, ymmword ptr [rdx + 2*rbx - 25568]
	vmovaps	ymm12, ymmword ptr [rdx + 2*rbx - 19168]
	vmovaps	ymm13, ymmword ptr [rdx + 2*rbx - 12768]
	vmovaps	ymm14, ymmword ptr [rdx + 2*rbx - 6368]
	vmovaps	ymm15, ymmword ptr [rdx + 2*rbx + 32]
	vmovaps	ymmword ptr [r12 + 2*rbx - 95968], ymm0
	vmovaps	ymmword ptr [r12 + 2*rbx - 89568], ymm1
	vmovaps	ymmword ptr [r12 + 2*rbx - 83168], ymm2
	vmovaps	ymmword ptr [r12 + 2*rbx - 76768], ymm3
	vmovaps	ymmword ptr [r12 + 2*rbx - 70368], ymm4
	vmovaps	ymmword ptr [r12 + 2*rbx - 63968], ymm5
	vmovaps	ymmword ptr [r12 + 2*rbx - 57568], ymm6
	vmovaps	ymmword ptr [r12 + 2*rbx - 51168], ymm7
	vmovaps	ymmword ptr [r12 + 2*rbx - 44768], ymm8
	vmovaps	ymmword ptr [r12 + 2*rbx - 38368], ymm9
	vmovaps	ymmword ptr [r12 + 2*rbx - 31968], ymm10
	vmovaps	ymmword ptr [r12 + 2*rbx - 25568], ymm11
	vmovaps	ymmword ptr [r12 + 2*rbx - 19168], ymm12
	vmovaps	ymmword ptr [r12 + 2*rbx - 12768], ymm13
	vmovaps	ymmword ptr [r12 + 2*rbx - 6368], ymm14
	vmovaps	ymmword ptr [r12 + 2*rbx + 32], ymm15
	add	rbx, 16
	cmp	rbx, 48
	jb	.LBB0_10
	lea	rbx, [r14 + 16]
	add	r12, 102400
	add	rdx, 102400
	cmp	r14, 48
	mov	r14, rbx
	jb	.LBB0_9
	inc	rsi
	add	r9, 55296000
	cmp	rsi, r11
	jne	.LBB0_8
	jmp	.LBB0_13
.LBB0_16:
	xor	eax, eax
	.loc	1 4 3 epilogue_begin
	pop	rbx
	pop	r12
	pop	r13
	pop	r14
	pop	r15
	pop	rbp
	.cfi_def_cfa rsp, 8
	vzeroupper
	ret
.Ltmp1:
.Lfunc_end0:
	.size	broadcast_dispatch_0_generic_Dx8640x3200_f16, .Lfunc_end0-broadcast_dispatch_0_generic_Dx8640x3200_f16
	.cfi_endproc

	.section	.text.iree_hal_executable_library_query,"ax",@progbits
	.globl	iree_hal_executable_library_query
	.p2align	4, 0x90
	.type	iree_hal_executable_library_query,@function
iree_hal_executable_library_query:
.Liree_hal_executable_library_query$local:
	.type	.Liree_hal_executable_library_query$local,@function
.Lfunc_begin1:
	.cfi_startproc
	xor	eax, eax
	cmp	edi, 4
	lea	rcx, [rip + iree_hal_executable_library_query_v0]
	cmove	rax, rcx
	ret
.Lfunc_end1:
	.size	iree_hal_executable_library_query, .Lfunc_end1-iree_hal_executable_library_query
	.size	.Liree_hal_executable_library_query$local, .Lfunc_end1-iree_hal_executable_library_query
	.cfi_endproc

	.section	.text.iree_h2f_ieee,"ax",@progbits
	.p2align	4, 0x90
	.type	iree_h2f_ieee,@function
iree_h2f_ieee:
.Lfunc_begin2:
	.cfi_startproc
	mov	eax, edi
	and	eax, 32768
	mov	edx, edi
	mov	ecx, edi
	and	ecx, 1023
	shl	eax, 16
	and	dx, 31744
	je	.LBB2_6
	and	edi, 31744
	cmp	edi, 31744
	jne	.LBB2_5
	test	cx, cx
	je	.LBB2_4
	or	eax, 2143289344
	vmovd	xmm0, eax
	ret
.LBB2_6:
	movzx	ecx, cx
	or	eax, 864026624
	vcvtsi2ss	xmm0, xmm0, ecx
	vmovd	xmm1, eax
	vmulss	xmm0, xmm0, xmm1
	ret
.LBB2_5:
	movzx	ecx, cx
	movzx	edx, dx
	add	edx, ecx
	shl	edx, 13
	lea	eax, [rdx + rax + 939524096]
	vmovd	xmm0, eax
	ret
.LBB2_4:
	or	eax, 2139095040
	vmovd	xmm0, eax
	ret
.Lfunc_end2:
	.size	iree_h2f_ieee, .Lfunc_end2-iree_h2f_ieee
	.cfi_endproc

	.section	.text.iree_f2h_ieee,"ax",@progbits
	.p2align	4, 0x90
	.type	iree_f2h_ieee,@function
iree_f2h_ieee:
.Lfunc_begin3:
	.cfi_startproc
	vmovd	edi, xmm0
	mov	edx, 2071
	bextr	esi, edi, edx
	mov	eax, edi
	mov	ecx, edi
	shr	ecx, 31
	and	eax, 8388607
	lea	edx, [rsi - 112]
	lea	r8d, [rsi - 113]
	cmp	r8d, 28
	ja	.LBB3_2
	add	eax, 4096
	shl	ecx, 15
	shl	edx, 10
	shr	eax, 13
	or	eax, ecx
	or	eax, edx
	ret
.LBB3_2:
	test	edi, edi
	je	.LBB3_6
	cmp	esi, 112
	ja	.LBB3_7
	cmp	esi, 102
	jae	.LBB3_10
.LBB3_6:
	xor	eax, eax
	ret
.LBB3_7:
	cmp	edx, 143
	jne	.LBB3_11
	shl	ecx, 15
	test	eax, eax
	je	.LBB3_14
	shr	eax, 13
	or	ecx, eax
	or	ecx, 31744
	mov	eax, ecx
	ret
.LBB3_10:
	mov	dl, 113
	or	eax, 8388608
	shl	ecx, 15
	sub	dl, sil
	shrx	eax, eax, edx
	mov	edx, eax
	and	edx, 4096
	lea	eax, [rax + 2*rdx]
	shr	eax, 13
	or	eax, ecx
	ret
.LBB3_11:
	test	edi, 4096
	je	.LBB3_13
	lea	edi, [rax + 8192]
	add	esi, -111
	xor	r8d, r8d
	cmp	eax, 8380416
	cmovb	esi, edx
	cmovb	r8d, edi
	mov	eax, r8d
	mov	edx, esi
.LBB3_13:
	shl	ecx, 15
	cmp	edx, 31
	jb	.LBB3_15
.LBB3_14:
	or	ecx, 31744
	mov	eax, ecx
	ret
.LBB3_15:
	shr	eax, 13
	shl	edx, 10
	or	eax, ecx
	or	eax, edx
	ret
.Lfunc_end3:
	.size	iree_f2h_ieee, .Lfunc_end3-iree_f2h_ieee
	.cfi_endproc

	.section	.text.__gnu_h2f_ieee,"ax",@progbits
	.p2align	4, 0x90
	.type	__gnu_h2f_ieee,@function
__gnu_h2f_ieee:
.Lfunc_begin4:
	.cfi_startproc
	mov	eax, edi
	and	eax, 32768
	mov	edx, edi
	mov	ecx, edi
	and	ecx, 1023
	shl	eax, 16
	and	dx, 31744
	je	.LBB4_6
	and	edi, 31744
	cmp	edi, 31744
	jne	.LBB4_5
	test	cx, cx
	je	.LBB4_4
	or	eax, 2143289344
	vmovd	xmm0, eax
	ret
.LBB4_6:
	movzx	ecx, cx
	or	eax, 864026624
	vcvtsi2ss	xmm0, xmm0, ecx
	vmovd	xmm1, eax
	vmulss	xmm0, xmm0, xmm1
	ret
.LBB4_5:
	movzx	ecx, cx
	movzx	edx, dx
	add	edx, ecx
	shl	edx, 13
	lea	eax, [rdx + rax + 939524096]
	vmovd	xmm0, eax
	ret
.LBB4_4:
	or	eax, 2139095040
	vmovd	xmm0, eax
	ret
.Lfunc_end4:
	.size	__gnu_h2f_ieee, .Lfunc_end4-__gnu_h2f_ieee
	.cfi_endproc

	.section	.text.__extendhfsf2,"ax",@progbits
	.p2align	4, 0x90
	.type	__extendhfsf2,@function
__extendhfsf2:
.Lfunc_begin5:
	.cfi_startproc
	vmovd	ecx, xmm0
	mov	eax, ecx
	shl	eax, 16
	mov	edx, ecx
	and	edx, 1023
	mov	esi, ecx
	and	eax, -2147483648
	and	esi, 31744
	je	.LBB5_6
	cmp	esi, 31744
	jne	.LBB5_5
	test	dx, dx
	je	.LBB5_4
	or	eax, 2143289344
	vmovd	xmm0, eax
	ret
.LBB5_6:
	movzx	ecx, dx
	or	eax, 864026624
	vcvtsi2ss	xmm0, xmm1, ecx
	vmovd	xmm1, eax
	vmulss	xmm0, xmm0, xmm1
	ret
.LBB5_5:
	and	ecx, 32767
	shl	ecx, 13
	lea	eax, [rcx + rax + 939524096]
	vmovd	xmm0, eax
	ret
.LBB5_4:
	or	eax, 2139095040
	vmovd	xmm0, eax
	ret
.Lfunc_end5:
	.size	__extendhfsf2, .Lfunc_end5-__extendhfsf2
	.cfi_endproc

	.section	.text.__gnu_f2h_ieee,"ax",@progbits
	.p2align	4, 0x90
	.type	__gnu_f2h_ieee,@function
__gnu_f2h_ieee:
.Lfunc_begin6:
	.cfi_startproc
	vmovd	edi, xmm0
	mov	edx, 2071
	bextr	esi, edi, edx
	mov	eax, edi
	mov	ecx, edi
	shr	ecx, 31
	and	eax, 8388607
	lea	edx, [rsi - 112]
	lea	r8d, [rsi - 113]
	cmp	r8d, 28
	ja	.LBB6_2
	add	eax, 4096
	shl	ecx, 15
	shl	edx, 10
	shr	eax, 13
	or	eax, ecx
	or	eax, edx
	ret
.LBB6_2:
	test	edi, edi
	je	.LBB6_6
	cmp	esi, 112
	ja	.LBB6_7
	cmp	esi, 102
	jae	.LBB6_10
.LBB6_6:
	xor	eax, eax
	ret
.LBB6_7:
	cmp	edx, 143
	jne	.LBB6_11
	shl	ecx, 15
	test	eax, eax
	je	.LBB6_14
	shr	eax, 13
	or	ecx, eax
	or	ecx, 31744
	mov	eax, ecx
	ret
.LBB6_10:
	mov	dl, 113
	or	eax, 8388608
	shl	ecx, 15
	sub	dl, sil
	shrx	eax, eax, edx
	mov	edx, eax
	and	edx, 4096
	lea	eax, [rax + 2*rdx]
	shr	eax, 13
	or	eax, ecx
	ret
.LBB6_11:
	test	edi, 4096
	je	.LBB6_13
	lea	edi, [rax + 8192]
	add	esi, -111
	xor	r8d, r8d
	cmp	eax, 8380416
	cmovb	esi, edx
	cmovb	r8d, edi
	mov	eax, r8d
	mov	edx, esi
.LBB6_13:
	shl	ecx, 15
	cmp	edx, 31
	jb	.LBB6_15
.LBB6_14:
	or	ecx, 31744
	mov	eax, ecx
	ret
.LBB6_15:
	shr	eax, 13
	shl	edx, 10
	or	eax, ecx
	or	eax, edx
	ret
.Lfunc_end6:
	.size	__gnu_f2h_ieee, .Lfunc_end6-__gnu_f2h_ieee
	.cfi_endproc

	.section	.text.__truncsfhf2,"ax",@progbits
	.p2align	4, 0x90
	.type	__truncsfhf2,@function
__truncsfhf2:
.Lfunc_begin7:
	.cfi_startproc
	vmovd	edi, xmm0
	mov	edx, 2071
	bextr	esi, edi, edx
	mov	eax, edi
	mov	ecx, edi
	shr	ecx, 31
	and	eax, 8388607
	lea	edx, [rsi - 112]
	lea	r8d, [rsi - 113]
	cmp	r8d, 28
	ja	.LBB7_2
	add	eax, 4096
	shl	ecx, 15
	shl	edx, 10
	shr	eax, 13
.LBB7_16:
	or	eax, ecx
	or	eax, edx
	mov	word ptr [rsp - 4], ax
	vmovss	xmm0, dword ptr [rsp - 4]
	ret
.LBB7_2:
	test	edi, edi
	je	.LBB7_6
	cmp	esi, 112
	ja	.LBB7_7
	cmp	esi, 102
	jae	.LBB7_10
.LBB7_6:
	xor	eax, eax
	mov	word ptr [rsp - 4], ax
	vmovss	xmm0, dword ptr [rsp - 4]
	ret
.LBB7_7:
	cmp	edx, 143
	jne	.LBB7_11
	shl	ecx, 15
	test	eax, eax
	je	.LBB7_14
	shr	eax, 13
	or	ecx, eax
	jmp	.LBB7_14
.LBB7_10:
	mov	dl, 113
	or	eax, 8388608
	shl	ecx, 15
	sub	dl, sil
	shrx	eax, eax, edx
	mov	edx, eax
	and	edx, 4096
	lea	eax, [rax + 2*rdx]
	shr	eax, 13
	or	eax, ecx
	mov	word ptr [rsp - 4], ax
	vmovss	xmm0, dword ptr [rsp - 4]
	ret
.LBB7_11:
	test	edi, 4096
	je	.LBB7_13
	lea	edi, [rax + 8192]
	add	esi, -111
	xor	r8d, r8d
	cmp	eax, 8380416
	cmovb	esi, edx
	cmovb	r8d, edi
	mov	eax, r8d
	mov	edx, esi
.LBB7_13:
	shl	ecx, 15
	cmp	edx, 31
	jb	.LBB7_15
.LBB7_14:
	or	ecx, 31744
	mov	eax, ecx
	mov	word ptr [rsp - 4], ax
	vmovss	xmm0, dword ptr [rsp - 4]
	ret
.LBB7_15:
	shr	eax, 13
	shl	edx, 10
	jmp	.LBB7_16
.Lfunc_end7:
	.size	__truncsfhf2, .Lfunc_end7-__truncsfhf2
	.cfi_endproc

	.section	.rodata.cst4,"aM",@progbits,4
	.p2align	2, 0x0
.LCPI8_0:
	.long	0x7b800000
.LCPI8_1:
	.long	0x80000000
.LCPI8_2:
	.long	0x3f800000
	.section	.text.ceilf,"ax",@progbits
	.p2align	4, 0x90
	.type	ceilf,@function
ceilf:
.Lfunc_begin8:
	.cfi_startproc
	vmovd	eax, xmm0
	mov	ecx, 2071
	bextr	ecx, eax, ecx
	cmp	ecx, 149
	ja	.LBB8_7
	cmp	ecx, 127
	jb	.LBB8_4
	add	ecx, -127
	mov	edx, 8388607
	shrx	edx, edx, ecx
	test	edx, eax
	je	.LBB8_7
	vaddss	xmm0, xmm0, dword ptr [rip + .LCPI8_0]
	xor	esi, esi
	test	eax, eax
	mov	edi, -8388608
	cmovs	edx, esi
	sarx	ecx, edi, ecx
	add	edx, eax
	and	edx, ecx
	vmovss	dword ptr [rsp - 8], xmm0
	vmovd	xmm0, edx
	ret
.LBB8_4:
	vaddss	xmm1, xmm0, dword ptr [rip + .LCPI8_0]
	vmovss	dword ptr [rsp - 4], xmm1
	test	eax, eax
	js	.LBB8_5
	vmovss	xmm1, dword ptr [rip + .LCPI8_2]
	sete	al
	kmovd	k1, eax
	vmovss	xmm1 {k1}, xmm1, xmm0
	vmovaps	xmm0, xmm1
.LBB8_7:
	ret
.LBB8_5:
	vmovss	xmm0, dword ptr [rip + .LCPI8_1]
	ret
.Lfunc_end8:
	.size	ceilf, .Lfunc_end8-ceilf
	.cfi_endproc

	.section	.rodata.cst4,"aM",@progbits,4
	.p2align	2, 0x0
.LCPI9_0:
	.long	0x42b17217
.LCPI9_1:
	.long	0xc2cff1b4
.LCPI9_2:
	.long	0x10000000
.LCPI9_3:
	.long	0x70000000
	.section	.rodata.cst8,"aM",@progbits,8
	.p2align	3, 0x0
.LCPI9_4:
	.quad	0x40471547652b82fe
.LCPI9_5:
	.quad	0x4338000000000000
.LCPI9_6:
	.quad	0xc338000000000000
.LCPI9_7:
	.quad	0x3ebc6af84b912394
.LCPI9_8:
	.quad	0x3f2ebfce50fac4f3
.LCPI9_9:
	.quad	0x3f962e42ff0c52d6
.LCPI9_10:
	.quad	0x3ff0000000000000
	.section	.text.expf,"ax",@progbits
	.p2align	4, 0x90
	.type	expf,@function
expf:
.Lfunc_begin9:
	.cfi_startproc
	vmovd	ecx, xmm0
	mov	eax, 2836
	bextr	eax, ecx, eax
	cmp	eax, 1067
	jae	.LBB9_1
.LBB9_8:
	vcvtss2sd	xmm0, xmm0, xmm0
	vmulsd	xmm0, xmm0, qword ptr [rip + .LCPI9_4]
	lea	rdx, [rip + __exp2f_data]
	vaddsd	xmm1, xmm0, qword ptr [rip + .LCPI9_5]
	vmovq	rax, xmm1
	vaddsd	xmm1, xmm1, qword ptr [rip + .LCPI9_6]
	mov	ecx, eax
	and	ecx, 31
	shl	rax, 47
	add	rax, qword ptr [rdx + 8*rcx]
	vsubsd	xmm0, xmm0, xmm1
	vmulsd	xmm2, xmm0, qword ptr [rip + .LCPI9_7]
	vmovq	xmm1, rax
	vmulsd	xmm3, xmm0, xmm0
	vaddsd	xmm2, xmm2, qword ptr [rip + .LCPI9_8]
	vmulsd	xmm0, xmm0, qword ptr [rip + .LCPI9_9]
	vaddsd	xmm0, xmm0, qword ptr [rip + .LCPI9_10]
	vmulsd	xmm2, xmm3, xmm2
	vaddsd	xmm0, xmm0, xmm2
	vmulsd	xmm0, xmm0, xmm1
	vcvtsd2ss	xmm1, xmm0, xmm0
.LBB9_9:
	vmovaps	xmm0, xmm1
	ret
.LBB9_1:
	vxorps	xmm1, xmm1, xmm1
	cmp	ecx, -8388608
	je	.LBB9_9
	cmp	eax, 2040
	jae	.LBB9_3
	vucomiss	xmm0, dword ptr [rip + .LCPI9_0]
	jbe	.LBB9_6
	mov	dword ptr [rsp - 8], 1879048192
	vmovss	xmm0, dword ptr [rsp - 8]
	vmulss	xmm0, xmm0, dword ptr [rip + .LCPI9_3]
	ret
.LBB9_3:
	vaddss	xmm0, xmm0, xmm0
	ret
.LBB9_6:
	vmovss	xmm1, dword ptr [rip + .LCPI9_1]
	vucomiss	xmm1, xmm0
	jbe	.LBB9_8
	mov	dword ptr [rsp - 4], 268435456
	vmovss	xmm0, dword ptr [rsp - 4]
	vmulss	xmm0, xmm0, dword ptr [rip + .LCPI9_2]
	ret
.Lfunc_end9:
	.size	expf, .Lfunc_end9-expf
	.cfi_endproc

	.section	.rodata.cst8,"aM",@progbits,8
	.p2align	2, 0x0
.LCPI10_0:
	.long	0xf0000000
	.long	0x70000000
	.section	.rodata.cst4,"aM",@progbits,4
	.p2align	2, 0x0
.LCPI10_1:
	.long	0x70000000
	.section	.text.__math_oflowf,"ax",@progbits
	.p2align	4, 0x90
	.type	__math_oflowf,@function
__math_oflowf:
.Lfunc_begin10:
	.cfi_startproc
	xor	eax, eax
	test	edi, edi
	lea	rcx, [rip + .LCPI10_0]
	sete	al
	vmovss	xmm0, dword ptr [rcx + 4*rax]
	vmovss	dword ptr [rsp - 4], xmm0
	vmovss	xmm0, dword ptr [rsp - 4]
	vmulss	xmm0, xmm0, dword ptr [rip + .LCPI10_1]
	ret
.Lfunc_end10:
	.size	__math_oflowf, .Lfunc_end10-__math_oflowf
	.cfi_endproc

	.section	.rodata.cst8,"aM",@progbits,8
	.p2align	2, 0x0
.LCPI11_0:
	.long	0x90000000
	.long	0x10000000
	.section	.rodata.cst4,"aM",@progbits,4
	.p2align	2, 0x0
.LCPI11_1:
	.long	0x10000000
	.section	.text.__math_uflowf,"ax",@progbits
	.p2align	4, 0x90
	.type	__math_uflowf,@function
__math_uflowf:
.Lfunc_begin11:
	.cfi_startproc
	xor	eax, eax
	test	edi, edi
	lea	rcx, [rip + .LCPI11_0]
	sete	al
	vmovss	xmm0, dword ptr [rcx + 4*rax]
	vmovss	dword ptr [rsp - 4], xmm0
	vmovss	xmm0, dword ptr [rsp - 4]
	vmulss	xmm0, xmm0, dword ptr [rip + .LCPI11_1]
	ret
.Lfunc_end11:
	.size	__math_uflowf, .Lfunc_end11-__math_uflowf
	.cfi_endproc

	.section	.rodata.cst4,"aM",@progbits,4
	.p2align	2, 0x0
.LCPI12_0:
	.long	0x80000000
	.section	.text.__math_xflowf,"ax",@progbits
	.p2align	4, 0x90
	.type	__math_xflowf,@function
__math_xflowf:
.Lfunc_begin12:
	.cfi_startproc
	vxorps	xmm1, xmm0, dword ptr [rip + .LCPI12_0]{1to4}
	test	edi, edi
	sete	al
	kmovd	k1, eax
	vmovss	xmm1 {k1}, xmm1, xmm0
	vmovss	dword ptr [rsp - 4], xmm1
	vmulss	xmm0, xmm0, dword ptr [rsp - 4]
	ret
.Lfunc_end12:
	.size	__math_xflowf, .Lfunc_end12-__math_xflowf
	.cfi_endproc

	.section	.text.feclearexcept,"ax",@progbits
	.p2align	4, 0x90
	.type	feclearexcept,@function
feclearexcept:
.Lfunc_begin13:
	.cfi_startproc
	xor	eax, eax
	ret
.Lfunc_end13:
	.size	feclearexcept, .Lfunc_end13-feclearexcept
	.cfi_endproc

	.section	.text.feraiseexcept,"ax",@progbits
	.p2align	4, 0x90
	.type	feraiseexcept,@function
feraiseexcept:
.Lfunc_begin14:
	.cfi_startproc
	xor	eax, eax
	ret
.Lfunc_end14:
	.size	feraiseexcept, .Lfunc_end14-feraiseexcept
	.cfi_endproc

	.section	.text.fetestexcept,"ax",@progbits
	.p2align	4, 0x90
	.type	fetestexcept,@function
fetestexcept:
.Lfunc_begin15:
	.cfi_startproc
	xor	eax, eax
	ret
.Lfunc_end15:
	.size	fetestexcept, .Lfunc_end15-fetestexcept
	.cfi_endproc

	.section	.text.fegetround,"ax",@progbits
	.p2align	4, 0x90
	.type	fegetround,@function
fegetround:
.Lfunc_begin16:
	.cfi_startproc
	xor	eax, eax
	ret
.Lfunc_end16:
	.size	fegetround, .Lfunc_end16-fegetround
	.cfi_endproc

	.section	.text.__fesetround,"ax",@progbits
	.p2align	4, 0x90
	.type	__fesetround,@function
__fesetround:
.Lfunc_begin17:
	.cfi_startproc
	xor	eax, eax
	ret
.Lfunc_end17:
	.size	__fesetround, .Lfunc_end17-__fesetround
	.cfi_endproc

	.section	.text.fegetenv,"ax",@progbits
	.p2align	4, 0x90
	.type	fegetenv,@function
fegetenv:
.Lfunc_begin18:
	.cfi_startproc
	xor	eax, eax
	ret
.Lfunc_end18:
	.size	fegetenv, .Lfunc_end18-fegetenv
	.cfi_endproc

	.section	.text.fesetenv,"ax",@progbits
	.p2align	4, 0x90
	.type	fesetenv,@function
fesetenv:
.Lfunc_begin19:
	.cfi_startproc
	xor	eax, eax
	ret
.Lfunc_end19:
	.size	fesetenv, .Lfunc_end19-fesetenv
	.cfi_endproc

	.section	.rodata.cst4,"aM",@progbits,4
	.p2align	2, 0x0
.LCPI20_0:
	.long	0x7b800000
.LCPI20_1:
	.long	0xbf800000
	.section	.text.floorf,"ax",@progbits
	.p2align	4, 0x90
	.type	floorf,@function
floorf:
.Lfunc_begin20:
	.cfi_startproc
	vmovd	eax, xmm0
	mov	ecx, 2071
	bextr	ecx, eax, ecx
	cmp	ecx, 149
	jbe	.LBB20_1
	ret
.LBB20_1:
	cmp	ecx, 127
	jb	.LBB20_4
	add	ecx, -127
	mov	edx, 8388607
	shrx	edx, edx, ecx
	test	edx, eax
	je	.LBB20_6
	vaddss	xmm0, xmm0, dword ptr [rip + .LCPI20_0]
	mov	esi, -8388608
	sarx	ecx, esi, ecx
	mov	esi, eax
	sar	esi, 31
	and	esi, edx
	add	esi, eax
	and	esi, ecx
	vmovss	dword ptr [rsp - 8], xmm0
	vmovd	xmm0, esi
	ret
.LBB20_4:
	vaddss	xmm1, xmm0, dword ptr [rip + .LCPI20_0]
	vmovss	dword ptr [rsp - 4], xmm1
	vxorps	xmm1, xmm1, xmm1
	test	eax, eax
	js	.LBB20_7
	vmovaps	xmm0, xmm1
.LBB20_6:
	ret
.LBB20_7:
	vcmpeqss	k1, xmm0, xmm1
	vmovss	xmm1, dword ptr [rip + .LCPI20_1]
	vmovss	xmm1 {k1}, xmm1, xmm0
	vmovaps	xmm0, xmm1
	ret
.Lfunc_end20:
	.size	floorf, .Lfunc_end20-floorf
	.cfi_endproc

	.section	.text.fmaf,"ax",@progbits
	.p2align	4, 0x90
	.type	fmaf,@function
fmaf:
.Lfunc_begin21:
	.cfi_startproc
	vcvtss2sd	xmm0, xmm0, xmm0
	movabs	rdx, 9218868437227405312
	vcvtss2sd	xmm1, xmm1, xmm1
	vcvtss2sd	xmm2, xmm2, xmm2
	vmulsd	xmm1, xmm0, xmm1
	vaddsd	xmm0, xmm1, xmm2
	vmovq	rax, xmm0
	mov	ecx, eax
	and	ecx, 536870911
	cmp	ecx, 268435456
	setne	cl
	andn	rdx, rax, rdx
	sete	dl
	or	dl, cl
	jne	.LBB21_4
	vsubsd	xmm3, xmm0, xmm1
	vucomisd	xmm3, xmm2
	jne	.LBB21_3
	jp	.LBB21_3
	vsubsd	xmm3, xmm0, xmm2
	vucomisd	xmm3, xmm1
	jne	.LBB21_3
	jp	.LBB21_3
.LBB21_4:
	vcvtsd2ss	xmm0, xmm0, xmm0
	ret
.LBB21_3:
	test	rax, rax
	vsubsd	xmm3, xmm1, xmm0
	vsubsd	xmm0, xmm2, xmm0
	sets	cl
	vucomisd	xmm2, xmm1
	vaddsd	xmm0, xmm1, xmm0
	vaddsd	xmm3, xmm3, xmm2
	vxorpd	xmm1, xmm1, xmm1
	setbe	dl
	xor	dl, cl
	kmovd	k1, edx
	vmovsd	xmm0 {k1}, xmm0, xmm3
	vucomisd	xmm1, xmm0
	setbe	dl
	xor	dl, cl
	movzx	ecx, dl
	dec	rcx
	or	rcx, 1
	add	rcx, rax
	vmovq	xmm0, rcx
	vcvtsd2ss	xmm0, xmm0, xmm0
	ret
.Lfunc_end21:
	.size	fmaf, .Lfunc_end21-fmaf
	.cfi_endproc

	.section	.text.fmodf,"ax",@progbits
	.p2align	4, 0x90
	.type	fmodf,@function
fmodf:
.Lfunc_begin22:
	.cfi_startproc
	vmovd	edx, xmm1
	mov	esi, edx
	add	esi, edx
	je	.LBB22_2
	mov	r8d, edx
	vmovd	eax, xmm0
	mov	edi, 2071
	and	r8d, 2147483647
	bextr	ecx, eax, edi
	cmp	r8d, 2139095041
	setae	r8b
	cmp	ecx, 255
	sete	r9b
	or	r9b, r8b
	cmp	r9b, 1
	jne	.LBB22_3
.LBB22_2:
	vmulss	xmm0, xmm0, xmm1
	vdivss	xmm0, xmm0, xmm0
	ret
.LBB22_3:
	lea	r8d, [rax + rax]
	cmp	r8d, esi
	jbe	.LBB22_4
	bextr	edi, edx, edi
	test	ecx, ecx
	je	.LBB22_6
	mov	esi, eax
	and	esi, 8388607
	or	esi, 8388608
	test	edi, edi
	je	.LBB22_11
.LBB22_14:
	and	edx, 8388607
	or	edx, 8388608
	cmp	ecx, edi
	jg	.LBB22_16
.LBB22_21:
	mov	edi, esi
	sub	edi, edx
	jns	.LBB22_22
	jmp	.LBB22_23
.LBB22_4:
	vpxor	xmm1, xmm1, xmm1
	sete	al
	vmulss	xmm1, xmm0, xmm1
	kmovd	k1, eax
	vmovss	xmm0 {k1}, xmm0, xmm1
	ret
.LBB22_6:
	mov	esi, eax
	xor	ecx, ecx
	shl	esi, 9
	js	.LBB22_8
	.p2align	4, 0x90
.LBB22_7:
	dec	ecx
	add	esi, esi
	jns	.LBB22_7
.LBB22_8:
	mov	sil, 1
	sub	sil, cl
	shlx	esi, eax, esi
	test	edi, edi
	jne	.LBB22_14
.LBB22_11:
	mov	r8d, edx
	xor	edi, edi
	shl	r8d, 9
	js	.LBB22_13
	.p2align	4, 0x90
.LBB22_12:
	dec	edi
	add	r8d, r8d
	jns	.LBB22_12
.LBB22_13:
	mov	r8b, 1
	sub	r8b, dil
	shlx	edx, edx, r8d
	cmp	ecx, edi
	jg	.LBB22_16
	jmp	.LBB22_21
	.p2align	4, 0x90
.LBB22_19:
	add	esi, esi
	dec	ecx
	cmp	ecx, edi
	jle	.LBB22_20
.LBB22_16:
	mov	r8d, esi
	sub	r8d, edx
	js	.LBB22_19
	mov	esi, r8d
	jne	.LBB22_19
	jmp	.LBB22_18
.LBB22_20:
	mov	ecx, edi
	mov	edi, esi
	sub	edi, edx
	js	.LBB22_23
.LBB22_22:
	mov	esi, edi
	je	.LBB22_18
.LBB22_23:
	cmp	esi, 8388607
	ja	.LBB22_24
	.p2align	4, 0x90
.LBB22_25:
	lea	edx, [rsi + rsi]
	dec	ecx
	cmp	esi, 4194304
	mov	esi, edx
	jb	.LBB22_25
	and	eax, -2147483648
	test	ecx, ecx
	jle	.LBB22_28
.LBB22_27:
	add	edx, -8388608
	shl	ecx, 23
	or	ecx, edx
	or	ecx, eax
	vmovd	xmm0, ecx
	ret
.LBB22_18:
	vpxor	xmm1, xmm1, xmm1
	vmulss	xmm0, xmm0, xmm1
	ret
.LBB22_24:
	mov	edx, esi
	and	eax, -2147483648
	test	ecx, ecx
	jg	.LBB22_27
.LBB22_28:
	mov	sil, 1
	sub	sil, cl
	shrx	ecx, edx, esi
	or	ecx, eax
	vmovd	xmm0, ecx
	ret
.Lfunc_end22:
	.size	fmodf, .Lfunc_end22-fmodf
	.cfi_endproc

	.section	.text.__math_invalidf,"ax",@progbits
	.p2align	4, 0x90
	.type	__math_invalidf,@function
__math_invalidf:
.Lfunc_begin23:
	.cfi_startproc
	vsubss	xmm0, xmm0, xmm0
	vdivss	xmm0, xmm0, xmm0
	ret
.Lfunc_end23:
	.size	__math_invalidf, .Lfunc_end23-__math_invalidf
	.cfi_endproc

	.section	.rodata.cst4,"aM",@progbits,4
	.p2align	2, 0x0
.LCPI24_0:
	.long	0x3f800000
.LCPI24_1:
	.long	0x80000000
.LCPI24_2:
	.long	0x4b000000
.LCPI24_12:
	.long	0x10000000
.LCPI24_20:
	.long	0x70000000
	.section	.rodata.cst8,"aM",@progbits,8
	.p2align	3, 0x0
.LCPI24_3:
	.quad	0xbff0000000000000
.LCPI24_4:
	.quad	0x3fd27616c9496e0b
.LCPI24_5:
	.quad	0xbfd71969a075c67a
.LCPI24_6:
	.quad	0x3fdec70a6ca7badd
.LCPI24_7:
	.quad	0xbfe7154748bef6c8
.LCPI24_8:
	.quad	0x3ff71547652ab82b
.LCPI24_9:
	.quad	0x405fffffffd1d571
.LCPI24_10:
	.quad	0xc062c00000000000
.LCPI24_11:
	.long	0x90000000
	.long	0x10000000
.LCPI24_13:
	.quad	0x42e8000000000000
.LCPI24_14:
	.quad	0xc2e8000000000000
.LCPI24_15:
	.quad	0x3fac6af84b912394
.LCPI24_16:
	.quad	0x3fcebfce50fac4f3
.LCPI24_17:
	.quad	0x3fe62e42ff0c52d6
.LCPI24_18:
	.quad	0x3ff0000000000000
.LCPI24_19:
	.long	0xf0000000
	.long	0x70000000
	.section	.text.powf,"ax",@progbits
	.p2align	4, 0x90
	.type	powf,@function
powf:
.Lfunc_begin24:
	.cfi_startproc
	vmovd	edx, xmm0
	vmovd	eax, xmm1
	lea	ecx, [rdx - 2139095040]
	cmp	ecx, -2130706432
	jb	.LBB24_2
	lea	esi, [rax + rax + 16777216]
	xor	ecx, ecx
	cmp	esi, 16777216
	jbe	.LBB24_2
.LBB24_24:
	lea	eax, [rdx - 1060306944]
	mov	esi, eax
	mov	edi, eax
	shr	esi, 19
	and	edi, -8388608
	sar	eax, 23
	sub	edx, edi
	shl	esi, 4
	lea	rdi, [rip + __powf_log2_data]
	vmovd	xmm0, edx
	movzx	esi, sil
	movabs	rdx, 9223231299366420480
	vcvtss2sd	xmm0, xmm0, xmm0
	vmulsd	xmm0, xmm0, qword ptr [rsi + rdi]
	vaddsd	xmm0, xmm0, qword ptr [rip + .LCPI24_3]
	vmulsd	xmm5, xmm0, qword ptr [rip + .LCPI24_6]
	vmulsd	xmm4, xmm0, qword ptr [rip + .LCPI24_4]
	vaddsd	xmm5, xmm5, qword ptr [rip + .LCPI24_7]
	vaddsd	xmm4, xmm4, qword ptr [rip + .LCPI24_5]
	vcvtsi2sd	xmm2, xmm2, eax
	vmulsd	xmm3, xmm0, xmm0
	vaddsd	xmm2, xmm2, qword ptr [rsi + rdi + 8]
	vmulsd	xmm0, xmm0, qword ptr [rip + .LCPI24_8]
	movabs	rsi, 4638426141214900225
	vmulsd	xmm6, xmm3, xmm3
	vmulsd	xmm3, xmm3, xmm5
	vmulsd	xmm4, xmm4, xmm6
	vaddsd	xmm0, xmm2, xmm0
	vaddsd	xmm0, xmm0, xmm3
	vaddsd	xmm0, xmm4, xmm0
	vcvtss2sd	xmm1, xmm1, xmm1
	vmulsd	xmm0, xmm0, xmm1
	vmovq	rax, xmm0
	and	rdx, rax
	cmp	rdx, rsi
	jae	.LBB24_25
.LBB24_29:
	vaddsd	xmm1, xmm0, qword ptr [rip + .LCPI24_13]
	lea	rdx, [rip + __exp2f_data]
	vmovq	rax, xmm1
	vaddsd	xmm1, xmm1, qword ptr [rip + .LCPI24_14]
	add	ecx, eax
	and	eax, 31
	shl	rcx, 47
	add	rcx, qword ptr [rdx + 8*rax]
	vsubsd	xmm0, xmm0, xmm1
	vmulsd	xmm2, xmm0, qword ptr [rip + .LCPI24_15]
	vmovq	xmm1, rcx
	vmulsd	xmm3, xmm0, xmm0
	vaddsd	xmm2, xmm2, qword ptr [rip + .LCPI24_16]
	vmulsd	xmm0, xmm0, qword ptr [rip + .LCPI24_17]
	vaddsd	xmm0, xmm0, qword ptr [rip + .LCPI24_18]
	vmulsd	xmm2, xmm3, xmm2
	vaddsd	xmm0, xmm0, xmm2
	vmulsd	xmm0, xmm0, xmm1
	vcvtsd2ss	xmm0, xmm0, xmm0
.LBB24_30:
	ret
.LBB24_2:
	lea	ecx, [rax + rax]
	lea	esi, [rcx - 1]
	cmp	esi, -16777217
	jae	.LBB24_3
	lea	ecx, [rdx + rdx - 1]
	cmp	ecx, -16777217
	jae	.LBB24_10
	xor	ecx, ecx
	test	edx, edx
	js	.LBB24_16
	cmp	edx, 8388607
	ja	.LBB24_24
.LBB24_23:
	vmulss	xmm0, xmm0, dword ptr [rip + .LCPI24_2]
	vmovd	edx, xmm0
	and	edx, 2147483647
	add	edx, -192937984
	jmp	.LBB24_24
.LBB24_25:
	vucomisd	xmm0, qword ptr [rip + .LCPI24_9]
	jbe	.LBB24_27
	xor	eax, eax
	test	ecx, ecx
	lea	rcx, [rip + .LCPI24_19]
	sete	al
	vmovss	xmm0, dword ptr [rcx + 4*rax]
	vmovss	dword ptr [rsp - 8], xmm0
	vmovss	xmm0, dword ptr [rsp - 8]
	vmulss	xmm0, xmm0, dword ptr [rip + .LCPI24_20]
	ret
.LBB24_16:
	mov	ecx, 2071
	bextr	ecx, eax, ecx
	cmp	ecx, 127
	jb	.LBB24_31
	cmp	ecx, 150
	jbe	.LBB24_18
.LBB24_20:
	xor	ecx, ecx
.LBB24_21:
	vmovd	edx, xmm0
	and	edx, 2147483647
	cmp	edx, 8388607
	ja	.LBB24_24
	jmp	.LBB24_23
.LBB24_27:
	vmovsd	xmm1, qword ptr [rip + .LCPI24_10]
	vucomisd	xmm1, xmm0
	jb	.LBB24_29
	xor	eax, eax
	test	ecx, ecx
	lea	rcx, [rip + .LCPI24_11]
	sete	al
	vmovss	xmm0, dword ptr [rcx + 4*rax]
	vmovss	dword ptr [rsp - 4], xmm0
	vmovss	xmm0, dword ptr [rsp - 4]
	vmulss	xmm0, xmm0, dword ptr [rip + .LCPI24_12]
	ret
.LBB24_18:
	mov	dl, -106
	sub	dl, cl
	bzhi	ecx, eax, edx
	je	.LBB24_19
.LBB24_31:
	vsubss	xmm0, xmm0, xmm0
	vdivss	xmm0, xmm0, xmm0
	ret
.LBB24_19:
	mov	ecx, 1
	shlx	edx, ecx, edx
	mov	ecx, 65536
	test	edx, eax
	jne	.LBB24_21
	jmp	.LBB24_20
.LBB24_3:
	vmovss	xmm2, dword ptr [rip + .LCPI24_0]
	test	ecx, ecx
	sete	sil
	cmp	edx, 1065353216
	sete	dil
	or	dil, sil
	je	.LBB24_5
	vmovaps	xmm0, xmm2
	ret
.LBB24_10:
	vmulss	xmm0, xmm0, xmm0
	test	edx, edx
	jns	.LBB24_13
	mov	ecx, 2071
	bextr	ecx, eax, ecx
	lea	edx, [rcx - 151]
	cmp	edx, -24
	jb	.LBB24_13
	vxorps	xmm1, xmm0, dword ptr [rip + .LCPI24_1]{1to4}
	mov	dl, -106
	sub	dl, cl
	bzhi	ecx, eax, edx
	movzx	edx, dl
	setne	cl
	bt	eax, edx
	setae	dl
	kmovd	k1, ecx
	kmovd	k2, edx
	vmovss	xmm1 {k2}, xmm1, xmm0
	vmovss	xmm1 {k1}, xmm1, xmm0
	vmovaps	xmm0, xmm1
.LBB24_13:
	test	eax, eax
	jns	.LBB24_30
	vmovss	xmm1, dword ptr [rip + .LCPI24_0]
	vdivss	xmm0, xmm1, xmm0
	vmovss	dword ptr [rsp - 12], xmm0
	vmovss	xmm0, dword ptr [rsp - 12]
	ret
.LBB24_5:
	add	edx, edx
	cmp	edx, -16777215
	setae	sil
	cmp	ecx, -16777215
	setae	cl
	or	cl, sil
	cmp	cl, 1
	jne	.LBB24_7
	vaddss	xmm0, xmm0, xmm1
	ret
.LBB24_7:
	vmovaps	xmm0, xmm2
	cmp	edx, 2130706432
	je	.LBB24_30
	setae	cl
	test	eax, eax
	vmulss	xmm0, xmm1, xmm1
	vxorps	xmm1, xmm1, xmm1
	setns	al
	xor	al, cl
	kmovd	k1, eax
	vmovss	xmm0 {k1}, xmm0, xmm1
	ret
.Lfunc_end24:
	.size	powf, .Lfunc_end24-powf
	.cfi_endproc

	.section	.rodata.cst4,"aM",@progbits,4
	.p2align	2, 0x0
.LCPI25_0:
	.long	0x7fffffff
.LCPI25_1:
	.long	0x4b000000
.LCPI25_2:
	.long	0xcb000000
.LCPI25_3:
	.long	0x3f000000
.LCPI25_4:
	.long	0xbf000000
.LCPI25_5:
	.long	0x3f800000
.LCPI25_6:
	.long	0xbf800000
.LCPI25_7:
	.long	0x80000000
	.section	.text.roundf,"ax",@progbits
	.p2align	4, 0x90
	.type	roundf,@function
roundf:
.Lfunc_begin25:
	.cfi_startproc
	vmovd	eax, xmm0
	mov	ecx, 2071
	bextr	ecx, eax, ecx
	cmp	ecx, 149
	ja	.LBB25_8
	vpandd	xmm1, xmm0, dword ptr [rip + .LCPI25_0]{1to4}
	vaddss	xmm2, xmm1, dword ptr [rip + .LCPI25_1]
	cmp	ecx, 125
	ja	.LBB25_3
	vxorps	xmm1, xmm1, xmm1
	vmovss	dword ptr [rsp - 4], xmm2
	vmulss	xmm0, xmm0, xmm1
	ret
.LBB25_3:
	vaddss	xmm0, xmm2, dword ptr [rip + .LCPI25_2]
	vsubss	xmm0, xmm0, xmm1
	vucomiss	xmm0, dword ptr [rip + .LCPI25_3]
	jbe	.LBB25_5
	vaddss	xmm0, xmm1, xmm0
	vaddss	xmm0, xmm0, dword ptr [rip + .LCPI25_6]
	jmp	.LBB25_7
.LBB25_5:
	vmovss	xmm2, dword ptr [rip + .LCPI25_4]
	vucomiss	xmm2, xmm0
	vaddss	xmm0, xmm1, xmm0
	jb	.LBB25_7
	vaddss	xmm0, xmm0, dword ptr [rip + .LCPI25_5]
.LBB25_7:
	vxorps	xmm1, xmm0, dword ptr [rip + .LCPI25_7]{1to4}
	test	eax, eax
	sets	al
	kmovd	k1, eax
	vmovss	xmm0 {k1}, xmm0, xmm1
.LBB25_8:
	ret
.Lfunc_end25:
	.size	roundf, .Lfunc_end25-roundf
	.cfi_endproc

	.type	__unnamed_1,@object
	.section	.rodata.__unnamed_1,"a",@progbits
__unnamed_1:
	.asciz	"broadcast_dispatch_0"
	.size	__unnamed_1, 21

	.type	iree_hal_executable_library_query_v0_header,@object
	.section	.data.rel.ro.iree_hal_executable_library_query_v0_header,"aw",@progbits
	.p2align	4, 0x0
iree_hal_executable_library_query_v0_header:
	.long	4
	.zero	4
	.quad	__unnamed_1
	.long	0
	.long	0
	.size	iree_hal_executable_library_query_v0_header, 24

	.type	iree_hal_executable_library_query_v0_funcs,@object
	.section	.data.rel.ro.iree_hal_executable_library_query_v0_funcs,"aw",@progbits
	.p2align	3, 0x0
iree_hal_executable_library_query_v0_funcs:
	.quad	broadcast_dispatch_0_generic_Dx8640x3200_f16
	.size	iree_hal_executable_library_query_v0_funcs, 8

	.type	iree_hal_executable_library_query_v0_attrs,@object
	.section	.rodata.iree_hal_executable_library_query_v0_attrs,"a",@progbits
	.p2align	3, 0x0
iree_hal_executable_library_query_v0_attrs:
	.zero	4
	.size	iree_hal_executable_library_query_v0_attrs, 4

	.type	__unnamed_2,@object
	.section	.rodata.__unnamed_2,"a",@progbits
__unnamed_2:
	.asciz	"broadcast_dispatch_0_generic_Dx8640x3200_f16"
	.size	__unnamed_2, 45

	.type	iree_hal_executable_library_query_v0_names,@object
	.section	.data.rel.ro.iree_hal_executable_library_query_v0_names,"aw",@progbits
	.p2align	3, 0x0
iree_hal_executable_library_query_v0_names:
	.quad	__unnamed_2
	.size	iree_hal_executable_library_query_v0_names, 8

	.type	__unnamed_3,@object
	.section	.rodata.__unnamed_3,"a",@progbits
__unnamed_3:
	.asciz	"broadcast.mlir"
	.size	__unnamed_3, 15

	.type	iree_hal_executable_library_query_v0_source_locations,@object
	.section	.data.rel.ro.iree_hal_executable_library_query_v0_source_locations,"aw",@progbits
	.p2align	3, 0x0
iree_hal_executable_library_query_v0_source_locations:
	.long	4
	.long	14
	.quad	__unnamed_3
	.size	iree_hal_executable_library_query_v0_source_locations, 16

	.type	iree_hal_executable_library_query_v0_broadcast_dispatch_0_generic_Dx8640x3200_f16_stage_names,@object
	.section	.rodata.iree_hal_executable_library_query_v0_broadcast_dispatch_0_generic_Dx8640x3200_f16_stage_names,"a",@progbits
	.p2align	3, 0x0
iree_hal_executable_library_query_v0_broadcast_dispatch_0_generic_Dx8640x3200_f16_stage_names:
	.size	iree_hal_executable_library_query_v0_broadcast_dispatch_0_generic_Dx8640x3200_f16_stage_names, 0

	.type	iree_hal_executable_library_query_v0_broadcast_dispatch_0_generic_Dx8640x3200_f16_stage_source_locations,@object
	.section	.rodata.iree_hal_executable_library_query_v0_broadcast_dispatch_0_generic_Dx8640x3200_f16_stage_source_locations,"a",@progbits
	.p2align	3, 0x0
iree_hal_executable_library_query_v0_broadcast_dispatch_0_generic_Dx8640x3200_f16_stage_source_locations:
	.size	iree_hal_executable_library_query_v0_broadcast_dispatch_0_generic_Dx8640x3200_f16_stage_source_locations, 0

	.type	iree_hal_executable_library_query_v0_stage_location_tables,@object
	.section	.data.rel.ro.iree_hal_executable_library_query_v0_stage_location_tables,"aw",@progbits
	.p2align	4, 0x0
iree_hal_executable_library_query_v0_stage_location_tables:
	.long	0
	.zero	4
	.quad	iree_hal_executable_library_query_v0_broadcast_dispatch_0_generic_Dx8640x3200_f16_stage_names
	.quad	iree_hal_executable_library_query_v0_broadcast_dispatch_0_generic_Dx8640x3200_f16_stage_source_locations
	.size	iree_hal_executable_library_query_v0_stage_location_tables, 24

	.type	iree_hal_executable_library_query_v0,@object
	.section	.data.rel.ro.iree_hal_executable_library_query_v0,"aw",@progbits
	.p2align	4, 0x0
iree_hal_executable_library_query_v0:
	.quad	iree_hal_executable_library_query_v0_header
	.zero	16
	.long	1
	.zero	4
	.quad	iree_hal_executable_library_query_v0_funcs
	.quad	iree_hal_executable_library_query_v0_attrs
	.quad	iree_hal_executable_library_query_v0_names
	.quad	0
	.quad	iree_hal_executable_library_query_v0_source_locations
	.quad	iree_hal_executable_library_query_v0_stage_location_tables
	.zero	4
	.zero	4
	.zero	16
	.size	iree_hal_executable_library_query_v0, 104

	.type	__powf_log2_data,@object
	.section	.rodata.__powf_log2_data,"a",@progbits
	.p2align	3, 0x0
__powf_log2_data:
	.quad	0x3ff661ec79f8f3be
	.quad	0xbfdefec65b963019
	.quad	0x3ff571ed4aaf883d
	.quad	0xbfdb0b6832d4fca4
	.quad	0x3ff49539f0f010b0
	.quad	0xbfd7418b0a1fb77b
	.quad	0x3ff3c995b0b80385
	.quad	0xbfd39de91a6dcf7b
	.quad	0x3ff30d190c8864a5
	.quad	0xbfd01d9bf3f2b631
	.quad	0x3ff25e227b0b8ea0
	.quad	0xbfc97c1d1b3b7af0
	.quad	0x3ff1bb4a4a1a343f
	.quad	0xbfc2f9e393af3c9f
	.quad	0x3ff12358f08ae5ba
	.quad	0xbfb960cbbf788d5c
	.quad	0x3ff0953f419900a7
	.quad	0xbfaa6f9db6475fce
	.quad	0x3ff0000000000000
	.quad	0x0000000000000000
	.quad	0x3fee608cfd9a47ac
	.quad	0x3fb338ca9f24f53d
	.quad	0x3feca4b31f026aa0
	.quad	0x3fc476a9543891ba
	.quad	0x3feb2036576afce6
	.quad	0x3fce840b4ac4e4d2
	.quad	0x3fe9c2d163a1aa2d
	.quad	0x3fd40645f0c6651c
	.quad	0x3fe886e6037841ed
	.quad	0x3fd88e9c2c1b9ff8
	.quad	0x3fe767dcf5534862
	.quad	0x3fdce0a44eb17bcc
	.quad	0x3fd27616c9496e0b
	.quad	0xbfd71969a075c67a
	.quad	0x3fdec70a6ca7badd
	.quad	0xbfe7154748bef6c8
	.quad	0x3ff71547652ab82b
	.size	__powf_log2_data, 296

	.type	__exp2f_data,@object
	.section	.rodata.__exp2f_data,"a",@progbits
	.p2align	3, 0x0
__exp2f_data:
	.quad	4607182418800017408
	.quad	4607140297302181236
	.quad	4607100335213349135
	.quad	4607062579818421073
	.quad	4607027079437701499
	.quad	4606993883449571754
	.quad	4606963042313658936
	.quad	4606934607594512097
	.quad	4606908631985796885
	.quad	4606885169335019979
	.quad	4606864274668794914
	.quad	4606846004218661165
	.quad	4606830415447468583
	.quad	4606817567076339586
	.quad	4606807519112221737
	.quad	4606800332876043653
	.quad	4606796071031487437
	.quad	4606794797614391156
	.quad	4606796578062795143
	.quad	4606801479247646227
	.quad	4606809569504174299
	.quad	4606820918663955941
	.quad	4606835598087680144
	.quad	4606853680698631517
	.quad	4606875241016906669
	.quad	4606900355194379847
	.quad	4606929101050434204
	.quad	4606961558108475497
	.quad	4606997807633245319
	.quad	4607037932668951391
	.quad	4607082018078232794
	.quad	4607130150581978432
	.quad	0x42e8000000000000
	.quad	0x3fac6af84b912394
	.quad	0x3fcebfce50fac4f3
	.quad	0x3fe62e42ff0c52d6
	.quad	0x4338000000000000
	.quad	0x40471547652b82fe
	.quad	0x3ebc6af84b912394
	.quad	0x3f2ebfce50fac4f3
	.quad	0x3f962e42ff0c52d6
	.size	__exp2f_data, 328

	.section	.debug_abbrev,"",@progbits
	.byte	1
	.byte	17
	.byte	1
	.byte	37
	.byte	14
	.byte	19
	.byte	5
	.byte	3
	.byte	14
	.byte	16
	.byte	23
	.ascii	"\264B"
	.byte	25
	.byte	17
	.byte	1
	.byte	18
	.byte	6
	.byte	0
	.byte	0
	.byte	2
	.byte	46
	.byte	0
	.byte	17
	.byte	1
	.byte	18
	.byte	6
	.byte	64
	.byte	24
	.byte	110
	.byte	14
	.byte	3
	.byte	14
	.byte	58
	.byte	11
	.byte	59
	.byte	11
	.byte	73
	.byte	19
	.byte	63
	.byte	25
	.byte	0
	.byte	0
	.byte	3
	.byte	36
	.byte	0
	.byte	3
	.byte	14
	.byte	62
	.byte	11
	.byte	11
	.byte	11
	.byte	0
	.byte	0
	.byte	0
	.section	.debug_info,"",@progbits
.Lcu_begin0:
	.long	.Ldebug_info_end0-.Ldebug_info_start0
.Ldebug_info_start0:
	.short	4
	.long	.debug_abbrev
	.byte	8
	.byte	1
	.long	.Linfo_string0
	.short	44
	.long	.Linfo_string1
	.long	.Lline_table_start0

	.quad	.Lfunc_begin0
	.long	.Lfunc_end0-.Lfunc_begin0
	.byte	2
	.quad	.Lfunc_begin0
	.long	.Lfunc_end0-.Lfunc_begin0
	.byte	1
	.byte	86
	.long	.Linfo_string2
	.long	.Linfo_string2
	.byte	1
	.byte	1
	.long	67

	.byte	3
	.long	.Linfo_string3
	.byte	5
	.byte	4
	.byte	0
.Ldebug_info_end0:
	.section	.debug_str,"MS",@progbits,1
.Linfo_string0:
	.asciz	"IREE"
.Linfo_string1:
	.asciz	"-"
.Linfo_string2:
	.asciz	"broadcast_dispatch_0_generic_Dx8640x3200_f16"
.Linfo_string3:
	.asciz	"int"
	.section	.debug_pubnames,"",@progbits
	.long	.LpubNames_end0-.LpubNames_start0
.LpubNames_start0:
	.short	2
	.long	.Lcu_begin0
	.long	75
	.long	38
	.asciz	"broadcast_dispatch_0_generic_Dx8640x3200_f16"
	.long	0
.LpubNames_end0:
	.section	.debug_pubtypes,"",@progbits
	.long	.LpubTypes_end0-.LpubTypes_start0
.LpubTypes_start0:
	.short	2
	.long	.Lcu_begin0
	.long	75
	.long	67
	.asciz	"int"
	.long	0
.LpubTypes_end0:
	.section	".note.GNU-stack","",@progbits
	.section	.debug_line,"",@progbits
.Lline_table_start0: