123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501 |
- .file "tanh_gen.c"
- .text
- ..TXTST0:
- # -- Begin tanh
- .text
- .align 16,0x90
- .globl tanh
- tanh:
- # parameter 1: %xmm0
- ..B1.1:
- .cfi_startproc
- ..___tag_value_tanh.1:
- ..L2:
- movsd %xmm0, -8(%rsp)
- movl -4(%rsp), %eax
- movl %eax, %ecx
- andl $2147483647, %ecx
- cmpl $2146435072, %ecx
- jae ..B1.16
- ..B1.2:
- cmpl $1077088193, %ecx
- jae ..B1.15
- ..B1.3:
- cmpl $1075078759, %ecx
- jae ..B1.14
- ..B1.4:
- cmpl $1066695393, %ecx
- jae ..B1.13
- ..B1.5:
- cmpl $1012924416, %ecx
- jae ..B1.12
- ..B1.6:
- shrl $31, %eax
- cmpl $1048576, %ecx
- jb ..B1.8
- ..B1.7:
- movl %eax, %eax
- lea _small_value_64(%rip), %rdx
- movsd (%rdx,%rax,8), %xmm1
- mulsd (%rdx), %xmm1
- subsd %xmm1, %xmm0
- ret
- ..B1.8:
- movl -8(%rsp), %edx
- orl %edx, %ecx
- je ..B1.10
- ..B1.9:
- lea _small_value_64(%rip), %rdx
- movsd -8(%rsp), %xmm0
- movsd (%rdx,%rax,8), %xmm1
- mulsd (%rdx), %xmm1
- movsd %xmm1, -32(%rsp)
- subsd %xmm1, %xmm0
- ret
- ..B1.10:
- movsd -8(%rsp), %xmm0
- ..B1.11:
- ret
- ..B1.12:
- movsd -8(%rsp), %xmm4
- movaps %xmm4, %xmm2
- mulsd %xmm4, %xmm2
- movaps %xmm2, %xmm1
- mulsd %xmm2, %xmm1
- movsd .L_2il0floatpacket.20(%rip), %xmm0
- movsd .L_2il0floatpacket.22(%rip), %xmm3
- mulsd %xmm1, %xmm0
- mulsd %xmm1, %xmm3
- addsd .L_2il0floatpacket.21(%rip), %xmm0
- addsd .L_2il0floatpacket.23(%rip), %xmm3
- mulsd %xmm1, %xmm0
- mulsd %xmm2, %xmm3
- addsd %xmm3, %xmm0
- mulsd %xmm4, %xmm0
- addsd %xmm4, %xmm0
- ret
- ..B1.13:
- movsd -8(%rsp), %xmm0
- lea _TWO_52(%rip), %rdx
- andps .L_2il0floatpacket.24(%rip), %xmm0
- xorl %r10d, %r10d
- mulsd .L_2il0floatpacket.10(%rip), %xmm0
- movq $0x3ff0000000000000, %rsi
- movsd .L_2il0floatpacket.11(%rip), %xmm4
- lea _TWO_32(%rip), %r9
- mulsd %xmm0, %xmm4
- lea ones(%rip), %r11
- movsd .L_2il0floatpacket.12(%rip), %xmm5
- movsd .L_2il0floatpacket.13(%rip), %xmm6
- movsd .L_2il0floatpacket.25(%rip), %xmm3
- movaps %xmm3, %xmm8
- addsd (%rdx), %xmm4
- movsd %xmm4, -32(%rsp)
- movsd -32(%rsp), %xmm7
- movsd .L_2il0floatpacket.15(%rip), %xmm12
- movsd .L_2il0floatpacket.17(%rip), %xmm14
- movsd .L_2il0floatpacket.14(%rip), %xmm2
- movaps %xmm2, %xmm9
- subsd (%rdx), %xmm7
- mulsd %xmm7, %xmm5
- mulsd %xmm6, %xmm7
- subsd %xmm5, %xmm0
- movaps %xmm0, %xmm10
- movaps %xmm3, %xmm6
- movl -32(%rsp), %ecx
- movl %ecx, %edi
- shll $25, %edi
- subsd %xmm7, %xmm10
- movaps %xmm10, %xmm11
- subsd %xmm10, %xmm0
- addsd %xmm10, %xmm8
- subsd %xmm7, %xmm0
- movsd %xmm8, -40(%rsp)
- addsd %xmm0, %xmm11
- movaps %xmm11, %xmm13
- mulsd %xmm11, %xmm13
- mulsd %xmm13, %xmm12
- mulsd %xmm13, %xmm14
- addsd .L_2il0floatpacket.16(%rip), %xmm12
- addsd .L_2il0floatpacket.18(%rip), %xmm14
- mulsd %xmm13, %xmm12
- mulsd %xmm13, %xmm14
- mulsd %xmm11, %xmm12
- addsd .L_2il0floatpacket.19(%rip), %xmm14
- movl %r10d, -40(%rsp)
- movsd -40(%rsp), %xmm4
- mulsd %xmm13, %xmm14
- addsd %xmm4, %xmm9
- sarl $25, %edi
- subsd %xmm9, %xmm10
- subl %edi, %ecx
- addsd %xmm10, %xmm0
- movslq %edi, %rdi
- addsd %xmm12, %xmm0
- shlq $4, %rdi
- addsd %xmm14, %xmm0
- movq __libm_exp_table_128@GOTPCREL(%rip), %r8
- shrl $7, %ecx
- movsd (%r9), %xmm5
- movsd 1032(%r8,%rdi), %xmm15
- movaps %xmm15, %xmm1
- mulsd %xmm0, %xmm1
- addsd %xmm4, %xmm0
- mulsd %xmm15, %xmm4
- mulsd 1024(%r8,%rdi), %xmm0
- shlq $52, %rcx
- addsd %xmm0, %xmm1
- addq %rsi, %rcx
- movaps %xmm2, %xmm0
- movq %rcx, -16(%rsp)
- movsd -16(%rsp), %xmm7
- mulsd %xmm7, %xmm1
- mulsd %xmm4, %xmm7
- movaps %xmm7, %xmm10
- addsd %xmm7, %xmm0
- addsd %xmm7, %xmm6
- addsd (%r9), %xmm0
- addsd %xmm5, %xmm6
- movsd %xmm0, -32(%rsp)
- movsd -32(%rsp), %xmm0
- movsd %xmm6, -32(%rsp)
- subsd %xmm5, %xmm0
- movsd -32(%rsp), %xmm9
- subsd %xmm0, %xmm10
- subsd (%r9), %xmm9
- addsd %xmm2, %xmm10
- movsd %xmm7, -40(%rsp)
- movaps %xmm9, %xmm8
- shrl $31, %eax
- subsd %xmm9, %xmm7
- addsd %xmm1, %xmm10
- addsd %xmm3, %xmm7
- movsd (%r11,%rax,8), %xmm12
- addsd %xmm7, %xmm1
- addsd %xmm1, %xmm8
- divsd %xmm8, %xmm3
- movsd %xmm3, -24(%rsp)
- movl %r10d, -24(%rsp)
- movsd -24(%rsp), %xmm11
- mulsd %xmm11, %xmm1
- subsd %xmm11, %xmm3
- mulsd %xmm11, %xmm9
- mulsd %xmm3, %xmm8
- addsd %xmm2, %xmm9
- addsd %xmm8, %xmm1
- movaps %xmm11, %xmm2
- addsd %xmm1, %xmm9
- addsd %xmm3, %xmm2
- mulsd %xmm9, %xmm2
- movaps %xmm0, %xmm1
- mulsd %xmm11, %xmm0
- subsd %xmm2, %xmm3
- mulsd %xmm3, %xmm1
- addsd %xmm11, %xmm3
- mulsd %xmm3, %xmm10
- movsd %xmm0, -8(%rsp)
- addsd %xmm10, %xmm1
- mulsd %xmm12, %xmm0
- mulsd %xmm1, %xmm12
- addsd %xmm12, %xmm0
- ret
- ..B1.14:
- lea _TWO_32(%rip), %rdx
- lea _TWO_52(%rip), %rcx
- movsd -8(%rsp), %xmm1
- movq $0x3ff0000000000000, %rsi
- andps .L_2il0floatpacket.24(%rip), %xmm1
- lea twos(%rip), %r8
- movsd .L_2il0floatpacket.0(%rip), %xmm3
- movaps %xmm1, %xmm7
- movsd (%rdx), %xmm0
- lea ones(%rip), %r9
- movsd (%rcx), %xmm4
- addsd %xmm1, %xmm0
- movsd %xmm0, -32(%rsp)
- movsd -32(%rsp), %xmm6
- movsd .L_2il0floatpacket.1(%rip), %xmm2
- mulsd %xmm1, %xmm2
- subsd (%rdx), %xmm6
- movsd .L_2il0floatpacket.2(%rip), %xmm12
- subsd %xmm6, %xmm7
- mulsd %xmm3, %xmm6
- mulsd %xmm3, %xmm7
- addsd %xmm6, %xmm4
- addsd %xmm2, %xmm7
- movsd %xmm4, -32(%rsp)
- movsd -32(%rsp), %xmm5
- movsd .L_2il0floatpacket.6(%rip), %xmm9
- movsd .L_2il0floatpacket.25(%rip), %xmm11
- movslq -32(%rsp), %rdi
- shlq $52, %rdi
- subsd (%rcx), %xmm5
- addq %rsi, %rdi
- subsd %xmm5, %xmm6
- movq %rdi, -16(%rsp)
- addsd %xmm6, %xmm7
- movaps %xmm7, %xmm8
- mulsd %xmm7, %xmm8
- mulsd %xmm8, %xmm12
- mulsd %xmm8, %xmm9
- addsd .L_2il0floatpacket.3(%rip), %xmm12
- addsd .L_2il0floatpacket.7(%rip), %xmm9
- mulsd %xmm8, %xmm12
- mulsd %xmm8, %xmm9
- addsd .L_2il0floatpacket.4(%rip), %xmm12
- addsd .L_2il0floatpacket.8(%rip), %xmm9
- mulsd %xmm8, %xmm12
- mulsd %xmm8, %xmm9
- addsd .L_2il0floatpacket.5(%rip), %xmm12
- addsd .L_2il0floatpacket.9(%rip), %xmm9
- mulsd %xmm7, %xmm12
- movsd -16(%rsp), %xmm10
- subsd %xmm12, %xmm9
- divsd %xmm9, %xmm12
- mulsd .L_2il0floatpacket.10(%rip), %xmm12
- shrl $31, %eax
- addsd %xmm11, %xmm12
- movl %eax, %r10d
- movsd %xmm12, -24(%rsp)
- xorl $1, %eax
- mulsd %xmm10, %xmm12
- movsd (%r8,%rax,8), %xmm0
- addsd %xmm11, %xmm12
- divsd %xmm12, %xmm0
- movsd %xmm7, -8(%rsp)
- addsd (%r9,%r10,8), %xmm0
- ret
- ..B1.15:
- lea ones(%rip), %rdx
- lea _small_value_64(%rip), %rcx
- shrl $31, %eax
- movsd (%rdx,%rax,8), %xmm0
- subsd (%rcx,%rax,8), %xmm0
- ret
- ..B1.16:
- andl $-2146435073, %ecx
- orl -8(%rsp), %ecx
- jne ..B1.18
- ..B1.17:
- lea ones(%rip), %rdx
- shrl $31, %eax
- movsd (%rdx,%rax,8), %xmm0
- ret
- ..B1.18:
- movsd -8(%rsp), %xmm0
- ..B1.19:
- ret
- .align 16,0x90
- .cfi_endproc
- .type tanh,@function
- .size tanh,.-tanh
- .data
- # -- End tanh
- .section .rodata, "a"
- .align 16
- .align 16
- .L_2il0floatpacket.24:
- .long 0xffffffff,0x7fffffff,0x00000000,0x00000000
- .type .L_2il0floatpacket.24,@object
- .size .L_2il0floatpacket.24,16
- .align 8
- .L_2il0floatpacket.0:
- .long 0x60000000,0x40071547
- .type .L_2il0floatpacket.0,@object
- .size .L_2il0floatpacket.0,8
- .align 8
- .L_2il0floatpacket.1:
- .long 0xf85ddf44,0x3e64ae0b
- .type .L_2il0floatpacket.1,@object
- .size .L_2il0floatpacket.1,8
- .align 8
- .L_2il0floatpacket.2:
- .long 0xfc0798c2,0x3fb3adff
- .type .L_2il0floatpacket.2,@object
- .size .L_2il0floatpacket.2,8
- .align 8
- .L_2il0floatpacket.3:
- .long 0x0e243699,0x406e3d92
- .type .L_2il0floatpacket.3,@object
- .size .L_2il0floatpacket.3,8
- .align 8
- .L_2il0floatpacket.4:
- .long 0xefbe8c98,0x40f689a6
- .type .L_2il0floatpacket.4,@object
- .size .L_2il0floatpacket.4,8
- .align 8
- .L_2il0floatpacket.5:
- .long 0x1ba09c9e,0x4156de47
- .type .L_2il0floatpacket.5,@object
- .size .L_2il0floatpacket.5,8
- .align 8
- .L_2il0floatpacket.6:
- .long 0xa57ad784,0x4018d7c2
- .type .L_2il0floatpacket.6,@object
- .size .L_2il0floatpacket.6,8
- .align 8
- .L_2il0floatpacket.7:
- .long 0x63529aa4,0x40b6b90b
- .type .L_2il0floatpacket.7,@object
- .size .L_2il0floatpacket.7,8
- .align 8
- .L_2il0floatpacket.8:
- .long 0xafc4d247,0x412d4376
- .type .L_2il0floatpacket.8,@object
- .size .L_2il0floatpacket.8,8
- .align 8
- .L_2il0floatpacket.9:
- .long 0x00000000,0x41707ef8
- .type .L_2il0floatpacket.9,@object
- .size .L_2il0floatpacket.9,8
- .align 8
- .L_2il0floatpacket.10:
- .long 0x00000000,0x40000000
- .type .L_2il0floatpacket.10,@object
- .size .L_2il0floatpacket.10,8
- .align 8
- .L_2il0floatpacket.11:
- .long 0x652b82fe,0x40671547
- .type .L_2il0floatpacket.11,@object
- .size .L_2il0floatpacket.11,8
- .align 8
- .L_2il0floatpacket.12:
- .long 0x00000000,0x3f762e42
- .type .L_2il0floatpacket.12,@object
- .size .L_2il0floatpacket.12,8
- .align 8
- .L_2il0floatpacket.13:
- .long 0x3de6af28,0x3e2fdf47
- .type .L_2il0floatpacket.13,@object
- .size .L_2il0floatpacket.13,8
- .align 8
- .L_2il0floatpacket.14:
- .long 0x00000000,0xbff00000
- .type .L_2il0floatpacket.14,@object
- .size .L_2il0floatpacket.14,8
- .align 8
- .L_2il0floatpacket.15:
- .long 0x6887cd7c,0x3f811111
- .type .L_2il0floatpacket.15,@object
- .size .L_2il0floatpacket.15,8
- .align 8
- .L_2il0floatpacket.16:
- .long 0x55555405,0x3fc55555
- .type .L_2il0floatpacket.16,@object
- .size .L_2il0floatpacket.16,8
- .align 8
- .L_2il0floatpacket.17:
- .long 0x87372663,0x3f56c16c
- .type .L_2il0floatpacket.17,@object
- .size .L_2il0floatpacket.17,8
- .align 8
- .L_2il0floatpacket.18:
- .long 0x5555541d,0x3fa55555
- .type .L_2il0floatpacket.18,@object
- .size .L_2il0floatpacket.18,8
- .align 8
- .L_2il0floatpacket.19:
- .long 0x00000000,0x3fe00000
- .type .L_2il0floatpacket.19,@object
- .size .L_2il0floatpacket.19,8
- .align 8
- .L_2il0floatpacket.20:
- .long 0xbe3b77e0,0x3f9662a1
- .type .L_2il0floatpacket.20,@object
- .size .L_2il0floatpacket.20,8
- .align 8
- .L_2il0floatpacket.21:
- .long 0x1110d2a4,0x3fc11111
- .type .L_2il0floatpacket.21,@object
- .size .L_2il0floatpacket.21,8
- .align 8
- .L_2il0floatpacket.22:
- .long 0x00fa2d61,0xbfaba1ba
- .type .L_2il0floatpacket.22,@object
- .size .L_2il0floatpacket.22,8
- .align 8
- .L_2il0floatpacket.23:
- .long 0x55555555,0xbfd55555
- .type .L_2il0floatpacket.23,@object
- .size .L_2il0floatpacket.23,8
- .align 8
- .L_2il0floatpacket.25:
- .long 0x00000000,0x3ff00000
- .type .L_2il0floatpacket.25,@object
- .size .L_2il0floatpacket.25,8
- .align 8
- ones:
- .long 0x00000000,0x3ff00000
- .long 0x00000000,0xbff00000
- .type ones,@object
- .size ones,16
- .align 8
- twos:
- .long 0x00000000,0x40000000
- .long 0x00000000,0xc0000000
- .type twos,@object
- .size twos,16
- .align 4
- _small_value_64:
- .long 0
- .long 24117248
- .long 0
- .long 2171600896
- .type _small_value_64,@object
- .size _small_value_64,16
- .align 4
- _TWO_52:
- .long 0
- .long 1127219200
- .type _TWO_52,@object
- .size _TWO_52,8
- .align 4
- _TWO_32:
- .long 0
- .long 1106247680
- .type _TWO_32,@object
- .size _TWO_32,8
- .data
- .section .note.GNU-stack, ""
- .section .eh_frame,"a",@progbits
- .eh_frame_seg:
- .align 1
- # End
|