moebiusband/NuSiF-Solver
2
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
b4bef94066
NuSiF-Solver/PoissonSolver/2D-mpi/ICC/solver.s

4207 lines
262 KiB
ArmAsm
Raw Blame History

# mark_description "Intel(R) C Intel(R) 64 Compiler Classic for applications running on Intel(R) 64, Version 2021.10.0 Build 202";
# mark_description "30609_000000";
# mark_description "-I./src/includes -I./ICC -I/apps/SPACK/0.19.1/opt/linux-almalinux8-icelake/gcc-8.5.0/intel-oneapi-mpi-2021.1";
# mark_description "0.0-ki6gcj4pr5rekayazvm4lsmvcyosce25/mpi/2021.10.0/include -Wl,-rpath,/apps/SPACK/0.19.1/opt/linux-almalinux";
# mark_description "8-icelake/gcc-8.5.0/intel-oneapi-compilers-2023.2.1-axze7ocbc5zpklxb6kr5h2d5hbpf2lxk/compiler/2023.2.1/linux";
# mark_description "/lib -Wl,-rpath,/apps/SPACK/0.19.1/opt/linux-almalinux8-icelake/gcc-8.5.0/intel-oneapi-compilers-2023.2.1-ax";
# mark_description "ze7ocbc5zpklxb6kr5h2d5hbpf2lxk/compiler/2023.2.1/linux/lib/x64 -Wl,-rpath,/apps/SPACK/0.19.1/opt/linux-almal";
# mark_description "inux8-icelake/gcc-8.5.0/intel-oneapi-compilers-2023.2.1-axze7ocbc5zpklxb6kr5h2d5hbpf2lxk/compiler/2023.2.1/l";
# mark_description "inux/lib/oclfpga/host/linux64/lib -Wl,-rpath,/apps/SPACK/0.19.1/opt/linux-almalinux8-icelake/gcc-8.5.0/intel";
# mark_description "-oneapi-compilers-2023.2.1-axze7ocbc5zpklxb6kr5h2d5hbpf2lxk/compiler/2023.2.1/linux/lib/oclfpga/linux64/lib ";
# mark_description "-Wl,-rpath,/apps/SPACK/0.19.1/opt/linux-almalinux8-icelake/gcc-8.5.0/intel-oneapi-compilers-2023.2.1-axze7oc";
# mark_description "bc5zpklxb6kr5h2d5hbpf2lxk/compiler/2023.2.1/linux/compiler/lib/intel64_lin -gcc-name=/usr/bin/gcc -gxx-name=";
# mark_description "/usr/bin/g++ -S -fverbose-asm -D_GNU_SOURCE -DARRAY_ALIGNMENT=64 -O3 -xHost -qopt-zmm-usage=high -std=c99 -o";
# mark_description " ICC/solver.s";
.file "solver.c"
.text
..TXTST0:
.L_2__routine_start_getResult_0:
# -- Begin getResult
.text
# mark_begin;
.align 16,0x90
.globl getResult
# --- getResult(Solver *)
getResult:
# parameter 1: %rdi
..B1.1: # Preds ..B1.0
# Execution count [1.00e+00]
.cfi_startproc
..___tag_value_getResult.1:
..L2:
#73.1
pushq %rbp #73.1
.cfi_def_cfa_offset 16
movq %rsp, %rbp #73.1
.cfi_def_cfa 6, 16
.cfi_offset 6, -16
andq $-64, %rsp #73.1
pushq %r12 #73.1
pushq %r13 #73.1
pushq %r14 #73.1
pushq %r15 #73.1
pushq %rbx #73.1
subq $88, %rsp #73.1
.cfi_escape 0x10, 0x03, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xd8, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xf8, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xf0, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xe8, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0f, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x22
movq %rdi, %r13 #73.1
xorl %r12d, %r12d #74.18
cmpl $0, 36(%r13) #77.25
je ..B1.21 # Prob 5% #77.25
# LOE rbx r9 r12 r13
..B1.2: # Preds ..B1.1
# Execution count [9.50e-01]
movl 24(%r13), %esi #93.39
movl 32(%r13), %r15d #92.26
# LOE rbx r9 r12 r13 esi r15d
..B1.3: # Preds ..B1.24 ..B1.46 ..B1.45 ..B1.2
# Execution count [1.00e+00]
movslq %esi, %rsi #94.5
addq $-32, %rsp #94.5
movq 48(%r13), %rax #94.5
movl $1275070475, %edx #94.5
movq %r12, %rcx #94.5
movq %rbx, %r8 #94.5
movl $1275070475, (%rsp) #94.5
lea 16(%rax,%rsi,8), %rdi #94.5
addl $2, %esi #94.5
imull %r15d, %esi #94.5
movl $0, 8(%rsp) #94.5
movl $1140850688, 16(%rsp) #94.5
vzeroupper #94.5
..___tag_value_getResult.11:
# MPI_Gatherv(const void *, int, MPI_Datatype, void *, const int *, const int *, MPI_Datatype, int, MPI_Comm)
call MPI_Gatherv #94.5
..___tag_value_getResult.12:
# LOE r12 r13
..B1.51: # Preds ..B1.3
# Execution count [1.00e+00]
addq $32, %rsp #94.5
# LOE r12 r13
..B1.4: # Preds ..B1.51
# Execution count [1.00e+00]
cmpl $0, 36(%r13) #104.25
jne ..B1.18 # Prob 78% #104.25
# LOE r12 r13
..B1.5: # Preds ..B1.4
# Execution count [2.20e-01]
movl $.L_2__STRING.4, %edi #105.9
movl $.L_2__STRING.9, %esi #105.9
movl 24(%r13), %ebx #105.9
movl 28(%r13), %r15d #105.9
# fopen(const char *__restrict__, const char *__restrict__)
call fopen #105.9
# LOE rax r12 ebx r15d
..B1.6: # Preds ..B1.5
# Execution count [2.20e-01]
testq %rax, %rax #105.9
je ..B1.19 # Prob 0% #105.9
# LOE rax r12 ebx r15d
..B1.7: # Preds ..B1.6
# Execution count [2.18e-01]
addl $2, %r15d #105.9
xorl %edx, %edx #105.9
testl %r15d, %r15d #105.9
jle ..B1.17 # Prob 10% #105.9
# LOE rax r12 edx ebx r15d
..B1.8: # Preds ..B1.7
# Execution count [1.97e-01]
movslq %ebx, %rsi #105.9
lea 2(%rbx), %r13d #105.9
addq $2, %rsi #105.9
xorl %r14d, %r14d #105.9
movq %rsi, 24(%rsp) #105.9[spill]
movl %r15d, 40(%rsp) #105.9[spill]
movl %edx, %r15d #105.9
movq %r12, 32(%rsp) #105.9[spill]
movq %rax, %r12 #105.9
# LOE r12 ebx r13d r14d r15d
..B1.9: # Preds ..B1.15 ..B1.8
# Execution count [1.09e+00]
xorl %edx, %edx #105.9
testl %r13d, %r13d #105.9
jle ..B1.14 # Prob 10% #105.9
# LOE rdx r12 ebx r13d r14d r15d
..B1.10: # Preds ..B1.9
# Execution count [9.83e-01]
movslq %r14d, %r14 #105.9
movq 32(%rsp), %rcx #105.9[spill]
movl %r15d, 8(%rsp) #105.9[spill]
movl %r13d, 16(%rsp) #105.9[spill]
movq %rdx, %r13 #105.9
movl %ebx, (%rsp) #105.9[spill]
lea (%rcx,%r14,8), %rax #105.9
movq 24(%rsp), %r15 #105.9[spill]
movq %rax, %rbx #105.9
# LOE rbx r12 r13 r15 r14d
..B1.11: # Preds ..B1.12 ..B1.10
# Execution count [5.46e+00]
movq %r12, %rdi #105.9
movl $.L_2__STRING.11, %esi #105.9
movl $1, %eax #105.9
vmovsd (%rbx,%r13,8), %xmm0 #105.9
# fprintf(FILE *__restrict__, const char *__restrict__, ...)
call fprintf #105.9
# LOE rbx r12 r13 r15 r14d
..B1.12: # Preds ..B1.11
# Execution count [5.46e+00]
incq %r13 #105.9
cmpq %r15, %r13 #105.9
jl ..B1.11 # Prob 82% #105.9
# LOE rbx r12 r13 r15 r14d
..B1.13: # Preds ..B1.12
# Execution count [9.83e-01]
movl 16(%rsp), %r13d #[spill]
movl 8(%rsp), %r15d #[spill]
movl (%rsp), %ebx #[spill]
# LOE r12 ebx r13d r14d r15d
..B1.14: # Preds ..B1.9 ..B1.13
# Execution count [1.09e+00]
movl $10, %edi #105.9
movq %r12, %rsi #105.9
call fputc #105.9
# LOE r12 ebx r13d r14d r15d
..B1.15: # Preds ..B1.14
# Execution count [1.09e+00]
incl %r15d #105.9
lea 2(%r14,%rbx), %r14d #105.9
cmpl 40(%rsp), %r15d #105.9[spill]
jl ..B1.9 # Prob 82% #105.9
# LOE r12 ebx r13d r14d r15d
..B1.16: # Preds ..B1.15
# Execution count [1.97e-01]
movq %r12, %rax #
# LOE rax
..B1.17: # Preds ..B1.7 ..B1.16
# Execution count [2.18e-01]
movq %rax, %rdi #105.9
addq $88, %rsp #105.9
.cfi_restore 3
popq %rbx #105.9
.cfi_restore 15
popq %r15 #105.9
.cfi_restore 14
popq %r14 #105.9
.cfi_restore 13
popq %r13 #105.9
.cfi_restore 12
popq %r12 #105.9
movq %rbp, %rsp #105.9
popq %rbp #105.9
.cfi_def_cfa 7, 8
.cfi_restore 6
# fclose(FILE *)
jmp fclose #105.9
.cfi_def_cfa 6, 16
.cfi_escape 0x10, 0x03, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xd8, 0xff, 0xff, 0xff, 0x22
.cfi_offset 6, -16
.cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xf8, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xf0, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xe8, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0f, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x22
# LOE
..B1.18: # Preds ..B1.4
# Execution count [1.00e+00]
addq $88, %rsp #107.1
.cfi_restore 3
popq %rbx #107.1
.cfi_restore 15
popq %r15 #107.1
.cfi_restore 14
popq %r14 #107.1
.cfi_restore 13
popq %r13 #107.1
.cfi_restore 12
popq %r12 #107.1
movq %rbp, %rsp #107.1
popq %rbp #107.1
.cfi_def_cfa 7, 8
.cfi_restore 6
ret #107.1
.cfi_def_cfa 6, 16
.cfi_escape 0x10, 0x03, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xd8, 0xff, 0xff, 0xff, 0x22
.cfi_offset 6, -16
.cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xf8, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xf0, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xe8, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0f, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x22
# LOE
..B1.19: # Preds ..B1.6
# Execution count [1.57e-03]: Infreq
movl $il0_peep_printf_format_0, %edi #105.9
call puts #105.9
# LOE
..B1.20: # Preds ..B1.19
# Execution count [1.57e-03]: Infreq
movl $1, %edi #105.9
# exit(int)
call exit #105.9
# LOE
..B1.21: # Preds ..B1.1
# Execution count [5.00e-02]: Infreq
movl 28(%r13), %eax #78.16
movl $64, %edi #78.16
movl 24(%r13), %esi #78.16
addl $2, %eax #78.16
addl $2, %esi #78.16
imull %eax, %esi #78.16
movslq %esi, %rsi #78.16
shlq $3, %rsi #78.16
..___tag_value_getResult.53:
# allocate(int, size_t)
call allocate #78.16
..___tag_value_getResult.54:
# LOE rax r13
..B1.53: # Preds ..B1.21
# Execution count [5.00e-02]: Infreq
movq %rax, %r12 #78.16
# LOE r12 r13
..B1.22: # Preds ..B1.53
# Execution count [5.00e-02]: Infreq
movslq 40(%r13), %rdi #79.30
shlq $2, %rdi #79.30
# malloc(size_t)
call malloc #79.30
# LOE rax r12 r13
..B1.54: # Preds ..B1.22
# Execution count [5.00e-02]: Infreq
movq %rax, %rbx #79.30
# LOE rbx r12 r13
..B1.23: # Preds ..B1.54
# Execution count [5.00e-02]: Infreq
movslq 40(%r13), %rdi #80.30
shlq $2, %rdi #80.30
# malloc(size_t)
call malloc #80.30
# LOE rax rbx r12 r13
..B1.55: # Preds ..B1.23
# Execution count [5.00e-02]: Infreq
movq %rax, %r9 #80.30
# LOE rbx r9 r12 r13
..B1.24: # Preds ..B1.55
# Execution count [4.50e-02]: Infreq
movl 24(%r13), %esi #81.45
movl 32(%r13), %r15d #81.24
movl 40(%r13), %r8d #85.29
movl $0, (%r9) #82.9
lea 2(%rsi), %r14d #81.60
movl %r14d, %eax #81.60
imull %r15d, %eax #81.60
movl %eax, (%rsp) #81.60[spill]
movl %eax, (%rbx) #81.9
cmpl $1, %r8d #85.29
jle ..B1.3 # Prob 0% #85.29
# LOE rbx r9 r12 r13 esi r8d r14d r15d
..B1.25: # Preds ..B1.24
# Execution count [4.50e-02]: Infreq
movl 28(%r13), %r10d #86.56
lea -1(%r8), %ecx #85.29
cmpl $16, %ecx #85.9
jl ..B1.48 # Prob 10% #85.9
# LOE rbx r9 r12 r13 ecx esi r8d r10d r14d r15d
..B1.26: # Preds ..B1.25
# Execution count [4.50e-02]: Infreq
cmpl $123, %ecx #85.9
jl ..B1.47 # Prob 10% #85.9
# LOE rbx r9 r12 r13 ecx esi r8d r10d r14d r15d
..B1.27: # Preds ..B1.26
# Execution count [4.50e-02]: Infreq
lea 4(%rbx), %r11 #86.13
andq $63, %r11 #85.9
testl $3, %r11d #85.9
je ..B1.29 # Prob 50% #85.9
# LOE rbx r9 r12 r13 ecx esi r8d r10d r11d r14d r15d
..B1.28: # Preds ..B1.27
# Execution count [2.25e-02]: Infreq
xorl %r11d, %r11d #85.9
jmp ..B1.31 # Prob 100% #85.9
# LOE rbx r9 r12 r13 ecx esi r8d r10d r11d r14d r15d
..B1.29: # Preds ..B1.27
# Execution count [2.25e-02]: Infreq
testl %r11d, %r11d #85.9
je ..B1.31 # Prob 50% #85.9
# LOE rbx r9 r12 r13 ecx esi r8d r10d r11d r14d r15d
..B1.30: # Preds ..B1.29
# Execution count [2.50e-01]: Infreq
negl %r11d #85.9
addl $64, %r11d #85.9
shrl $2, %r11d #85.9
cmpl %r11d, %ecx #85.9
cmovl %ecx, %r11d #85.9
# LOE rbx r9 r12 r13 ecx esi r8d r10d r11d r14d r15d
..B1.31: # Preds ..B1.28 ..B1.30 ..B1.29
# Execution count [5.00e-02]: Infreq
movl %ecx, %edi #85.9
subl %r11d, %edi #85.9
andl $15, %edi #85.9
negl %edi #85.9
addl %ecx, %edi #85.9
vmovups .L_2il0floatpacket.0(%rip), %zmm7 #85.9
cmpl $1, %r11d #85.9
jb ..B1.35 # Prob 50% #85.9
# LOE rbx r9 r12 r13 ecx esi edi r8d r10d r11d r14d r15d zmm7
..B1.32: # Preds ..B1.31
# Execution count [4.50e-02]: Infreq
movl %r10d, %eax #86.28
cltd #86.28
vmovups .L_2il0floatpacket.2(%rip), %zmm6 #85.9
vmovups .L_2il0floatpacket.3(%rip), %zmm5 #86.13
vmovups .L_2il0floatpacket.1(%rip), %zmm1 #86.28
idivl %r8d #86.28
movq $0, 8(%rsp) #85.9[spill]
imull %r14d, %eax #86.88
vpbroadcastd %r14d, %zmm4 #86.88
vpbroadcastd %r11d, %zmm0 #85.9
vpbroadcastd %edx, %zmm2 #86.28
vpbroadcastd %eax, %zmm3 #86.88
movslq %r11d, %rdx #85.9
movq 8(%rsp), %rax #85.9[spill]
# LOE rax rdx rbx r9 r12 r13 ecx esi edi r8d r10d r11d r14d r15d zmm0 zmm1 zmm2 zmm3 zmm4 zmm5 zmm6 zmm7
..B1.33: # Preds ..B1.33 ..B1.32
# Execution count [2.50e-01]: Infreq
vpcmpgtd %zmm5, %zmm2, %k1 #86.28
vpcmpgtd %zmm6, %zmm0, %k2 #85.9
vpaddd %zmm7, %zmm6, %zmm6 #85.9
vpaddd %zmm7, %zmm5, %zmm5 #86.13
vmovdqa32 %zmm1, %zmm8{%k1}{z} #86.28
vpmulld %zmm8, %zmm4, %zmm9 #86.88
vpaddd %zmm9, %zmm3, %zmm10 #86.88
vmovdqu32 %zmm10, 4(%rbx,%rax,4){%k2} #86.13
addq $16, %rax #85.9
cmpq %rdx, %rax #85.9
jb ..B1.33 # Prob 82% #85.9
# LOE rax rdx rbx r9 r12 r13 ecx esi edi r8d r10d r11d r14d r15d zmm0 zmm1 zmm2 zmm3 zmm4 zmm5 zmm6 zmm7
..B1.34: # Preds ..B1.33
# Execution count [4.50e-02]: Infreq
cmpl %r11d, %ecx #85.9
je ..B1.41 # Prob 10% #85.9
# LOE rbx r9 r12 r13 ecx esi edi r8d r10d r11d r14d r15d zmm7
..B1.35: # Preds ..B1.31 ..B1.34 ..B1.47
# Execution count [2.50e-01]: Infreq
lea 1(%r11), %edx #86.13
vpbroadcastd %edx, %zmm0 #86.13
lea 16(%r11), %edx #85.9
vpaddd .L_2il0floatpacket.4(%rip), %zmm0, %zmm4 #86.13
cmpl %edx, %edi #85.9
jl ..B1.39 # Prob 50% #85.9
# LOE rbx r9 r12 r13 ecx esi edi r8d r10d r11d r14d r15d zmm4 zmm7
..B1.36: # Preds ..B1.35
# Execution count [4.50e-02]: Infreq
movslq %r11d, %r11 #85.9
movl %r10d, %eax #86.28
vpbroadcastd %r14d, %zmm3 #86.88
vmovups .L_2il0floatpacket.1(%rip), %zmm0 #86.28
lea (%rbx,%r11,4), %rdx #79.30
movq %rdx, 8(%rsp) #79.30[spill]
cltd #86.28
idivl %r8d #86.28
vpbroadcastd %edx, %zmm1 #86.28
imull %r14d, %eax #86.88
vpbroadcastd %eax, %zmm2 #86.88
movq 8(%rsp), %rdx #86.88[spill]
# LOE rdx rbx r9 r12 r13 ecx esi edi r8d r10d r11d r14d r15d zmm0 zmm1 zmm2 zmm3 zmm4 zmm7
..B1.37: # Preds ..B1.37 ..B1.36
# Execution count [2.50e-01]: Infreq
vpcmpgtd %zmm4, %zmm1, %k1 #86.28
addl $16, %r11d #85.9
vpaddd %zmm7, %zmm4, %zmm4 #86.13
vmovdqa32 %zmm0, %zmm5{%k1}{z} #86.28
vpmulld %zmm5, %zmm3, %zmm6 #86.88
vpaddd %zmm6, %zmm2, %zmm8 #86.88
vmovdqu32 %zmm8, 4(%rdx) #86.13
addq $64, %rdx #85.9
cmpl %edi, %r11d #85.9
jb ..B1.37 # Prob 82% #85.9
# LOE rdx rbx r9 r12 r13 ecx esi edi r8d r10d r11d r14d r15d zmm0 zmm1 zmm2 zmm3 zmm4 zmm7
..B1.39: # Preds ..B1.37 ..B1.35 ..B1.48
# Execution count [5.00e-02]: Infreq
lea 1(%rdi), %edx #85.9
cmpl %ecx, %edx #85.9
ja ..B1.41 # Prob 50% #85.9
# LOE rbx r9 r12 r13 edx ecx esi edi r8d r10d r14d r15d
..B1.40: # Preds ..B1.39
# Execution count [2.50e-01]: Infreq
movl %r10d, %eax #86.28
movl %ecx, %r10d #85.9
vpbroadcastd %edx, %zmm0 #86.13
subl %edi, %r10d #85.9
cltd #86.28
vpbroadcastd %r14d, %zmm3 #86.88
vpaddd .L_2il0floatpacket.4(%rip), %zmm0, %zmm2 #86.13
vpbroadcastd %r10d, %zmm7 #85.9
idivl %r8d #86.28
vpbroadcastd %edx, %zmm1 #86.28
imull %eax, %r14d #86.88
vpcmpgtd .L_2il0floatpacket.2(%rip), %zmm7, %k2 #85.9
vpcmpgtd %zmm2, %zmm1, %k1 #86.28
vpbroadcastd %r14d, %zmm5 #86.88
vmovdqu32 .L_2il0floatpacket.1(%rip), %zmm4{%k1}{z} #86.28
movslq %edi, %rdi #86.13
vpmulld %zmm4, %zmm3, %zmm6 #86.88
vpaddd %zmm6, %zmm5, %zmm8 #86.88
vmovdqu32 %zmm8, 4(%rbx,%rdi,4){%k2} #86.13
# LOE rbx r9 r12 r13 ecx esi r15d
..B1.41: # Preds ..B1.34 ..B1.39 ..B1.40
# Execution count [5.00e-02]: Infreq
movl %ecx, %eax #85.9
movl $1, %edi #85.9
xorl %edx, %edx #85.9
shrl $1, %eax #85.9
je ..B1.45 # Prob 10% #85.9
# LOE rax rdx rbx r9 r12 r13 ecx esi edi r15d
..B1.42: # Preds ..B1.41
# Execution count [4.50e-02]: Infreq
movl (%rsp), %r11d #85.9[spill]
# LOE rax rdx rbx r9 r12 r13 ecx esi r11d r15d
..B1.43: # Preds ..B1.43 ..B1.42
# Execution count [1.25e-01]: Infreq
movl 4(%rbx,%rdx,8), %edi #88.23
movl %r11d, 4(%r9,%rdx,8) #87.13
addl %edi, %r11d #88.13
movl %r11d, 8(%r9,%rdx,8) #87.13
movl 8(%rbx,%rdx,8), %r10d #88.23
incq %rdx #85.9
addl %r10d, %r11d #88.13
cmpq %rax, %rdx #85.9
jb ..B1.43 # Prob 64% #85.9
# LOE rax rdx rbx r9 r12 r13 ecx esi r11d r15d
..B1.44: # Preds ..B1.43
# Execution count [4.50e-02]: Infreq
movl %r11d, (%rsp) #[spill]
lea 1(%rdx,%rdx), %edi #86.13
# LOE rbx r9 r12 r13 ecx esi edi r15d
..B1.45: # Preds ..B1.44 ..B1.41
# Execution count [5.00e-02]: Infreq
lea -1(%rdi), %eax #85.9
cmpl %ecx, %eax #85.9
jae ..B1.3 # Prob 10% #85.9
# LOE rbx r9 r12 r13 esi edi r15d
..B1.46: # Preds ..B1.45
# Execution count [4.50e-02]: Infreq
movslq %edi, %rdi #86.13
movl (%rsp), %eax #87.13[spill]
movl %eax, (%r9,%rdi,4) #87.13
jmp ..B1.3 # Prob 100% #87.13
# LOE rbx r9 r12 r13 esi r15d
..B1.47: # Preds ..B1.26
# Execution count [4.50e-03]: Infreq
movl %ecx, %edi #85.9
xorl %r11d, %r11d #85.9
vmovups .L_2il0floatpacket.0(%rip), %zmm7 #86.13
andl $-16, %edi #85.9
jmp ..B1.35 # Prob 100% #85.9
# LOE rbx r9 r12 r13 ecx esi edi r8d r10d r11d r14d r15d zmm7
..B1.48: # Preds ..B1.25
# Execution count [4.50e-03]: Infreq
xorl %edi, %edi #85.9
jmp ..B1.39 # Prob 100% #85.9
.align 16,0x90
# LOE rbx r9 r12 r13 ecx esi edi r8d r10d r14d r15d
.cfi_endproc
# mark_end;
.type getResult,@function
.size getResult,.-getResult
..LNgetResult.0:
.section .rodata.str1.4, "aMS",@progbits,1
.align 4
.align 4
il0_peep_printf_format_0:
.long 1869771333
.word 8562
.byte 0
.data
# -- End getResult
.text
.L_2__routine_start_debug_1:
# -- Begin debug
.text
# mark_begin;
.align 16,0x90
.globl debug
# --- debug(Solver *)
debug:
# parameter 1: %rdi
..B2.1: # Preds ..B2.0
# Execution count [1.00e+00]
.cfi_startproc
..___tag_value_debug.64:
..L65:
#162.1
subq $72, %rsp #162.1
.cfi_def_cfa_offset 80
xorl %eax, %eax #187.16
movl 36(%rdi), %edx #164.17
cmpl $0, 40(%rdi) #187.25
jle ..B2.20 # Prob 10% #187.25
# LOE rbx rbp rdi r12 r13 r14 r15 eax edx
..B2.2: # Preds ..B2.1
# Execution count [9.00e-01]
movq %r12, (%rsp) #[spill]
.cfi_offset 12, -80
movl %eax, %r12d #
movq %r13, 8(%rsp) #[spill]
.cfi_offset 13, -72
movl %edx, %r13d #
movq %r14, 16(%rsp) #[spill]
movq %r15, 24(%rsp) #[spill]
movq %rbx, 32(%rsp) #[spill]
.cfi_offset 3, -48
.cfi_offset 14, -64
.cfi_offset 15, -56
movq %rdi, %rbx #
movq %rbp, 40(%rsp) #[spill]
.cfi_offset 6, -40
# LOE rbx r12d r13d
..B2.3: # Preds ..B2.18 ..B2.2
# Execution count [5.00e+00]
cmpl %r13d, %r12d #188.18
jne ..B2.17 # Prob 78% #188.18
# LOE rbx r12d r13d
..B2.4: # Preds ..B2.3
# Execution count [1.10e+00]
movl $.L_2__STRING.0, %edi #189.13
xorl %eax, %eax #189.13
movl 36(%rbx), %esi #189.13
movq 48(%rbx), %r15 #189.13
movl 24(%rbx), %r14d #189.13
..___tag_value_debug.79:
# printf(const char *__restrict__, ...)
call printf #189.13
..___tag_value_debug.80:
# LOE rbx r15 r12d r13d r14d
..B2.5: # Preds ..B2.4
# Execution count [1.10e+00]
movl 32(%rbx), %ecx #189.13
xorl %esi, %esi #189.13
addl $2, %ecx #189.13
testl %ecx, %ecx #189.13
jle ..B2.16 # Prob 10% #189.13
# LOE rbx r15 esi r12d r13d r14d
..B2.6: # Preds ..B2.5
# Execution count [9.90e-01]
movl %r12d, 56(%rsp) #189.13[spill]
xorl %ebp, %ebp #189.13
movl %r13d, 48(%rsp) #189.13[spill]
movl %r14d, %r13d #189.13
movl %esi, %r14d #189.13
# LOE rbx r15 ebp r13d r14d
..B2.7: # Preds ..B2.14 ..B2.6
# Execution count [5.50e+00]
movl $.L_2__STRING.1, %edi #189.13
movl %r14d, %esi #189.13
xorl %eax, %eax #189.13
..___tag_value_debug.83:
# printf(const char *__restrict__, ...)
call printf #189.13
..___tag_value_debug.84:
# LOE rbx r15 ebp r13d r14d
..B2.8: # Preds ..B2.7
# Execution count [5.50e+00]
movl 24(%rbx), %edx #189.13
xorl %eax, %eax #189.13
addl $2, %edx #189.13
testl %edx, %edx #189.13
jle ..B2.13 # Prob 10% #189.13
# LOE rax rbx r15 ebp r13d r14d
..B2.9: # Preds ..B2.8
# Execution count [4.95e+00]
movslq %ebp, %rbp #189.13
movq %r15, 64(%rsp) #189.13[spill]
lea (%r15,%rbp,8), %r12 #189.13
movq %rax, %r15 #189.13
# LOE rbx r12 r15 ebp r13d r14d
..B2.10: # Preds ..B2.11 ..B2.9
# Execution count [2.75e+01]
vmovsd (%r12,%r15,8), %xmm0 #189.13
movl $.L_2__STRING.2, %edi #189.13
movl $1, %eax #189.13
..___tag_value_debug.86:
# printf(const char *__restrict__, ...)
call printf #189.13
..___tag_value_debug.87:
# LOE rbx r12 r15 ebp r13d r14d
..B2.11: # Preds ..B2.10
# Execution count [2.75e+01]
movslq 24(%rbx), %rdx #189.13
incq %r15 #189.13
addq $2, %rdx #189.13
cmpq %rdx, %r15 #189.13
jl ..B2.10 # Prob 82% #189.13
# LOE rbx r12 r15 ebp r13d r14d
..B2.12: # Preds ..B2.11
# Execution count [4.95e+00]
movq 64(%rsp), %r15 #[spill]
# LOE rbx r15 ebp r13d r14d
..B2.13: # Preds ..B2.12 ..B2.8
# Execution count [5.50e+00]
movl $10, %edi #189.13
call putchar #189.13
# LOE rbx r15 ebp r13d r14d
..B2.14: # Preds ..B2.13
# Execution count [5.50e+00]
movl 32(%rbx), %eax #189.13
incl %r14d #189.13
addl $2, %eax #189.13
lea 2(%rbp,%r13), %ebp #189.13
cmpl %eax, %r14d #189.13
jl ..B2.7 # Prob 82% #189.13
# LOE rbx r15 ebp r13d r14d
..B2.15: # Preds ..B2.14
# Execution count [9.90e-01]
movl 56(%rsp), %r12d #[spill]
movl 48(%rsp), %r13d #[spill]
# LOE rbx r12d r13d
..B2.16: # Preds ..B2.15 ..B2.5
# Execution count [1.10e+00]
movq stdout(%rip), %rdi #189.13
# fflush(FILE *)
call fflush #189.13
# LOE rbx r12d r13d
..B2.17: # Preds ..B2.16 ..B2.3
# Execution count [5.00e+00]
movl $1140850688, %edi #191.9
..___tag_value_debug.91:
# MPI_Barrier(MPI_Comm)
call MPI_Barrier #191.9
..___tag_value_debug.92:
# LOE rbx r12d r13d
..B2.18: # Preds ..B2.17
# Execution count [5.00e+00]
incl %r12d #187.39
cmpl 40(%rbx), %r12d #187.25
jl ..B2.3 # Prob 82% #187.25
# LOE rbx r12d r13d
..B2.19: # Preds ..B2.18
# Execution count [9.00e-01]
movq (%rsp), %r12 #[spill]
.cfi_restore 12
movq 8(%rsp), %r13 #[spill]
.cfi_restore 13
movq 16(%rsp), %r14 #[spill]
.cfi_restore 14
movq 24(%rsp), %r15 #[spill]
.cfi_restore 15
movq 32(%rsp), %rbx #[spill]
.cfi_restore 3
movq 40(%rsp), %rbp #[spill]
.cfi_restore 6
# LOE rbx rbp r12 r13 r14 r15
..B2.20: # Preds ..B2.19 ..B2.1
# Execution count [1.00e+00]
addq $72, %rsp #193.1
.cfi_def_cfa_offset 8
ret #193.1
.align 16,0x90
# LOE
.cfi_endproc
# mark_end;
.type debug,@function
.size debug,.-debug
..LNdebug.1:
.data
# -- End debug
.text
.L_2__routine_start_solve_2:
# -- Begin solve
.text
# mark_begin;
.align 16,0x90
.globl solve
# --- solve(Solver *)
solve:
# parameter 1: %rdi
..B3.1: # Preds ..B3.0
# Execution count [1.00e+00]
.cfi_startproc
..___tag_value_solve.106:
..L107:
#196.1
pushq %rbp #196.1
.cfi_def_cfa_offset 16
movq %rsp, %rbp #196.1
.cfi_def_cfa 6, 16
.cfi_offset 6, -16
andq $-64, %rsp #196.1
pushq %r12 #196.1
pushq %r13 #196.1
pushq %r14 #196.1
pushq %r15 #196.1
pushq %rbx #196.1
subq $408, %rsp #196.1
movq %rdi, %r11 #196.1
vmovsd .L_2il0floatpacket.5(%rip), %xmm2 #212.29
xorl %esi, %esi #198.12
vmovsd .L_2il0floatpacket.9(%rip), %xmm9 #210.27
vmovsd (%r11), %xmm6 #208.21
vmulsd %xmm6, %xmm6, %xmm4 #208.34
vmulsd 72(%r11), %xmm2, %xmm3 #212.29
vdivsd %xmm4, %xmm9, %xmm11 #210.27
vmovsd 8(%r11), %xmm1 #209.21
vmulsd %xmm1, %xmm1, %xmm8 #209.34
vmulsd %xmm4, %xmm3, %xmm5 #212.42
vdivsd %xmm8, %xmm9, %xmm10 #211.27
vmovsd 64(%r11), %xmm0 #204.21
vmulsd %xmm8, %xmm5, %xmm7 #212.42
vaddsd %xmm9, %xmm0, %xmm1 #217.17
vfmadd231sd %xmm6, %xmm6, %xmm8 #212.56
vmulsd %xmm0, %xmm0, %xmm5 #215.27
vdivsd %xmm8, %xmm7, %xmm6 #212.56
movl 88(%r11), %eax #206.21
movl %eax, 288(%rsp) #206.21[spill]
.cfi_escape 0x10, 0x03, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xd8, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xf8, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xf0, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xe8, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0f, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x22
movl 24(%r11), %r13d #201.21
movl 28(%r11), %edx #202.21
movl 32(%r11), %r10d #203.21
movq 48(%r11), %r9 #213.21
movq 56(%r11), %rax #214.21
vcomisd %xmm5, %xmm1 #219.20
jb ..B3.32 # Prob 10% #219.20
# LOE rax r9 r11 edx esi r10d r13d xmm0 xmm1 xmm5 xmm6 xmm10 xmm11
..B3.2: # Preds ..B3.1
# Execution count [9.00e-01]
movl %r13d, %ecx #203.19
movl %r13d, %r8d #241.13
imull %r10d, %ecx #203.19
vxorpd %xmm4, %xmm4, %xmm4 #253.34
imull %r13d, %edx #253.34
vmovsd .L_2il0floatpacket.8(%rip), %xmm2 #226.49
vmovsd %xmm5, 264(%rsp) #223.9[spill]
vmovsd %xmm6, 328(%rsp) #223.9[spill]
vmovsd %xmm10, 336(%rsp) #223.9[spill]
shrl $1, %r8d #241.13
lea (%r13,%rcx), %edi #203.19
vcvtsi2sd %edx, %xmm4, %xmm4 #253.34
vmovdqu .L_2il0floatpacket.6(%rip), %ymm3 #235.13
vmovsd %xmm4, 272(%rsp) #223.9[spill]
movl %r8d, 176(%rsp) #241.13[spill]
lea (%rdi,%r10,2), %edx #203.19
movl %edx, 48(%rsp) #203.19[spill]
lea (%rcx,%r10,2), %r8d #203.19
movslq %edx, %rdx #242.17
movl %r10d, %ecx #246.9
movslq %r8d, %r8 #242.39
shrl $1, %ecx #246.9
movl %ecx, 240(%rsp) #246.9[spill]
movl $0, 280(%rsp) #219.5[spill]
lea (%r9,%rdx,8), %r15 #242.17
movq %r15, 88(%rsp) #242.17[spill]
lea (%r9,%r8,8), %r12 #242.39
movq %r12, 80(%rsp) #242.39[spill]
lea 24(%r15), %r14 #242.17
movq %r14, 128(%rsp) #242.17[spill]
lea 8(%r12), %rdi #242.39
movq %rdi, 104(%rsp) #242.39[spill]
subq %rdi, %r14 #241.13
movslq %r13d, %rdi #226.21
lea (%r13,%r13,2), %r12d #248.30
movl %r12d, 208(%rsp) #248.30[spill]
lea 4(%r13,%r13), %r12d #227.68
movl %r12d, 200(%rsp) #227.68[spill]
vmovsd %xmm11, 344(%rsp) #223.9[spill]
vmovsd %xmm0, 32(%rsp) #223.9[spill]
lea (,%rdi,8), %rcx #227.58
cmpq %rcx, %r14 #241.13
lea (%rax,%rdi,8), %rdx #226.21
movq %rdx, 304(%rsp) #226.21[spill]
movl $0, %edx #241.13
setg %dl #241.13
negq %r14 #241.13
xorl %eax, %eax #241.13
cmpq %rcx, %r14 #241.13
movq %rcx, 144(%rsp) #223.9[spill]
lea (%rcx,%rdi,8), %rbx #248.30
setg %al #241.13
lea (%r9,%rdi,8), %r14 #226.35
orl %eax, %edx #241.13
movl %r13d, %eax #235.13
addq %r9, %rbx #248.30
andl $-8, %eax #235.13
movq %rbx, 216(%rsp) #248.30[spill]
lea 24(%r14), %rbx #236.27
movl %eax, 40(%rsp) #235.13[spill]
lea 8(%r9), %rax #236.17
movq %rbx, 120(%rsp) #236.27[spill]
subq %rax, %rbx #235.13
movq %rbx, %r12 #235.13
lea 16(,%rdi,8), %r15 #226.21
negq %r12 #235.13
cmpq %rcx, %r12 #235.13
movl $0, %r12d #235.13
movq %rax, 136(%rsp) #236.17[spill]
setg %r12b #235.13
cmpq %rcx, %rbx #235.13
movl $0, %ebx #235.13
setg %bl #235.13
andq $63, %rax #235.13
movl %eax, 112(%rsp) #235.13[spill]
orl %ebx, %r12d #235.13
movl %eax, %ebx #235.13
negl %eax #235.13
andl $7, %ebx #235.13
movq %r15, 312(%rsp) #226.21[spill]
movl %r13d, %r15d #247.30
addl $64, %eax #235.13
negl %r15d #247.30
shrl $3, %eax #235.13
movl %r15d, 224(%rsp) #247.30[spill]
lea 2(%r13), %r15d #227.68
movl %ebx, 96(%rsp) #235.13[spill]
movl %eax, 296(%rsp) #235.13[spill]
movslq %r10d, %rbx #223.9
movl %r12d, 160(%rsp) #223.9[spill]
xorl %r12d, %r12d #223.9
movl %edx, 152(%rsp) #223.9[spill]
movq %rdi, 352(%rsp) #223.9[spill]
movl %r8d, 56(%rsp) #223.9[spill]
movq %r9, 320(%rsp) #223.9[spill]
movl %r10d, 248(%rsp) #223.9[spill]
movq %r11, 256(%rsp) #223.9[spill]
# LOE rbx r12 r14 esi r13d r15d xmm1
..B3.3: # Preds ..B3.30 ..B3.2
# Execution count [2.62e+00]
movl 280(%rsp), %eax #219.5[spill]
incl %eax #219.5
movl %eax, 280(%rsp) #219.5[spill]
lea -1(%rax), %edx #252.9
cmpl 288(%rsp), %edx #219.36[spill]
jge ..B3.31 # Prob 20% #219.36
# LOE rbx r12 r14 esi r13d r15d xmm1
..B3.4: # Preds ..B3.3
# Execution count [2.09e+00]
movq 256(%rsp), %rax #221.9[spill]
movl $738197504, %ecx #221.9
movl %ecx, 64(%rsp) #221.9
movl %ecx, 68(%rsp) #221.9
movl %ecx, 72(%rsp) #221.9
movl %ecx, 76(%rsp) #221.9
movl 36(%rax), %ecx #221.9
movq %r12, 368(%rsp) #220.9
lea 1(%rcx), %r10d #221.9
cmpl 40(%rax), %r10d #221.9
jge ..B3.8 # Prob 78% #221.9
# LOE rax rbx r12 r14 ecx r10d r13d r15d
..B3.5: # Preds ..B3.4
# Execution count [4.60e-01]
addq $-16, %rsp #221.9
movl %r10d, %ecx #221.9
lea 80(%rsp), %r8 #221.9
movl $1140850688, %r9d #221.9
movl 24(%rax), %esi #221.9
movl 32(%rax), %edx #221.9
movq 48(%rax), %rax #221.9
movl %edx, -56(%r8) #221.9[spill]
lea 2(%rsi), %r11d #221.9
movl %r11d, 16(%rsp) #221.9[spill]
imull %edx, %r11d #221.9
movl $1275070475, %edx #221.9
movslq %r11d, %r11 #221.9
movq %r8, (%rsp) #221.9
movl $1, %r8d #221.9
movq %rax, 32(%rsp) #221.9[spill]
movl %r10d, 40(%rsp) #221.9[spill]
lea 8(%rax,%r11,8), %rdi #221.9
vzeroupper #221.9
..___tag_value_solve.159:
# MPI_Isend(const void *, int, MPI_Datatype, int, int, MPI_Comm, MPI_Request *)
call MPI_Isend #221.9
..___tag_value_solve.160:
# LOE rbx r12 r14 r13d r15d
..B3.82: # Preds ..B3.5
# Execution count [4.60e-01]
movl 40(%rsp), %r10d #[spill]
movq 32(%rsp), %rax #[spill]
# LOE rax rbx r12 r14 r10d r13d r15d
..B3.6: # Preds ..B3.82
# Execution count [4.60e-01]
movl $1275070475, %edx #221.9
movl 24(%rsp), %r11d #221.9[spill]
movl %r10d, %ecx #221.9
incl %r11d #221.9
movl $2, %r8d #221.9
imull 16(%rsp), %r11d #221.9[spill]
movl $1140850688, %r9d #221.9
movslq %r11d, %r11 #221.9
lea 8(%rax,%r11,8), %rdi #221.9
lea 84(%rsp), %rax #221.9
movq %rax, (%rsp) #221.9
movq 188(%rax), %rax #221.9[spill]
movl 24(%rax), %esi #221.9
..___tag_value_solve.166:
# MPI_Irecv(void *, int, MPI_Datatype, int, int, MPI_Comm, MPI_Request *)
call MPI_Irecv #221.9
..___tag_value_solve.167:
# LOE rbx r12 r14 r13d r15d
..B3.83: # Preds ..B3.6
# Execution count [4.60e-01]
addq $16, %rsp #221.9
# LOE rbx r12 r14 r13d r15d
..B3.7: # Preds ..B3.83
# Execution count [4.60e-01]
movq 256(%rsp), %rax #221.9[spill]
movl 36(%rax), %ecx #221.9
# LOE rbx r12 r14 ecx r13d r15d
..B3.8: # Preds ..B3.4 ..B3.7
# Execution count [2.09e+00]
testl %ecx, %ecx #221.9
jle ..B3.11 # Prob 40% #221.9
# LOE rbx r12 r14 ecx r13d r15d
..B3.9: # Preds ..B3.8
# Execution count [1.25e+00]
addq $-16, %rsp #221.9
decl %ecx #221.9
movl $1275070475, %edx #221.9
lea 88(%rsp), %rax #221.9
movl $2, %r8d #221.9
movl $1140850688, %r9d #221.9
movq %rax, (%rsp) #221.9
movq 184(%rax), %rax #221.9[spill]
movl %ecx, 208(%rsp) #221.9[spill]
movq 48(%rax), %r10 #221.9
movslq 24(%rax), %r11 #221.9
movl 24(%rax), %esi #221.9
movq %r10, 200(%rsp) #221.9[spill]
lea 24(%r10,%r11,8), %rdi #221.9
vzeroupper #221.9
..___tag_value_solve.172:
# MPI_Isend(const void *, int, MPI_Datatype, int, int, MPI_Comm, MPI_Request *)
call MPI_Isend #221.9
..___tag_value_solve.173:
# LOE rbx r12 r14 r13d r15d
..B3.84: # Preds ..B3.9
# Execution count [1.25e+00]
movl 208(%rsp), %ecx #[spill]
# LOE rbx r12 r14 ecx r13d r15d
..B3.10: # Preds ..B3.84
# Execution count [1.25e+00]
movl $1275070475, %edx #221.9
movq 200(%rsp), %rdi #221.9[spill]
lea 92(%rsp), %rax #221.9
addq $8, %rdi #221.9
movl $1, %r8d #221.9
movl $1140850688, %r9d #221.9
movq %rax, (%rsp) #221.9
movq 180(%rax), %r10 #221.9[spill]
movl 24(%r10), %esi #221.9
..___tag_value_solve.177:
# MPI_Irecv(void *, int, MPI_Datatype, int, int, MPI_Comm, MPI_Request *)
call MPI_Irecv #221.9
..___tag_value_solve.178:
# LOE rbx r12 r14 r13d r15d
..B3.85: # Preds ..B3.10
# Execution count [1.25e+00]
addq $16, %rsp #221.9
# LOE rbx r12 r14 r13d r15d
..B3.11: # Preds ..B3.85 ..B3.8
# Execution count [2.09e+00]
movl $4, %edi #221.9
lea 64(%rsp), %rsi #221.9
movl $1, %edx #221.9
vzeroupper #221.9
..___tag_value_solve.179:
# MPI_Waitall(int, MPI_Request *, MPI_Status *)
call MPI_Waitall #221.9
..___tag_value_solve.180:
# LOE rbx r12 r14 r13d r15d
..B3.12: # Preds ..B3.11
# Execution count [2.09e+00]
movq %r12, %rcx #223.9
cmpl $0, 248(%rsp) #223.41[spill]
jle ..B3.75 # Prob 10% #223.41
# LOE rcx rbx r12 r14 r13d r15d
..B3.13: # Preds ..B3.12
# Execution count [9.42e+00]
testl %r13d, %r13d #224.40
jle ..B3.72 # Prob 10% #224.40
# LOE rcx rbx r12 r14 r13d r15d
..B3.14: # Preds ..B3.13
# Execution count [1.83e+00]
vmovsd .L_2il0floatpacket.8(%rip), %xmm0 #
vmovsd 328(%rsp), %xmm1 #[spill]
vmovsd 336(%rsp), %xmm2 #[spill]
vmovsd 344(%rsp), %xmm3 #[spill]
movq 304(%rsp), %rdi #[spill]
movq 312(%rsp), %rdx #[spill]
movq 320(%rsp), %rsi #[spill]
# LOE rdx rcx rbx rsi rdi r12 r14 r13d r15d xmm0 xmm1 xmm2 xmm3
..B3.15: # Preds ..B3.19 ..B3.14
# Execution count [9.42e+00]
movq %rcx, %r8 #227.58
movl %ecx, %eax #226.17
imulq %rdx, %r8 #227.58
imull %r15d, %eax #227.68
movl %r13d, 360(%rsp) #227.68[spill]
lea (%rax,%r13,2), %eax #227.68
movslq %eax, %rax #227.68
movq %r12, %r9 #227.58
movq 352(%rsp), %r13 #227.68[spill]
lea (%rsi,%r8), %r10 #227.38
lea (%rdi,%r8), %r11 #226.21
addq %r14, %r8 #227.58
lea (%rsi,%rax,8), %rax #227.68
.align 16,0x90
# LOE rax rdx rcx rbx rsi rdi r8 r9 r10 r11 r12 r13 r14 r15d xmm0 xmm1 xmm2 xmm3
..B3.16: # Preds ..B3.16 ..B3.15
# Execution count [5.23e+01]
vmovsd 24(%r8,%r9,8), %xmm10 #226.55
vmulsd %xmm10, %xmm0, %xmm7 #227.58
vmovsd 16(%r8,%r9,8), %xmm6 #226.35
vmovsd 8(%r10,%r9,8), %xmm4 #227.38
vsubsd %xmm7, %xmm6, %xmm8 #226.55
vsubsd %xmm7, %xmm4, %xmm5 #227.58
vaddsd 32(%r8,%r9,8), %xmm8, %xmm12 #226.65
vaddsd 40(%rax,%r9,8), %xmm5, %xmm9 #227.68
vfnmadd213sd 24(%r11,%r9,8), %xmm3, %xmm12 #226.17
vfnmadd231sd %xmm2, %xmm9, %xmm12 #226.17
vfnmadd231sd %xmm1, %xmm12, %xmm10 #229.17
vmovsd %xmm10, 24(%r8,%r9,8) #229.17
vmovsd 368(%rsp), %xmm11 #230.17
incq %r9 #224.13
vfmadd213sd %xmm11, %xmm12, %xmm12 #230.17
vmovsd %xmm12, 368(%rsp) #230.17
cmpq %r13, %r9 #224.13
jb ..B3.16 # Prob 82% #224.13
# LOE rax rdx rcx rbx rsi rdi r8 r9 r10 r11 r12 r13 r14 r15d xmm0 xmm1 xmm2 xmm3
..B3.17: # Preds ..B3.16
# Execution count [9.42e+00]
incq %rcx #223.9
movl 360(%rsp), %r13d #[spill]
cmpq %rbx, %rcx #223.9
jae ..B3.35 # Prob 18% #223.9
# LOE rdx rcx rbx rsi rdi r12 r14 r13d r15d xmm0 xmm1 xmm2 xmm3
..B3.19: # Preds ..B3.74 ..B3.17 ..B3.21
# Execution count [8.43e+00]
testl %r13d, %r13d #224.40
jg ..B3.15 # Prob 90% #224.40
# LOE rdx rcx rbx rsi rdi r12 r14 r13d r15d xmm0 xmm1 xmm2 xmm3
..B3.21: # Preds ..B3.19
# Execution count [7.19e-01]: Infreq
incq %rcx #223.9
cmpq %rbx, %rcx #223.9
jb ..B3.19 # Prob 82% #223.9
# LOE rdx rcx rbx rsi rdi r12 r14 r13d r15d xmm0 xmm1 xmm2 xmm3
..B3.22: # Preds ..B3.21
# Execution count [1.29e-01]: Infreq
vmovsd %xmm1, 328(%rsp) #[spill]
vmovsd %xmm2, 336(%rsp) #[spill]
vmovsd %xmm3, 344(%rsp) #[spill]
movq %rsi, 320(%rsp) #[spill]
# LOE rbx r12 r14 r13d r15d
..B3.23: # Preds ..B3.73 ..B3.72 ..B3.22 ..B3.44 ..B3.40
#
# Execution count [2.09e+00]: Infreq
xorl %r8d, %r8d #246.9
movl $1, %edx #246.9
xorl %ecx, %ecx #247.13
cmpl $0, 240(%rsp) #246.9[spill]
je ..B3.27 # Prob 10% #246.9
# LOE rbx r12 r14 edx ecx r8d r13d r15d
..B3.24: # Preds ..B3.23
# Execution count [1.88e+00]: Infreq
movl 200(%rsp), %edi #[spill]
movl 208(%rsp), %esi #[spill]
movl 240(%rsp), %eax #[spill]
movq 320(%rsp), %rdx #[spill]
.align 16,0x90
# LOE rdx rbx r12 r14 eax ecx esi edi r8d r13d r15d
..B3.25: # Preds ..B3.25 ..B3.24
# Execution count [5.23e+00]: Infreq
movl %ecx, %r9d #247.30
incl %r8d #246.9
subl %r13d, %r9d #247.30
lea (%r9,%r13,2), %r10d #247.30
movslq %r10d, %r10 #247.30
lea (%rcx,%r13,2), %r9d #248.30
movslq %r9d, %r9 #248.30
movq 24(%rdx,%r10,8), %r11 #247.30
movq %r11, 16(%rdx,%r10,8) #247.13
lea (%rsi,%rcx), %r10d #227.68
movq 16(%rdx,%r9,8), %r11 #248.30
addl %edi, %ecx #246.9
movslq %r10d, %r10 #248.30
movq %r11, 24(%rdx,%r9,8) #248.13
movq 40(%rdx,%r9,8), %r11 #247.30
movq %r11, 32(%rdx,%r9,8) #247.13
movq 32(%rdx,%r10,8), %r9 #248.30
movq %r9, 40(%rdx,%r10,8) #248.13
cmpl %eax, %r8d #246.9
jb ..B3.25 # Prob 64% #246.9
# LOE rdx rbx r12 r14 eax ecx esi edi r8d r13d r15d
..B3.26: # Preds ..B3.25
# Execution count [1.88e+00]: Infreq
lea 1(%r8,%r8), %edx #247.13
# LOE rbx r12 r14 edx r13d r15d
..B3.27: # Preds ..B3.26 ..B3.23
# Execution count [2.09e+00]: Infreq
lea -1(%rdx), %eax #246.9
cmpl 248(%rsp), %eax #246.9[spill]
jae ..B3.29 # Prob 10% #246.9
# LOE rbx r12 r14 edx r13d r15d
..B3.28: # Preds ..B3.27
# Execution count [1.88e+00]: Infreq
imull %r15d, %edx #247.30
addl 224(%rsp), %edx #247.30[spill]
movslq %edx, %rdx #247.30
movq 216(%rsp), %rsi #248.30[spill]
movq 8(%r14,%rdx,8), %rax #247.30
movq %rax, (%r14,%rdx,8) #247.13
movq (%rsi,%rdx,8), %rcx #248.30
movq %rcx, 8(%rsi,%rdx,8) #248.13
# LOE rbx r12 r14 r13d r15d
..B3.29: # Preds ..B3.27 ..B3.44 ..B3.76 ..B3.77 ..B3.78
# ..B3.28
# Execution count [2.09e+00]: Infreq
movl $1, %edx #251.9
lea 368(%rsp), %rdi #251.9
movl $1275070475, %ecx #251.9
lea 232(%rsp), %rsi #251.9
movl $1476395011, %r8d #251.9
movl $1140850688, %r9d #251.9
vzeroupper #251.9
..___tag_value_solve.203:
# MPI_Allreduce(const void *, void *, int, MPI_Datatype, MPI_Op, MPI_Comm)
call MPI_Allreduce #251.9
..___tag_value_solve.204:
# LOE rbx r12 r14 r13d r15d
..B3.30: # Preds ..B3.29
# Execution count [2.09e+00]: Infreq
vmovsd 232(%rsp), %xmm0 #252.15
vdivsd 272(%rsp), %xmm0, %xmm1 #253.34[spill]
vsqrtsd %xmm1, %xmm1, %xmm1 #253.15
movl 280(%rsp), %esi #259.9[spill]
vcomisd 264(%rsp), %xmm1 #219.20[spill]
jae ..B3.3 # Prob 82% #219.20
# LOE rbx r12 r14 esi r13d r15d xmm1
..B3.31: # Preds ..B3.3 ..B3.30
# Execution count [3.77e-01]: Infreq
vmovsd 32(%rsp), %xmm0 #[spill]
movq 256(%rsp), %r11 #[spill]
# LOE r11 esi xmm0 xmm1
..B3.32: # Preds ..B3.31 ..B3.1
# Execution count [1.00e+00]: Infreq
cmpl $0, 36(%r11) #262.25
jne ..B3.34 # Prob 78% #262.25
# LOE esi xmm0 xmm1
..B3.33: # Preds ..B3.32
# Execution count [2.20e-01]: Infreq
movl $.L_2__STRING.6, %edi #263.9
xorl %eax, %eax #263.9
vmovsd %xmm1, 368(%rsp) #217.5
vzeroupper #263.9
vmovsd %xmm1, (%rsp) #263.9[spill]
vmovsd %xmm0, 32(%rsp) #263.9[spill]
..___tag_value_solve.211:
# printf(const char *__restrict__, ...)
call printf #263.9
..___tag_value_solve.212:
# LOE
..B3.89: # Preds ..B3.33
# Execution count [2.20e-01]: Infreq
vmovsd 32(%rsp), %xmm0 #[spill]
vmovsd (%rsp), %xmm1 #[spill]
# LOE xmm0 xmm1
..B3.34: # Preds ..B3.89 ..B3.32
# Execution count [1.00e+00]: Infreq
vcmpgtsd %xmm1, %xmm0, %xmm0 #266.16
vmovd %xmm0, %eax #266.16
negl %eax #266.16
vzeroupper #266.16
addq $408, %rsp #266.16
.cfi_restore 3
popq %rbx #266.16
.cfi_restore 15
popq %r15 #266.16
.cfi_restore 14
popq %r14 #266.16
.cfi_restore 13
popq %r13 #266.16
.cfi_restore 12
popq %r12 #266.16
movq %rbp, %rsp #266.16
popq %rbp #266.16
.cfi_def_cfa 7, 8
.cfi_restore 6
ret #266.16
.cfi_def_cfa 6, 16
.cfi_escape 0x10, 0x03, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xd8, 0xff, 0xff, 0xff, 0x22
.cfi_offset 6, -16
.cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xf8, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xf0, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xe8, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0f, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x22
# LOE
..B3.35: # Preds ..B3.17
# Execution count [1.70e+00]: Infreq
movq 256(%rsp), %rax #234.13[spill]
vmovsd %xmm1, 328(%rsp) #[spill]
vmovsd %xmm2, 336(%rsp) #[spill]
movl 36(%rax), %ecx #234.13
vmovsd %xmm3, 344(%rsp) #[spill]
movq %rsi, 320(%rsp) #[spill]
testl %ecx, %ecx #234.29
jne ..B3.40 # Prob 50% #234.29
# LOE rax rbx r12 r14 ecx r13d r15d
..B3.36: # Preds ..B3.35 ..B3.76
# Execution count [5.23e-03]: Infreq
cmpl $12, %r13d #235.13
jle ..B3.64 # Prob 10% #235.13
# LOE rbx r12 r14 ecx r13d r15d
..B3.37: # Preds ..B3.36
# Execution count [1.05e+00]: Infreq
cmpl $0, 160(%rsp) #235.13[spill]
je ..B3.45 # Prob 10% #235.13
# LOE rbx r12 r14 ecx r13d r15d
..B3.38: # Preds ..B3.37
# Execution count [1.05e+00]: Infreq
movq 136(%rsp), %rdi #235.13[spill]
movq 120(%rsp), %rsi #235.13[spill]
movq 144(%rsp), %rdx #235.13[spill]
movl %ecx, 168(%rsp) #235.13[spill]
call __intel_skx_avx512_memcpy #235.13
# LOE rbx r12 r14 r13d r15d
..B3.88: # Preds ..B3.38
# Execution count [1.05e+00]: Infreq
movl 168(%rsp), %ecx #[spill]
# LOE rbx r12 r14 ecx r13d r15d
..B3.39: # Preds ..B3.88 ..B3.53
# Execution count [1.05e+00]: Infreq
movq 256(%rsp), %rax #240.30[spill]
movl 40(%rax), %edx #240.30
decl %edx #240.45
cmpl %edx, %ecx #240.45
je ..B3.42 # Prob 50% #240.45
jmp ..B3.44 # Prob 100% #240.45
# LOE rbx r12 r14 r13d r15d
..B3.40: # Preds ..B3.35
# Execution count [8.48e-01]: Infreq
movl 40(%rax), %edx #240.30
decl %edx #240.45
cmpl %edx, %ecx #240.45
jne ..B3.23 # Prob 50% #240.45
# LOE rbx r12 r14 r13d r15d
..B3.41: # Preds ..B3.40 ..B3.62 ..B3.78
# Execution count [0.00e+00]: Infreq
cmpl $12, %r13d #241.13
jle ..B3.66 # Prob 10% #241.13
# LOE rbx r12 r14 r13d r15d
..B3.42: # Preds ..B3.41 ..B3.39
# Execution count [1.05e+00]: Infreq
cmpl $0, 152(%rsp) #241.13[spill]
je ..B3.66 # Prob 10% #241.13
# LOE rbx r12 r14 r13d r15d
..B3.43: # Preds ..B3.42
# Execution count [1.05e+00]: Infreq
vzeroupper #241.13
movq 128(%rsp), %rdi #241.13[spill]
movq 104(%rsp), %rsi #241.13[spill]
movq 144(%rsp), %rdx #241.13[spill]
call __intel_skx_avx512_memcpy #241.13
# LOE rbx r12 r14 r13d r15d
..B3.44: # Preds ..B3.43 ..B3.70 ..B3.39 ..B3.62 ..B3.71
#
# Execution count [5.38e-01]: Infreq
cmpl $0, 248(%rsp) #246.41[spill]
jg ..B3.23 # Prob 50% #246.41
jmp ..B3.29 # Prob 100% #246.41
# LOE rbx r12 r14 r13d r15d
..B3.45: # Preds ..B3.37
# Execution count [1.99e-01]: Infreq
cmpl $30, %r13d #235.13
jl ..B3.63 # Prob 10% #235.13
# LOE rbx r12 r14 ecx r13d r15d
..B3.46: # Preds ..B3.45
# Execution count [1.05e+00]: Infreq
movl 112(%rsp), %edi #235.13[spill]
cmpl $0, 96(%rsp) #235.13[spill]
je ..B3.48 # Prob 50% #235.13
# LOE rbx r12 r14 ecx edi r13d r15d
..B3.47: # Preds ..B3.46
# Execution count [5.23e-01]: Infreq
xorl %edi, %edi #235.13
jmp ..B3.50 # Prob 100% #235.13
# LOE rbx r12 r14 ecx edi r13d r15d
..B3.48: # Preds ..B3.46
# Execution count [5.23e-01]: Infreq
cmpl $0, 112(%rsp) #235.13[spill]
je ..B3.50 # Prob 50% #235.13
# LOE rbx r12 r14 ecx edi r13d r15d
..B3.49: # Preds ..B3.48
# Execution count [5.81e+00]: Infreq
movl 296(%rsp), %edi #235.13[spill]
cmpl %edi, %r13d #235.13
cmovl %r13d, %edi #235.13
# LOE rbx r12 r14 ecx edi r13d r15d
..B3.50: # Preds ..B3.47 ..B3.49 ..B3.48
# Execution count [1.16e+00]: Infreq
movl %r13d, %esi #235.13
subl %edi, %esi #235.13
andl $7, %esi #235.13
negl %esi #235.13
addl %r13d, %esi #235.13
cmpl $1, %edi #235.13
jb ..B3.54 # Prob 50% #235.13
# LOE rbx r12 r14 ecx esi edi r13d r15d
..B3.51: # Preds ..B3.50
# Execution count [1.05e+00]: Infreq
vmovdqu .L_2il0floatpacket.7(%rip), %ymm1 #235.13
movq %r12, %rdx #235.13
vmovdqu .L_2il0floatpacket.6(%rip), %ymm3 #235.13
vpbroadcastd %edi, %ymm0 #235.13
movslq %edi, %rax #235.13
movq 320(%rsp), %r8 #235.13[spill]
# LOE rax rdx rbx r8 r12 r14 ecx esi edi r13d r15d ymm0 ymm1 ymm3
..B3.52: # Preds ..B3.52 ..B3.51
# Execution count [5.81e+00]: Infreq
vpcmpgtd %ymm1, %ymm0, %k1 #235.13
vpaddd %ymm3, %ymm1, %ymm1 #235.13
vmovupd 24(%r14,%rdx,8), %zmm2{%k1}{z} #236.27
vmovupd %zmm2, 8(%r8,%rdx,8){%k1} #236.17
addq $8, %rdx #235.13
cmpq %rax, %rdx #235.13
jb ..B3.52 # Prob 82% #235.13
# LOE rax rdx rbx r8 r12 r14 ecx esi edi r13d r15d ymm0 ymm1 ymm3
..B3.53: # Preds ..B3.52
# Execution count [1.05e+00]: Infreq
cmpl %edi, %r13d #235.13
je ..B3.39 # Prob 10% #235.13
# LOE rbx r12 r14 ecx esi edi r13d r15d
..B3.54: # Preds ..B3.53 ..B3.50 ..B3.63
# Execution count [5.81e+00]: Infreq
movslq %esi, %rax #235.13
lea 8(%rdi), %edx #235.13
cmpl %edx, %esi #235.13
jl ..B3.58 # Prob 50% #235.13
# LOE rax rbx r12 r14 ecx edi r13d r15d
..B3.55: # Preds ..B3.54
# Execution count [1.05e+00]: Infreq
movslq %edi, %rdi #235.13
movq 320(%rsp), %rdx #235.13[spill]
# LOE rax rdx rbx rdi r12 r14 ecx r13d r15d
..B3.56: # Preds ..B3.56 ..B3.55
# Execution count [5.81e+00]: Infreq
vmovups 24(%r14,%rdi,8), %zmm0 #236.27
vmovupd %zmm0, 8(%rdx,%rdi,8) #236.17
addq $8, %rdi #235.13
cmpq %rax, %rdi #235.13
jb ..B3.56 # Prob 82% #235.13
# LOE rax rdx rbx rdi r12 r14 ecx r13d r15d
..B3.58: # Preds ..B3.56 ..B3.54 ..B3.65
# Execution count [1.16e+00]: Infreq
cmpq 352(%rsp), %rax #235.13[spill]
jae ..B3.62 # Prob 10% #235.13
# LOE rax rbx r12 r14 ecx r13d r15d
..B3.59: # Preds ..B3.58
# Execution count [1.05e+00]: Infreq
movq 352(%rsp), %rsi #[spill]
movq 320(%rsp), %rdi #[spill]
# LOE rax rbx rsi rdi r12 r14 ecx r13d r15d
..B3.60: # Preds ..B3.60 ..B3.59
# Execution count [5.81e+00]: Infreq
movq 24(%r14,%rax,8), %rdx #236.27
movq %rdx, 8(%rdi,%rax,8) #236.17
incq %rax #235.13
cmpq %rsi, %rax #235.13
jb ..B3.60 # Prob 82% #235.13
# LOE rax rbx rsi rdi r12 r14 ecx r13d r15d
..B3.62: # Preds ..B3.60 ..B3.58
# Execution count [0.00e+00]: Infreq
movq 256(%rsp), %rax #240.30[spill]
movl 40(%rax), %edx #240.30
decl %edx #240.45
cmpl %edx, %ecx #240.45
je ..B3.41 # Prob 50% #240.45
jmp ..B3.44 # Prob 100% #240.45
# LOE rbx r12 r14 r13d r15d
..B3.63: # Preds ..B3.64 ..B3.45
# Execution count [1.05e-01]: Infreq
movl 40(%rsp), %esi #235.13[spill]
xorl %edi, %edi #235.13
jmp ..B3.54 # Prob 100% #235.13
# LOE rbx r12 r14 ecx esi edi r13d r15d
..B3.64: # Preds ..B3.36
# Execution count [9.42e-01]: Infreq
cmpl $8, %r13d #235.13
jge ..B3.63 # Prob 90% #235.13
# LOE rbx r12 r14 ecx r13d r15d
..B3.65: # Preds ..B3.64
# Execution count [1.05e-01]: Infreq
movq %r12, %rax #235.13
jmp ..B3.58 # Prob 100% #235.13
# LOE rax rbx r12 r14 ecx r13d r15d
..B3.66: # Preds ..B3.42 ..B3.41
# Execution count [1.16e+00]: Infreq
movl $1, %edx #241.13
xorl %r8d, %r8d #241.13
cmpl $0, 176(%rsp) #241.13[spill]
je ..B3.70 # Prob 10% #241.13
# LOE rbx r12 r14 edx r8d r13d r15d
..B3.67: # Preds ..B3.66
# Execution count [1.05e+00]: Infreq
movl 56(%rsp), %edi #[spill]
movl 48(%rsp), %esi #[spill]
movl 176(%rsp), %ecx #[spill]
movq 320(%rsp), %rdx #[spill]
# LOE rdx rbx r12 r14 ecx esi edi r8d r13d r15d
..B3.68: # Preds ..B3.68 ..B3.67
# Execution count [2.91e+00]: Infreq
lea (%rdi,%r8,2), %r9d #203.19
movslq %r9d, %r9 #242.39
lea (%rsi,%r8,2), %eax #203.19
movslq %eax, %rax #242.17
incl %r8d #241.13
movq 8(%rdx,%r9,8), %r10 #242.39
movq %r10, 24(%rdx,%rax,8) #242.17
movq 16(%rdx,%r9,8), %r11 #242.39
movq %r11, 32(%rdx,%rax,8) #242.17
cmpl %ecx, %r8d #241.13
jb ..B3.68 # Prob 64% #241.13
# LOE rdx rbx r12 r14 ecx esi edi r8d r13d r15d
..B3.69: # Preds ..B3.68
# Execution count [1.05e+00]: Infreq
lea 1(%r8,%r8), %edx #242.17
# LOE rbx r12 r14 edx r13d r15d
..B3.70: # Preds ..B3.69 ..B3.66
# Execution count [1.16e+00]: Infreq
lea -1(%rdx), %eax #241.13
cmpl %r13d, %eax #241.13
jae ..B3.44 # Prob 10% #241.13
# LOE rbx r12 r14 edx r13d r15d
..B3.71: # Preds ..B3.70
# Execution count [1.05e+00]: Infreq
movslq %edx, %rdx #242.39
movq 80(%rsp), %rax #242.39[spill]
movq 88(%rsp), %rsi #242.17[spill]
movq (%rax,%rdx,8), %rcx #242.39
movq %rcx, 16(%rsi,%rdx,8) #242.17
jmp ..B3.44 # Prob 100% #242.17
# LOE rbx r12 r14 r13d r15d
..B3.72: # Preds ..B3.13
# Execution count [1.88e-01]: Infreq
cmpq $1, %rbx #223.9
jbe ..B3.23 # Prob 18% #223.9
# LOE rbx r12 r14 r13d r15d
..B3.73: # Preds ..B3.72
# Execution count [1.54e-01]: Infreq
movl $2, %ecx #223.9
cmpq $2, %rbx #223.9
jbe ..B3.23 # Prob 18% #223.9
# LOE rcx rbx r12 r14 r13d r15d
..B3.74: # Preds ..B3.73
# Execution count [7.84e+00]: Infreq
vmovsd .L_2il0floatpacket.8(%rip), %xmm0 #
vmovsd 328(%rsp), %xmm1 #[spill]
vmovsd 336(%rsp), %xmm2 #[spill]
vmovsd 344(%rsp), %xmm3 #[spill]
movq 304(%rsp), %rdi #[spill]
movq 312(%rsp), %rdx #[spill]
movq 320(%rsp), %rsi #[spill]
jmp ..B3.19 # Prob 100% #
# LOE rdx rcx rbx rsi rdi r12 r14 r13d r15d xmm0 xmm1 xmm2 xmm3
..B3.75: # Preds ..B3.12
# Execution count [2.09e-01]: Infreq
movq 256(%rsp), %rax #234.13[spill]
movl 36(%rax), %ecx #234.13
testl %ecx, %ecx #234.29
jne ..B3.77 # Prob 50% #234.29
# LOE rax rbx r12 r14 ecx r13d r15d
..B3.76: # Preds ..B3.75
# Execution count [1.05e-01]: Infreq
testl %r13d, %r13d #235.40
jg ..B3.36 # Prob 50% #235.40
jmp ..B3.29 # Prob 100% #235.40
# LOE rbx r12 r14 ecx r13d r15d
..B3.77: # Preds ..B3.75
# Execution count [6.23e-01]: Infreq
movl 40(%rax), %edx #240.30
decl %edx #240.45
cmpl %edx, %ecx #240.45
jne ..B3.29 # Prob 50% #240.45
# LOE rbx r12 r14 r13d r15d
..B3.78: # Preds ..B3.77
# Execution count [3.11e-01]: Infreq
testl %r13d, %r13d #241.40
jg ..B3.41 # Prob 50% #241.40
jmp ..B3.29 # Prob 100% #241.40
.align 16,0x90
# LOE rbx r12 r14 r13d r15d
.cfi_endproc
# mark_end;
.type solve,@function
.size solve,.-solve
..LNsolve.2:
.data
# -- End solve
.text
.L_2__routine_start_solveRB_3:
# -- Begin solveRB
.text
# mark_begin;
.align 16,0x90
.globl solveRB
# --- solveRB(Solver *)
solveRB:
# parameter 1: %rdi
..B4.1: # Preds ..B4.0
# Execution count [1.00e+00]
.cfi_startproc
..___tag_value_solveRB.273:
..L274:
#273.1
pushq %r14 #273.1
.cfi_def_cfa_offset 16
.cfi_offset 14, -16
pushq %r15 #273.1
.cfi_def_cfa_offset 24
.cfi_offset 15, -24
pushq %rbx #273.1
.cfi_def_cfa_offset 32
.cfi_offset 3, -32
pushq %rbp #273.1
.cfi_def_cfa_offset 40
.cfi_offset 6, -40
subq $280, %rsp #273.1
.cfi_def_cfa_offset 320
movq %rdi, %r10 #273.1
vmovsd .L_2il0floatpacket.5(%rip), %xmm2 #289.29
xorl %esi, %esi #275.12
vmovsd .L_2il0floatpacket.9(%rip), %xmm9 #287.27
vmovsd (%r10), %xmm6 #285.21
vmulsd %xmm6, %xmm6, %xmm4 #285.34
vmulsd 72(%r10), %xmm2, %xmm3 #289.29
vdivsd %xmm4, %xmm9, %xmm11 #287.27
vmovsd 8(%r10), %xmm1 #286.21
vmulsd %xmm1, %xmm1, %xmm8 #286.34
vmulsd %xmm4, %xmm3, %xmm5 #289.42
vdivsd %xmm8, %xmm9, %xmm10 #288.27
vmovsd 64(%r10), %xmm0 #281.21
vmulsd %xmm0, %xmm0, %xmm4 #293.26
vaddsd %xmm9, %xmm0, %xmm1 #295.17
vmulsd %xmm8, %xmm5, %xmm7 #289.42
vfmadd231sd %xmm6, %xmm6, %xmm8 #289.56
vcomisd %xmm4, %xmm1 #297.20
vdivsd %xmm8, %xmm7, %xmm5 #289.56
movl 24(%r10), %ecx #278.21
movl 28(%r10), %eax #279.21
movl 32(%r10), %r15d #280.21
movl 88(%r10), %r9d #283.21
movq 48(%r10), %r8 #290.21
movq 56(%r10), %rdi #291.21
jb ..B4.38 # Prob 10% #297.20
# LOE rdi r8 r10 r12 r13 eax ecx esi r9d r15d xmm0 xmm1 xmm4 xmm5 xmm10 xmm11
..B4.2: # Preds ..B4.1
# Execution count [9.00e-01]
imull %ecx, %eax #331.37
vxorpd %xmm3, %xmm3, %xmm3 #331.37
vmovsd .L_2il0floatpacket.8(%rip), %xmm2 #309.48
vmovsd %xmm4, 176(%rsp) #322.35[spill]
vmovsd %xmm5, 232(%rsp) #322.35[spill]
vmovsd %xmm10, 224(%rsp) #322.35[spill]
movslq %ecx, %rcx #321.35
movl %r15d, %ebx #325.9
movl %ecx, %edx #326.30
lea 2(%rcx), %r11d #310.67
negl %edx #326.30
xorl %r14d, %r14d #297.5
movl %edx, 104(%rsp) #326.30[spill]
lea 4(%rcx,%rcx), %edx #310.67
movl %edx, 48(%rsp) #310.67[spill]
movl %ecx, %edx #280.19
imull %r15d, %edx #280.19
lea (,%rcx,8), %rbp #321.35
vmovsd %xmm11, 216(%rsp) #322.35[spill]
vmovsd %xmm0, 24(%rsp) #322.35[spill]
vcvtsi2sd %eax, %xmm3, %xmm3 #331.37
movl %r11d, 88(%rsp) #310.67[spill]
lea (%rbp,%rcx,8), %rax #327.30
addq %r8, %rax #327.30
lea (%rcx,%rdx), %r11d #322.13
movq %rax, 96(%rsp) #327.30[spill]
lea (%rcx,%rcx,2), %eax #327.30
movl %eax, 40(%rsp) #327.30[spill]
movl %ecx, %eax #320.9
shrl $1, %eax #320.9
lea (%rdx,%r15,2), %edx #322.35
movl %eax, 160(%rsp) #320.9[spill]
lea (%r11,%r15,2), %eax #280.19
movslq %eax, %rax #322.13
lea (%r8,%rcx,8), %rbp #321.35
movslq %edx, %rdx #322.35
shrl $1, %ebx #325.9
vmovsd %xmm3, 168(%rsp) #322.35[spill]
movl %edx, 56(%rsp) #322.35[spill]
lea (%r8,%rax,8), %r11 #322.13
movq %r11, 80(%rsp) #322.13[spill]
lea (%r8,%rdx,8), %r11 #322.35
movq %r11, 72(%rsp) #322.35[spill]
movl %eax, 64(%rsp) #322.35[spill]
movq %rdi, 200(%rsp) #322.35[spill]
movq %r8, 240(%rsp) #322.35[spill]
movl %r9d, 184(%rsp) #322.35[spill]
movq %r10, 208(%rsp) #322.35[spill]
movq %r12, 8(%rsp) #322.35[spill]
.cfi_offset 12, -312
movl %ecx, %r12d #322.35
movq %r13, 16(%rsp) #322.35[spill]
.cfi_offset 13, -304
xorl %r13d, %r13d #322.35
# LOE rbp r13 ebx esi r12d r14d r15d xmm1
..B4.3: # Preds ..B4.36 ..B4.2
# Execution count [2.62e+00]
incl %r14d #297.5
lea -1(%r14), %eax #330.9
cmpl 184(%rsp), %eax #297.36[spill]
jge ..B4.37 # Prob 20% #297.36
# LOE rbp r13 ebx esi r12d r14d r15d xmm1
..B4.4: # Preds ..B4.3
# Execution count [2.09e+00]
movl $1, %eax #299.9
xorl %esi, %esi #301.9
movq %r13, 264(%rsp) #298.9
movl %esi, 192(%rsp) #303.13[spill]
movq %rbp, 112(%rsp) #303.13[spill]
movl %ebx, 120(%rsp) #303.13[spill]
movl %r14d, 128(%rsp) #303.13[spill]
movl %r12d, 248(%rsp) #303.13[spill]
movl %eax, %r12d #303.13
movq 208(%rsp), %r13 #303.13[spill]
# LOE r13 r12d r15d
..B4.5: # Preds ..B4.41 ..B4.20 ..B4.4
# Execution count [4.19e+00]
movl 36(%r13), %r14d #303.13
movl $738197504, %edx #303.13
movl %edx, 136(%rsp) #303.13
movl %r12d, %ebx #302.13
movl %edx, 140(%rsp) #303.13
movl %edx, 144(%rsp) #303.13
lea 1(%r14), %ebp #303.13
movl %edx, 148(%rsp) #303.13
cmpl 40(%r13), %ebp #303.13
jge ..B4.9 # Prob 78% #303.13
# LOE r13 ebx ebp r12d r14d r15d
..B4.6: # Preds ..B4.5
# Execution count [9.21e-01]
movl 24(%r13), %esi #303.13
addq $-16, %rsp #303.13
.cfi_def_cfa_offset 336
movq 48(%r13), %rax #303.13
lea 152(%rsp), %r14 #303.13
movl $1275070475, %edx #303.13
movl %ebp, %ecx #303.13
movl $1, %r8d #303.13
lea 2(%rsi), %r10d #303.13
movq %r14, (%rsp) #303.13
movl $1140850688, %r9d #303.13
movl 32(%r13), %r14d #303.13
movl %r10d, 16(%rsp) #303.13[spill]
imull %r14d, %r10d #303.13
movslq %r10d, %r10 #303.13
movq %rax, 48(%rsp) #303.13[spill]
lea 8(%rax,%r10,8), %rdi #303.13
..___tag_value_solveRB.318:
# MPI_Isend(const void *, int, MPI_Datatype, int, int, MPI_Comm, MPI_Request *)
call MPI_Isend #303.13
..___tag_value_solveRB.319:
# LOE r13 ebx ebp r12d r14d r15d
..B4.46: # Preds ..B4.6
# Execution count [9.21e-01]
movq 48(%rsp), %rax #[spill]
# LOE rax r13 ebx ebp r12d r14d r15d
..B4.7: # Preds ..B4.46
# Execution count [9.21e-01]
incl %r14d #303.13
imull 16(%rsp), %r14d #303.13[spill]
movl $1275070475, %edx #303.13
movslq %r14d, %r14 #303.13
movl %ebp, %ecx #303.13
movl $2, %r8d #303.13
movl $1140850688, %r9d #303.13
movl 24(%r13), %esi #303.13
lea 156(%rsp), %r10 #303.13
movq %r10, (%rsp) #303.13
lea 8(%rax,%r14,8), %rdi #303.13
..___tag_value_solveRB.322:
# MPI_Irecv(void *, int, MPI_Datatype, int, int, MPI_Comm, MPI_Request *)
call MPI_Irecv #303.13
..___tag_value_solveRB.323:
# LOE r13 ebx r12d r15d
..B4.47: # Preds ..B4.7
# Execution count [9.21e-01]
addq $16, %rsp #303.13
.cfi_def_cfa_offset 320
# LOE r13 ebx r12d r15d
..B4.8: # Preds ..B4.47
# Execution count [9.21e-01]
movl 36(%r13), %r14d #303.13
# LOE r13 ebx r12d r14d r15d
..B4.9: # Preds ..B4.8 ..B4.5
# Execution count [4.19e+00]
testl %r14d, %r14d #303.13
jle ..B4.12 # Prob 40% #303.13
# LOE r13 ebx r12d r14d r15d
..B4.10: # Preds ..B4.9
# Execution count [2.50e+00]
addq $-16, %rsp #303.13
.cfi_def_cfa_offset 336
decl %r14d #303.13
movslq 24(%r13), %r10 #303.13
lea 160(%rsp), %rbp #303.13
movl $1275070475, %edx #303.13
movl %r14d, %ecx #303.13
movl $2, %r8d #303.13
movl $1140850688, %r9d #303.13
movq %rbp, (%rsp) #303.13
movq 48(%r13), %rbp #303.13
movl 24(%r13), %esi #303.13
lea 24(%rbp,%r10,8), %rdi #303.13
..___tag_value_solveRB.326:
# MPI_Isend(const void *, int, MPI_Datatype, int, int, MPI_Comm, MPI_Request *)
call MPI_Isend #303.13
..___tag_value_solveRB.327:
# LOE rbp r13 ebx r12d r14d r15d
..B4.11: # Preds ..B4.10
# Execution count [2.50e+00]
addq $8, %rbp #303.13
movq %rbp, %rdi #303.13
movl $1275070475, %edx #303.13
movl %r14d, %ecx #303.13
movl $1, %r8d #303.13
movl $1140850688, %r9d #303.13
lea 164(%rsp), %r10 #303.13
movl 24(%r13), %esi #303.13
movq %r10, (%rsp) #303.13
..___tag_value_solveRB.328:
# MPI_Irecv(void *, int, MPI_Datatype, int, int, MPI_Comm, MPI_Request *)
call MPI_Irecv #303.13
..___tag_value_solveRB.329:
# LOE r13 ebx r12d r15d
..B4.49: # Preds ..B4.11
# Execution count [2.50e+00]
addq $16, %rsp #303.13
.cfi_def_cfa_offset 320
# LOE r13 ebx r12d r15d
..B4.12: # Preds ..B4.49 ..B4.9
# Execution count [4.19e+00]
movl $4, %edi #303.13
lea 136(%rsp), %rsi #303.13
movl $1, %edx #303.13
..___tag_value_solveRB.331:
# MPI_Waitall(int, MPI_Request *, MPI_Status *)
call MPI_Waitall #303.13
..___tag_value_solveRB.332:
# LOE r13 ebx r12d r15d
..B4.13: # Preds ..B4.12
# Execution count [4.19e+00]
xorl %edx, %edx #305.13
xorl %eax, %eax #306.30
testl %r15d, %r15d #305.45
jle ..B4.41 # Prob 10% #305.45
# LOE r13 eax edx ebx r12d r15d
..B4.14: # Preds ..B4.13
# Execution count [3.77e+00]
movl %r12d, 256(%rsp) #[spill]
movq %r13, 208(%rsp) #[spill]
vmovsd .L_2il0floatpacket.8(%rip), %xmm0 #
vmovsd 232(%rsp), %xmm1 #[spill]
vmovsd 224(%rsp), %xmm2 #[spill]
vmovsd 216(%rsp), %xmm3 #[spill]
movq 200(%rsp), %r8 #[spill]
movq 240(%rsp), %rdi #[spill]
movl 248(%rsp), %r9d #[spill]
# LOE rdi r8 eax edx ebx r9d r15d xmm0 xmm1 xmm2 xmm3
..B4.15: # Preds ..B4.19 ..B4.14
# Execution count [2.09e+01]
xorl %r10d, %r10d #306.17
cmpl %r9d, %ebx #306.46
jg ..B4.19 # Prob 10% #306.46
# LOE rdi r8 eax edx ebx r9d r10d r15d xmm0 xmm1 xmm2 xmm3
..B4.16: # Preds ..B4.15
# Execution count [1.88e+01]
movslq %ebx, %rbx #308.32
movl %r9d, %ebp #278.21
subl %ebx, %ebp #278.21
lea (%r9,%rax), %r11d #308.32
addl $2, %ebp #278.21
lea (%rax,%r9,2), %r12d #310.37
shrl $1, %ebp #278.21
lea (%rdi,%rbx,8), %r13 #309.34
lea (%r8,%rbx,8), %rsi #308.32
.align 16,0x90
# LOE rsi rdi r8 r13 eax edx ebx ebp r9d r10d r11d r12d r15d xmm0 xmm1 xmm2 xmm3
..B4.17: # Preds ..B4.17 ..B4.16
# Execution count [1.05e+02]
lea (%r11,%r10,2), %ecx #310.57
movslq %ecx, %rcx #308.32
lea (%r12,%r10,2), %r14d #310.37
movslq %r14d, %r14 #310.37
vmovsd 16(%r13,%rcx,8), %xmm10 #310.57
vmulsd %xmm10, %xmm0, %xmm7 #310.57
vmovsd 24(%r13,%rcx,8), %xmm4 #309.34
vmovsd 32(%r13,%r14,8), %xmm6 #310.37
lea (%rax,%r10,2), %r14d #310.67
movslq %r14d, %r14 #310.67
incl %r10d #306.17
vsubsd %xmm7, %xmm4, %xmm5 #309.54
vsubsd %xmm7, %xmm6, %xmm8 #310.57
vaddsd 8(%r13,%rcx,8), %xmm5, %xmm12 #309.64
vaddsd (%r13,%r14,8), %xmm8, %xmm9 #310.67
vfnmadd213sd 16(%rsi,%rcx,8), %xmm3, %xmm12 #310.82
vfnmadd231sd %xmm2, %xmm9, %xmm12 #310.82
vfnmadd231sd %xmm1, %xmm12, %xmm10 #312.21
vmovsd %xmm10, 16(%r13,%rcx,8) #312.21
vmovsd 264(%rsp), %xmm11 #313.21
vfmadd213sd %xmm11, %xmm12, %xmm12 #313.21
vmovsd %xmm12, 264(%rsp) #313.21
cmpl %ebp, %r10d #306.17
jb ..B4.17 # Prob 82% #306.17
# LOE rsi rdi r8 r13 eax edx ebx ebp r9d r10d r11d r12d r15d xmm0 xmm1 xmm2 xmm3
..B4.19: # Preds ..B4.17 ..B4.15
# Execution count [2.09e+01]
negl %ebx #315.27
incl %edx #305.13
addl $3, %ebx #315.27
lea 2(%rax,%r9), %eax #305.13
cmpl %r15d, %edx #305.13
jb ..B4.15 # Prob 82% #305.13
# LOE rdi r8 eax edx ebx r9d r15d xmm0 xmm1 xmm2 xmm3
..B4.20: # Preds ..B4.19
# Execution count [3.77e+00]
movl 256(%rsp), %r12d #[spill]
movl 192(%rsp), %edx #301.9[spill]
negl %r12d #317.23
incl %edx #301.9
addl $3, %r12d #317.23
movq 208(%rsp), %r13 #[spill]
movl %edx, 192(%rsp) #301.9[spill]
cmpl $2, %edx #301.9
jb ..B4.5 # Prob 50% #301.9
# LOE r13 r12d r15d
..B4.21: # Preds ..B4.20
# Execution count [1.88e+00]
movl 248(%rsp), %r12d #[spill]
movq 112(%rsp), %rbp #[spill]
movl 120(%rsp), %ebx #[spill]
movl 128(%rsp), %r14d #[spill]
movq %r13, 208(%rsp) #[spill]
xorl %r13d, %r13d #
testl %r12d, %r12d #320.36
jle ..B4.29 # Prob 50% #320.36
# LOE rbp r13 ebx r12d r14d r15d
..B4.22: # Preds ..B4.21 ..B4.42
# Execution count [2.09e+00]
movl $1, %edx #320.9
xorl %esi, %esi #320.9
cmpl $0, 160(%rsp) #320.9[spill]
je ..B4.26 # Prob 10% #320.9
# LOE rbp r13 edx ebx esi r12d r14d r15d
..B4.23: # Preds ..B4.22
# Execution count [1.88e+00]
movl 56(%rsp), %r9d #[spill]
movl 64(%rsp), %r8d #[spill]
movl 160(%rsp), %edi #[spill]
movq 240(%rsp), %rcx #[spill]
.align 16,0x90
# LOE rcx rbp r13 ebx esi edi r8d r9d r12d r14d r15d
..B4.24: # Preds ..B4.24 ..B4.23
# Execution count [5.23e+00]
movslq %esi, %r10 #321.35
lea (%r9,%rsi,2), %eax #280.19
shlq $4, %r10 #321.35
lea (%r8,%rsi,2), %edx #280.19
movslq %eax, %rax #322.35
incl %esi #320.9
movslq %edx, %rdx #322.13
movq 24(%rbp,%r10), %r11 #321.35
movq %r11, 8(%rcx,%r10) #321.13
movq 8(%rcx,%rax,8), %r11 #322.35
movq %r11, 24(%rcx,%rdx,8) #322.13
movq 32(%rbp,%r10), %r11 #321.35
movq %r11, 16(%rcx,%r10) #321.13
movq 16(%rcx,%rax,8), %r10 #322.35
movq %r10, 32(%rcx,%rdx,8) #322.13
cmpl %edi, %esi #320.9
jb ..B4.24 # Prob 64% #320.9
# LOE rcx rbp r13 ebx esi edi r8d r9d r12d r14d r15d
..B4.25: # Preds ..B4.24
# Execution count [1.88e+00]
lea 1(%rsi,%rsi), %edx #321.13
# LOE rbp r13 edx ebx r12d r14d r15d
..B4.26: # Preds ..B4.25 ..B4.22
# Execution count [2.09e+00]
lea -1(%rdx), %eax #320.9
cmpl %r12d, %eax #320.9
jae ..B4.28 # Prob 10% #320.9
# LOE rbp r13 edx ebx r12d r14d r15d
..B4.27: # Preds ..B4.26
# Execution count [1.88e+00]
movslq %edx, %rdx #321.35
movq 240(%rsp), %rsi #321.13[spill]
movq 72(%rsp), %rdi #322.35[spill]
movq 80(%rsp), %r9 #322.13[spill]
movq 16(%rbp,%rdx,8), %rax #321.35
movq %rax, (%rsi,%rdx,8) #321.13
movq (%rdi,%rdx,8), %r8 #322.35
movq %r8, 16(%r9,%rdx,8) #322.13
# LOE rbp r13 ebx r12d r14d r15d
..B4.28: # Preds ..B4.26 ..B4.27
# Execution count [1.05e+00]
testl %r15d, %r15d #325.41
jle ..B4.35 # Prob 50% #325.41
# LOE rbp r13 ebx r12d r14d r15d
..B4.29: # Preds ..B4.21 ..B4.28
# Execution count [2.09e+00]
xorl %edi, %edi #325.9
movl $1, %esi #325.9
xorl %edx, %edx #326.13
testl %ebx, %ebx #325.9
je ..B4.33 # Prob 10% #325.9
# LOE rbp r13 edx ebx esi edi r12d r14d r15d
..B4.30: # Preds ..B4.29
# Execution count [1.88e+00]
movl 48(%rsp), %esi #[spill]
movl 40(%rsp), %ecx #[spill]
movq 240(%rsp), %rax #[spill]
.align 16,0x90
# LOE rax rbp r13 edx ecx ebx esi edi r12d r14d r15d
..B4.31: # Preds ..B4.31 ..B4.30
# Execution count [5.23e+00]
movl %edx, %r8d #326.30
lea (%rdx,%r12,2), %r11d #327.30
subl %r12d, %r8d #326.30
incl %edi #325.9
movslq %r11d, %r11 #327.30
lea (%r8,%r12,2), %r9d #326.30
movslq %r9d, %r9 #326.30
movq 24(%rax,%r9,8), %r10 #326.30
movq %r10, 16(%rax,%r9,8) #326.13
lea (%rcx,%rdx), %r10d #310.67
movq 16(%rax,%r11,8), %r8 #327.30
addl %esi, %edx #325.9
movslq %r10d, %r10 #327.30
movq %r8, 24(%rax,%r11,8) #327.13
movq 40(%rax,%r11,8), %r9 #326.30
movq %r9, 32(%rax,%r11,8) #326.13
movq 32(%rax,%r10,8), %r8 #327.30
movq %r8, 40(%rax,%r10,8) #327.13
cmpl %ebx, %edi #325.9
jb ..B4.31 # Prob 64% #325.9
# LOE rax rbp r13 edx ecx ebx esi edi r12d r14d r15d
..B4.32: # Preds ..B4.31
# Execution count [1.88e+00]
lea 1(%rdi,%rdi), %esi #326.13
# LOE rbp r13 ebx esi r12d r14d r15d
..B4.33: # Preds ..B4.32 ..B4.29
# Execution count [2.09e+00]
lea -1(%rsi), %eax #325.9
cmpl %r15d, %eax #325.9
jae ..B4.35 # Prob 10% #325.9
# LOE rbp r13 ebx esi r12d r14d r15d
..B4.34: # Preds ..B4.33
# Execution count [1.88e+00]
imull 88(%rsp), %esi #326.30[spill]
addl 104(%rsp), %esi #326.30[spill]
movslq %esi, %rsi #326.30
movq 96(%rsp), %rdi #327.30[spill]
movq 8(%rbp,%rsi,8), %rax #326.30
movq %rax, (%rbp,%rsi,8) #326.13
movq (%rdi,%rsi,8), %rdx #327.30
movq %rdx, 8(%rdi,%rsi,8) #327.13
# LOE rbp r13 ebx r12d r14d r15d
..B4.35: # Preds ..B4.33 ..B4.28 ..B4.42 ..B4.34
# Execution count [2.09e+00]
movl $1, %edx #329.9
lea 264(%rsp), %rdi #329.9
movl $1275070475, %ecx #329.9
lea 152(%rsp), %rsi #329.9
movl $1476395011, %r8d #329.9
movl $1140850688, %r9d #329.9
..___tag_value_solveRB.364:
# MPI_Allreduce(const void *, void *, int, MPI_Datatype, MPI_Op, MPI_Comm)
call MPI_Allreduce #329.9
..___tag_value_solveRB.365:
# LOE rbp r13 ebx r12d r14d r15d
..B4.36: # Preds ..B4.35
# Execution count [2.09e+00]
vmovsd 152(%rsp), %xmm0 #330.15
movl %r14d, %esi #335.9
vdivsd 168(%rsp), %xmm0, %xmm1 #331.37[spill]
vcomisd 176(%rsp), %xmm1 #297.20[spill]
jae ..B4.3 # Prob 82% #297.20
# LOE rbp r13 ebx esi r12d r14d r15d xmm1
..B4.37: # Preds ..B4.3 ..B4.36
# Execution count [3.77e-01]
vmovsd 24(%rsp), %xmm0 #[spill]
movq 208(%rsp), %r10 #[spill]
movq 8(%rsp), %r12 #[spill]
.cfi_restore 12
movq 16(%rsp), %r13 #[spill]
.cfi_restore 13
# LOE r10 r12 r13 esi xmm0 xmm1
..B4.38: # Preds ..B4.37 ..B4.1
# Execution count [1.00e+00]
cmpl $0, 36(%r10) #338.25
jne ..B4.40 # Prob 78% #338.25
# LOE r12 r13 esi xmm0 xmm1
..B4.39: # Preds ..B4.38
# Execution count [2.20e-01]
movl $.L_2__STRING.6, %edi #339.9
xorl %eax, %eax #339.9
vmovsd %xmm1, 264(%rsp) #295.5
vmovsd %xmm1, (%rsp) #339.9[spill]
vmovsd %xmm0, 24(%rsp) #339.9[spill]
..___tag_value_solveRB.375:
# printf(const char *__restrict__, ...)
call printf #339.9
..___tag_value_solveRB.376:
# LOE r12 r13
..B4.50: # Preds ..B4.39
# Execution count [2.20e-01]
vmovsd 24(%rsp), %xmm0 #[spill]
vmovsd (%rsp), %xmm1 #[spill]
# LOE r12 r13 xmm0 xmm1
..B4.40: # Preds ..B4.50 ..B4.38
# Execution count [1.00e+00]
vcmpgtsd %xmm1, %xmm0, %xmm0 #342.16
vmovd %xmm0, %eax #342.16
negl %eax #342.16
addq $280, %rsp #342.16
.cfi_def_cfa_offset 40
.cfi_restore 6
popq %rbp #342.16
.cfi_def_cfa_offset 32
.cfi_restore 3
popq %rbx #342.16
.cfi_def_cfa_offset 24
.cfi_restore 15
popq %r15 #342.16
.cfi_def_cfa_offset 16
.cfi_restore 14
popq %r14 #342.16
.cfi_def_cfa_offset 8
ret #342.16
.cfi_def_cfa_offset 320
.cfi_offset 3, -32
.cfi_offset 6, -40
.cfi_offset 12, -312
.cfi_offset 13, -304
.cfi_offset 14, -16
.cfi_offset 15, -24
# LOE
..B4.41: # Preds ..B4.13
# Execution count [4.19e-01]: Infreq
movl 192(%rsp), %edx #301.9[spill]
negl %r12d #317.23
incl %edx #301.9
addl $3, %r12d #317.23
movl %edx, 192(%rsp) #301.9[spill]
cmpl $2, %edx #301.9
jb ..B4.5 # Prob 50% #301.9
# LOE r13 r12d r15d
..B4.42: # Preds ..B4.41
# Execution count [2.09e-01]: Infreq
movl 248(%rsp), %r12d #[spill]
movq 112(%rsp), %rbp #[spill]
movl 120(%rsp), %ebx #[spill]
movl 128(%rsp), %r14d #[spill]
movq %r13, 208(%rsp) #[spill]
xorl %r13d, %r13d #
testl %r12d, %r12d #320.36
jg ..B4.22 # Prob 50% #320.36
jmp ..B4.35 # Prob 100% #320.36
.align 16,0x90
# LOE rbp r13 ebx r12d r14d r15d
.cfi_endproc
# mark_end;
.type solveRB,@function
.size solveRB,.-solveRB
..LNsolveRB.3:
.data
# -- End solveRB
.text
.L_2__routine_start_solveRBA_4:
# -- Begin solveRBA
.text
# mark_begin;
.align 16,0x90
.globl solveRBA
# --- solveRBA(Solver *)
solveRBA:
# parameter 1: %rdi
..B5.1: # Preds ..B5.0
# Execution count [1.00e+00]
.cfi_startproc
..___tag_value_solveRBA.404:
..L405:
#349.1
pushq %rbp #349.1
.cfi_def_cfa_offset 16
movq %rsp, %rbp #349.1
.cfi_def_cfa 6, 16
.cfi_offset 6, -16
andq $-64, %rsp #349.1
pushq %r12 #349.1
pushq %r13 #349.1
pushq %r14 #349.1
pushq %r15 #349.1
pushq %rbx #349.1
subq $472, %rsp #349.1
.cfi_escape 0x10, 0x03, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xd8, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xf8, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xf0, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xe8, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0f, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x22
xorl %ebx, %ebx #351.12
vmovsd 72(%rdi), %xmm0 #358.21
vmovsd (%rdi), %xmm5 #361.21
vmovsd .L_2il0floatpacket.5(%rip), %xmm9 #365.29
vmulsd %xmm5, %xmm5, %xmm3 #361.34
vmulsd %xmm9, %xmm0, %xmm2 #365.29
vmovsd 8(%rdi), %xmm1 #362.21
vmulsd %xmm1, %xmm1, %xmm7 #362.34
vmulsd %xmm3, %xmm2, %xmm4 #365.42
vmovsd 64(%rdi), %xmm8 #357.21
vmovsd .L_2il0floatpacket.9(%rip), %xmm17 #363.27
vmulsd %xmm8, %xmm8, %xmm1 #370.26
vaddsd %xmm17, %xmm8, %xmm2 #372.17
vdivsd %xmm7, %xmm17, %xmm15 #364.27
vmulsd %xmm7, %xmm4, %xmm6 #365.42
vfmadd231sd %xmm5, %xmm5, %xmm7 #365.56
vdivsd %xmm3, %xmm17, %xmm16 #363.27
movq 56(%rdi), %rax #367.21
movl 24(%rdi), %ecx #354.21
movl 28(%rdi), %r8d #355.21
movl 32(%rdi), %r13d #356.21
movl 88(%rdi), %r12d #359.21
movq 48(%rdi), %rdx #366.21
vmovsd 80(%rdi), %xmm4 #369.20
movq %rdi, 376(%rsp) #349.1[spill]
vdivsd %xmm7, %xmm6, %xmm14 #365.56
movq %rax, 320(%rsp) #367.21[spill]
vcomisd %xmm1, %xmm2 #374.20
jb ..B5.58 # Prob 10% #374.20
# LOE rdx ecx ebx r8d r12d r13d xmm0 xmm1 xmm2 xmm4 xmm9 xmm14 xmm15 xmm16 xmm17
..B5.2: # Preds ..B5.1
# Execution count [9.00e-01]
vmulsd %xmm4, %xmm9, %xmm3 #395.64
movl %r13d, %edi #404.9
imull %ecx, %r8d #409.37
vmulsd .L_2il0floatpacket.10(%rip), %xmm4, %xmm5 #396.65
vmovups .L_2il0floatpacket.11(%rip), %zmm9 #386.34
vmovups .L_2il0floatpacket.12(%rip), %zmm8 #386.34
vfnmadd213sd %xmm17, %xmm4, %xmm3 #395.70
vmovups .L_2il0floatpacket.13(%rip), %zmm7 #386.48
vmulsd %xmm5, %xmm4, %xmm12 #396.71
vmovdqu .L_2il0floatpacket.6(%rip), %ymm5 #383.17
vmovdqu .L_2il0floatpacket.7(%rip), %ymm4 #383.17
vdivsd %xmm3, %xmm17, %xmm13 #395.70
movslq %ecx, %rcx #385.32
vxorpd %xmm3, %xmm3, %xmm3 #409.37
vbroadcastsd %xmm16, %zmm10 #363.19
lea (%rcx,%rcx,2), %r15d #406.30
vbroadcastsd %xmm15, %zmm6 #364.19
vmovsd .L_2il0floatpacket.8(%rip), %xmm11 #386.48
vmovsd %xmm0, 360(%rsp) #401.35[spill]
movl %r15d, 56(%rsp) #406.30[spill]
movl %ecx, %r15d #356.19
imull %r13d, %r15d #356.19
lea 2(%rcx), %r14d #387.67
vmovsd %xmm12, 312(%rsp) #401.35[spill]
vmovsd %xmm13, 40(%rsp) #401.35[spill]
vmovsd %xmm14, 352(%rsp) #401.35[spill]
vmovsd %xmm15, 336(%rsp) #401.35[spill]
movl %r14d, 208(%rsp) #387.67[spill]
lea (%rcx,%r15), %r14d #401.13
movl %ecx, %r10d #405.30
lea (%r14,%r13,2), %esi #356.19
movslq %esi, %rsi #401.13
lea (%r15,%r13,2), %r15d #401.35
movslq %r15d, %r15 #401.35
negl %r10d #405.30
vcvtsi2sd %r8d, %xmm3, %xmm3 #409.37
vmovups %zmm6, 64(%rsp) #401.35[spill]
vmovups %zmm10, 128(%rsp) #401.35[spill]
vmovsd %xmm16, 344(%rsp) #401.35[spill]
movl %r10d, 216(%rsp) #405.30[spill]
lea (,%rcx,8), %r8 #400.35
movl %ecx, %r10d #399.9
lea (%r8,%rcx,8), %r8 #406.30
shrl $1, %edi #404.9
lea 4(%rcx,%rcx), %eax #387.67
movq %rcx, 416(%rsp) #385.32[spill]
xorl %r11d, %r11d #374.5
movl %eax, 48(%rsp) #387.67[spill]
addq %rdx, %r8 #406.30
shrl $1, %r10d #399.9
lea (%rdx,%rcx,8), %r9 #400.35
movl %r15d, 192(%rsp) #401.35[spill]
lea (%rdx,%rsi,8), %r14 #401.13
movl %esi, 200(%rsp) #401.35[spill]
lea (%rdx,%r15,8), %rax #401.35
# LOE rax rdx r8 r9 r14 ecx ebx edi r10d r11d r12d r13d xmm1 xmm2 xmm3
..B5.3: # Preds ..B5.56 ..B5.2
# Execution count [2.62e+00]
incl %r11d #374.5
lea -1(%r11), %r15d #409.9
cmpl %r12d, %r15d #374.36
jge ..B5.57 # Prob 20% #374.36
# LOE rax rdx r8 r9 r14 ecx ebx edi r10d r11d r12d r13d xmm1 xmm2 xmm3
..B5.4: # Preds ..B5.3
# Execution count [2.09e+00]
xorb %r15b, %r15b #378.9
movl $1, %ebx #376.9
vxorpd %xmm0, %xmm0, %xmm0 #375.9
vmovsd %xmm0, 328(%rsp) #380.13[spill]
vmovsd %xmm3, 304(%rsp) #380.13[spill]
vmovsd %xmm1, 296(%rsp) #380.13[spill]
movb %r15b, 392(%rsp) #380.13[spill]
movq %rax, 248(%rsp) #380.13[spill]
movq %r14, 256(%rsp) #380.13[spill]
movl %r10d, 264(%rsp) #380.13[spill]
movq %r9, 272(%rsp) #380.13[spill]
movq %r8, 280(%rsp) #380.13[spill]
movl %edi, 288(%rsp) #380.13[spill]
movl %r11d, 384(%rsp) #380.13[spill]
movq %rdx, 408(%rsp) #380.13[spill]
movl %r12d, 240(%rsp) #380.13[spill]
movl %ecx, 400(%rsp) #380.13[spill]
# LOE ebx r13d
..B5.5: # Preds ..B5.41 ..B5.4
# Execution count [4.19e+00]
movq 376(%rsp), %rax #380.13[spill]
movl $738197504, %r12d #380.13
movl %r12d, 224(%rsp) #380.13
movl %r12d, 228(%rsp) #380.13
movl 36(%rax), %r15d #380.13
movl %r12d, 232(%rsp) #380.13
movl %r12d, 236(%rsp) #380.13
movl %ebx, %r12d #379.13
lea 1(%r15), %r14d #380.13
cmpl 40(%rax), %r14d #380.13
jge ..B5.9 # Prob 78% #380.13
# LOE rax ebx r12d r13d r14d r15d
..B5.6: # Preds ..B5.5
# Execution count [9.21e-01]
addq $-16, %rsp #380.13
movl $1275070475, %edx #380.13
movl %r14d, %ecx #380.13
lea 240(%rsp), %r15 #380.13
movl $1, %r8d #380.13
movl $1140850688, %r9d #380.13
movq %r15, (%rsp) #380.13
vzeroupper #380.13
movl 24(%rax), %esi #380.13
movl 32(%rax), %r15d #380.13
movq 48(%rax), %rax #380.13
movq %rax, 48(%rsp) #380.13[spill]
lea 2(%rsi), %r10d #380.13
movl %r10d, 40(%rsp) #380.13[spill]
imull %r15d, %r10d #380.13
movslq %r10d, %r10 #380.13
lea 8(%rax,%r10,8), %rdi #380.13
..___tag_value_solveRBA.448:
# MPI_Isend(const void *, int, MPI_Datatype, int, int, MPI_Comm, MPI_Request *)
call MPI_Isend #380.13
..___tag_value_solveRBA.449:
# LOE ebx r12d r13d r14d r15d
..B5.65: # Preds ..B5.6
# Execution count [9.21e-01]
movq 48(%rsp), %rax #[spill]
# LOE rax ebx r12d r13d r14d r15d
..B5.7: # Preds ..B5.65
# Execution count [9.21e-01]
incl %r15d #380.13
imull 40(%rsp), %r15d #380.13[spill]
movl $1275070475, %edx #380.13
movslq %r15d, %r15 #380.13
movl %r14d, %ecx #380.13
movl $2, %r8d #380.13
movl $1140850688, %r9d #380.13
lea 8(%rax,%r15,8), %rdi #380.13
lea 244(%rsp), %rax #380.13
movq %rax, (%rsp) #380.13
movq 148(%rax), %r10 #380.13[spill]
movl 24(%r10), %esi #380.13
..___tag_value_solveRBA.453:
# MPI_Irecv(void *, int, MPI_Datatype, int, int, MPI_Comm, MPI_Request *)
call MPI_Irecv #380.13
..___tag_value_solveRBA.454:
# LOE ebx r12d r13d
..B5.66: # Preds ..B5.7
# Execution count [9.21e-01]
addq $16, %rsp #380.13
# LOE ebx r12d r13d
..B5.8: # Preds ..B5.66
# Execution count [9.21e-01]
movq 376(%rsp), %rax #380.13[spill]
movl 36(%rax), %r15d #380.13
# LOE rax ebx r12d r13d r15d
..B5.9: # Preds ..B5.8 ..B5.5
# Execution count [4.19e+00]
testl %r15d, %r15d #380.13
jle ..B5.12 # Prob 40% #380.13
# LOE rax ebx r12d r13d r15d
..B5.10: # Preds ..B5.9
# Execution count [2.50e+00]
addq $-16, %rsp #380.13
decl %r15d #380.13
movl $1275070475, %edx #380.13
lea 248(%rsp), %r14 #380.13
movl %r15d, %ecx #380.13
movl $2, %r8d #380.13
movl $1140850688, %r9d #380.13
movq %r14, (%rsp) #380.13
movq %rax, %r10 #380.13
vzeroupper #380.13
movq 48(%r10), %r14 #380.13
movslq 24(%r10), %rax #380.13
movl 24(%r10), %esi #380.13
lea 24(%r14,%rax,8), %rdi #380.13
..___tag_value_solveRBA.456:
# MPI_Isend(const void *, int, MPI_Datatype, int, int, MPI_Comm, MPI_Request *)
call MPI_Isend #380.13
..___tag_value_solveRBA.457:
# LOE r14 ebx r12d r13d r15d
..B5.11: # Preds ..B5.10
# Execution count [2.50e+00]
addq $8, %r14 #380.13
movq %r14, %rdi #380.13
lea 252(%rsp), %rax #380.13
movl $1275070475, %edx #380.13
movl %r15d, %ecx #380.13
movl $1, %r8d #380.13
movl $1140850688, %r9d #380.13
movq %rax, (%rsp) #380.13
movq 140(%rax), %r10 #380.13[spill]
movl 24(%r10), %esi #380.13
..___tag_value_solveRBA.459:
# MPI_Irecv(void *, int, MPI_Datatype, int, int, MPI_Comm, MPI_Request *)
call MPI_Irecv #380.13
..___tag_value_solveRBA.460:
# LOE ebx r12d r13d
..B5.68: # Preds ..B5.11
# Execution count [2.50e+00]
addq $16, %rsp #380.13
# LOE ebx r12d r13d
..B5.12: # Preds ..B5.68 ..B5.9
# Execution count [4.19e+00]
movl $4, %edi #380.13
lea 224(%rsp), %rsi #380.13
movl $1, %edx #380.13
vzeroupper #380.13
..___tag_value_solveRBA.461:
# MPI_Waitall(int, MPI_Request *, MPI_Status *)
call MPI_Waitall #380.13
..___tag_value_solveRBA.462:
# LOE ebx r12d r13d
..B5.13: # Preds ..B5.12
# Execution count [4.19e+00]
xorl %r14d, %r14d #382.13
testl %r13d, %r13d #382.45
jle ..B5.37 # Prob 10% #382.45
# LOE ebx r12d r13d r14d
..B5.14: # Preds ..B5.13
# Execution count [3.77e+00]
vmovsd 360(%rsp), %xmm0 #389.41[spill]
vmulsd 352(%rsp), %xmm0, %xmm11 #389.41[spill]
vmovups 64(%rsp), %zmm5 #389.41[spill]
vmovups 128(%rsp), %zmm1 #389.41[spill]
vmovdqu .L_2il0floatpacket.7(%rip), %ymm8 #389.41
vmovdqu .L_2il0floatpacket.6(%rip), %ymm12 #389.41
vmovsd 328(%rsp), %xmm6 #389.41[spill]
vmovsd 336(%rsp), %xmm13 #389.41[spill]
vmovsd 344(%rsp), %xmm7 #389.41[spill]
vmovsd .L_2il0floatpacket.8(%rip), %xmm9 #389.41
vbroadcastsd %xmm11, %zmm0 #389.41
vmovups .L_2il0floatpacket.13(%rip), %zmm4 #389.41
vmovups .L_2il0floatpacket.12(%rip), %zmm3 #389.41
vmovups .L_2il0floatpacket.11(%rip), %zmm2 #389.41
movl %ebx, 368(%rsp) #389.41[spill]
movq 320(%rsp), %r10 #389.41[spill]
movq 408(%rsp), %rdi #389.41[spill]
movl 400(%rsp), %r9d #389.41[spill]
# LOE rdi r10 r9d r12d r13d r14d xmm6 xmm7 xmm9 xmm11 xmm13 ymm8 ymm12 zmm0 zmm1 zmm2 zmm3 zmm4 zmm5
..B5.15: # Preds ..B5.35 ..B5.14
# Execution count [2.09e+01]
cmpl %r9d, %r12d #383.46
jg ..B5.35 # Prob 50% #383.46
# LOE rdi r10 r9d r12d r13d r14d xmm6 xmm7 xmm9 xmm11 xmm13 ymm8 ymm12 zmm0 zmm1 zmm2 zmm3 zmm4 zmm5
..B5.16: # Preds ..B5.15
# Execution count [0.00e+00]
movq 416(%rsp), %rax #387.67[spill]
movl %r9d, %ecx #354.21
movslq %r14d, %r14 #385.32
subl %r12d, %ecx #354.21
movslq %r12d, %r11 #385.32
lea 2(%rax), %rdx #387.67
imulq %r14, %rdx #385.32
lea (%rax,%r11), %rbx #385.32
addq %rdx, %rbx #385.32
lea 8(%rdi,%rbx,8), %rsi #386.64
movq %rsi, 8(%rsp) #386.64[spill]
lea 16(%r10,%rbx,8), %rax #385.32
negq %rax #385.32
lea 2(%rcx), %ebx #354.21
movl %ebx, %r8d #383.17
addq %rsi, %rax #385.32
shrl $31, %r8d #383.17
lea 2(%rcx,%r8), %esi #354.21
sarl $1, %esi #354.21
movslq %esi, %r15 #354.21
shlq $4, %r15 #383.17
movq %r15, 16(%rsp) #383.17[spill]
lea -8(%r15), %rcx #383.17
movq %rcx, (%rsp) #383.17[spill]
cmpq %rcx, %rax #383.17
jge ..B5.18 # Prob 50% #383.17
# LOE rax rdx rdi r10 r11 r15 ebx esi r9d r12d r13d r14d xmm6 xmm7 xmm9 xmm11 xmm13 ymm8 ymm12 zmm0 zmm1 zmm2 zmm3 zmm4 zmm5
..B5.17: # Preds ..B5.16
# Execution count [0.00e+00]
movq %r15, %rcx #383.17
negq %rax #386.64
lea 8(%rcx), %r8 #383.17
cmpq %r8, %rax #383.17
jl ..B5.31 # Prob 50% #383.17
# LOE rdx rdi r10 r11 r15 ebx esi r9d r12d r13d r14d xmm6 xmm7 xmm9 xmm11 xmm13 ymm8 ymm12 zmm0 zmm1 zmm2 zmm3 zmm4 zmm5
..B5.18: # Preds ..B5.16 ..B5.17
# Execution count [0.00e+00]
movl 208(%rsp), %r8d #387.37[spill]
imull %r14d, %r8d #387.37
lea (%r8,%r9,2), %eax #387.37
movl %eax, 432(%rsp) #387.37[spill]
cltq #387.37
addq %r11, %rax #387.37
lea 32(%rdi,%rax,8), %rcx #387.37
negq %rcx #387.37
addq 8(%rsp), %rcx #387.37[spill]
cmpq (%rsp), %rcx #383.17[spill]
jge ..B5.20 # Prob 50% #383.17
# LOE rdx rcx rdi r10 r11 r15 ebx esi r8d r9d r12d r13d r14d xmm6 xmm7 xmm9 xmm11 xmm13 ymm8 ymm12 zmm0 zmm1 zmm2 zmm3 zmm4 zmm5
..B5.19: # Preds ..B5.18
# Execution count [0.00e+00]
movq %r15, %rax #383.17
negq %rcx #386.64
lea 8(%rax), %r15 #383.17
cmpq %r15, %rcx #383.17
jl ..B5.32 # Prob 50% #383.17
# LOE rdx rdi r10 r11 ebx esi r8d r9d r12d r13d r14d xmm6 xmm7 xmm9 xmm11 xmm13 ymm8 ymm12 zmm0 zmm1 zmm2 zmm3 zmm4 zmm5
..B5.20: # Preds ..B5.18 ..B5.19
# Execution count [0.00e+00]
addq %r11, %rdx #387.67
lea (%rdi,%rdx,8), %rax #387.67
movq 8(%rsp), %rdx #387.67[spill]
subq %rax, %rdx #387.67
movq %rdx, 8(%rsp) #387.67[spill]
cmpq (%rsp), %rdx #383.17[spill]
jge ..B5.22 # Prob 50% #383.17
# LOE rdx rdi r10 r11 ebx esi r8d r9d r12d r13d r14d xmm6 xmm7 xmm9 xmm11 xmm13 ymm8 ymm12 zmm0 zmm1 zmm2 zmm3 zmm4 zmm5
..B5.21: # Preds ..B5.20
# Execution count [0.00e+00]
movq %rdx, %rax #386.64
movq 16(%rsp), %rdx #383.17[spill]
negq %rax #386.64
addq $8, %rdx #383.17
cmpq %rdx, %rax #383.17
jl ..B5.32 # Prob 50% #383.17
# LOE rdi r10 r11 ebx esi r8d r9d r12d r13d r14d xmm6 xmm7 xmm9 xmm11 xmm13 ymm8 ymm12 zmm0 zmm1 zmm2 zmm3 zmm4 zmm5
..B5.22: # Preds ..B5.20 ..B5.21
# Execution count [9.42e+00]
vpxord %zmm10, %zmm10, %zmm10 #390.21
vmovaps %zmm10, %zmm14 #390.21
cmpl $16, %esi #383.17
jl ..B5.61 # Prob 10% #383.17
# LOE rdi r10 r11 esi r8d r9d r12d r13d r14d xmm6 xmm7 xmm9 xmm11 xmm13 ymm8 ymm12 zmm0 zmm1 zmm2 zmm3 zmm4 zmm5 zmm10 zmm14
..B5.23: # Preds ..B5.22
# Execution count [9.42e+00]
movslq %r12d, %r12 #386.64
lea (%rdi,%r11,8), %rbx #386.54
movl $85, %eax #385.32
lea (%r10,%r11,8), %r15 #385.32
movslq %r8d, %r11 #386.64
xorl %edx, %edx #383.17
addq %r12, %r11 #386.64
lea (%r9,%r8), %ecx #385.32
addq 416(%rsp), %r11 #386.64[spill]
kmovw %eax, %k1 #385.32
movl %esi, %eax #383.17
vmovdqu32 .L_2il0floatpacket.14(%rip), %zmm23 #385.32
vmovdqu32 .L_2il0floatpacket.15(%rip), %zmm24 #385.32
vmovdqu32 .L_2il0floatpacket.16(%rip), %zmm26 #385.32
movl %r13d, 424(%rsp) #366.21[spill]
andl $-16, %eax #383.17
movl 432(%rsp), %r9d #366.21[spill]
lea (%rdi,%r11,8), %r11 #366.21
# LOE rbx r10 r11 r15 eax edx ecx esi r8d r9d r12d r14d xmm6 xmm7 xmm9 xmm11 xmm13 ymm8 ymm12 zmm0 zmm1 zmm2 zmm3 zmm4 zmm5 zmm10 zmm14 zmm23 zmm24 zmm26 k1
..B5.24: # Preds ..B5.24 ..B5.23
# Execution count [5.23e+01]
vmovupd 16(%r11), %zmm19 #386.34
lea (%rcx,%rdx,2), %edi #385.32
vmovupd 80(%r11), %zmm27 #386.34
vmovupd 144(%r11), %zmm20 #386.34
vmovupd 208(%r11), %zmm30 #386.34
vmovupd 8(%r11), %zmm25 #386.34
vmovupd 72(%r11), %zmm18 #386.34
vmovupd 136(%r11), %zmm17 #386.34
vmovaps %zmm2, %zmm22 #386.34
lea (%r9,%rdx,2), %r13d #387.37
vpermi2pd %zmm27, %zmm19, %zmm22 #386.34
vpermt2pd %zmm27, %zmm3, %zmm19 #386.34
vmovupd 200(%r11), %zmm27 #386.34
vfnmadd231pd %zmm22, %zmm4, %zmm19 #386.54
movslq %edi, %rdi #385.32
addq $256, %r11 #383.17
vmovaps %zmm2, %zmm15 #386.34
vmovaps %zmm2, %zmm31 #386.64
vpermi2pd %zmm30, %zmm20, %zmm15 #386.34
vpermt2pd %zmm30, %zmm3, %zmm20 #386.34
vpermi2pd %zmm18, %zmm25, %zmm31 #386.64
vpermt2pd %zmm18, %zmm3, %zmm25 #386.64
vmovupd 16(%r15,%rdi,8), %zmm21{%k1}{z} #385.32
vmovupd 80(%r15,%rdi,8), %zmm28{%k1}{z} #385.32
vmovupd 144(%r15,%rdi,8), %zmm16{%k1}{z} #385.32
vmovupd 208(%r15,%rdi,8), %zmm29{%k1}{z} #385.32
vfnmadd231pd %zmm4, %zmm15, %zmm20 #386.54
vaddpd %zmm31, %zmm19, %zmm18 #386.64
vpermt2pd %zmm28, %zmm23, %zmm21 #385.32
vpermt2pd %zmm29, %zmm23, %zmm16 #385.32
vfmsub231pd %zmm1, %zmm18, %zmm21 #387.82
vmovaps %zmm2, %zmm19 #386.64
movslq %r13d, %r13 #387.37
vpermi2pd %zmm27, %zmm17, %zmm19 #386.64
vpermt2pd %zmm27, %zmm3, %zmm17 #386.64
vmovupd 32(%rbx,%r13,8), %zmm18{%k1}{z} #387.37
vmovupd 96(%rbx,%r13,8), %zmm28{%k1}{z} #387.37
vmovupd 224(%rbx,%r13,8), %zmm29{%k1}{z} #387.37
vaddpd %zmm19, %zmm20, %zmm20 #386.64
vpermt2pd %zmm28, %zmm23, %zmm18 #387.37
vmovupd 160(%rbx,%r13,8), %zmm19{%k1}{z} #387.37
vfmsub231pd %zmm1, %zmm20, %zmm16 #387.82
vpermt2pd %zmm29, %zmm23, %zmm19 #387.37
lea (%r8,%rdx,2), %r13d #387.67
addl $16, %edx #383.17
movslq %r13d, %r13 #387.67
vmovupd (%rbx,%r13,8), %zmm29{%k1}{z} #387.67
vmovupd 64(%rbx,%r13,8), %zmm30{%k1}{z} #387.67
vpermt2pd %zmm30, %zmm23, %zmm29 #387.67
vfmsub213pd %zmm29, %zmm4, %zmm22 #387.82
vsubpd %zmm18, %zmm22, %zmm22 #387.82
vmovupd 128(%rbx,%r13,8), %zmm18{%k1}{z} #387.67
vfmsub213pd %zmm21, %zmm5, %zmm22 #387.82
vmovupd 192(%rbx,%r13,8), %zmm21{%k1}{z} #387.67
vfnmadd231pd %zmm22, %zmm0, %zmm25 #389.21
vfmadd231pd %zmm22, %zmm22, %zmm10 #390.21
vpermt2pd %zmm21, %zmm23, %zmm18 #387.67
vfmsub231pd %zmm4, %zmm15, %zmm18 #387.82
vsubpd %zmm19, %zmm18, %zmm18 #387.82
vfmsub213pd %zmm16, %zmm5, %zmm18 #387.82
vpermpd %zmm25, %zmm24, %zmm16 #389.21
vpermpd %zmm25, %zmm26, %zmm25 #389.21
vfnmadd231pd %zmm18, %zmm0, %zmm17 #389.21
vmovupd %zmm16, 16(%rbx,%rdi,8){%k1} #389.21
vmovupd %zmm25, 80(%rbx,%rdi,8){%k1} #389.21
vfmadd231pd %zmm18, %zmm18, %zmm14 #390.21
vpermpd %zmm17, %zmm24, %zmm15 #389.21
vpermpd %zmm17, %zmm26, %zmm17 #389.21
vmovupd %zmm15, 144(%rbx,%rdi,8){%k1} #389.21
vmovupd %zmm17, 208(%rbx,%rdi,8){%k1} #389.21
cmpl %eax, %edx #383.17
jb ..B5.24 # Prob 82% #383.17
# LOE rbx r10 r11 r15 eax edx ecx esi r8d r9d r12d r14d xmm6 xmm7 xmm9 xmm11 xmm13 ymm8 ymm12 zmm0 zmm1 zmm2 zmm3 zmm4 zmm5 zmm10 zmm14 zmm23 zmm24 zmm26 k1
..B5.25: # Preds ..B5.24
# Execution count [9.42e+00]
movq 408(%rsp), %rdi #[spill]
movl 424(%rsp), %r13d #[spill]
movl 400(%rsp), %r9d #[spill]
# LOE rdi r10 eax esi r8d r9d r12d r13d r14d xmm6 xmm7 xmm9 xmm11 xmm13 ymm8 ymm12 zmm0 zmm1 zmm2 zmm3 zmm4 zmm5 zmm10 zmm14
..B5.26: # Preds ..B5.25 ..B5.61
# Execution count [1.05e+01]
lea 1(%rax), %edx #383.17
cmpl %esi, %edx #383.17
ja ..B5.30 # Prob 50% #383.17
# LOE rdi r10 eax esi r8d r9d r12d r13d r14d xmm6 xmm7 xmm9 xmm11 xmm13 ymm8 ymm12 zmm0 zmm1 zmm2 zmm3 zmm4 zmm5 zmm10 zmm14
..B5.27: # Preds ..B5.26
# Execution count [9.42e+00]
subl %eax, %esi #383.17
lea (%r12,%r9,2), %ebx #387.37
addl %r8d, %ebx #387.37
lea (%r9,%r12), %edx #385.32
vmovdqa32 %ymm8, %ymm19 #383.17
lea (%r12,%r8), %ecx #387.67
addl %edx, %r8d #385.32
xorl %r11d, %r11d #383.17
vpbroadcastd %esi, %ymm18 #383.17
lea (%rcx,%rax,2), %ecx #387.67
vmovdqu32 .L_2il0floatpacket.14(%rip), %zmm17 #385.32
vmovdqu32 .L_2il0floatpacket.15(%rip), %zmm20 #385.32
vmovdqu32 .L_2il0floatpacket.16(%rip), %zmm21 #385.32
lea (%rbx,%rax,2), %ebx #387.37
lea (%r8,%rax,2), %edx #385.32
# LOE rdi r10 edx ecx ebx esi r9d r11d r12d r13d r14d xmm6 xmm7 xmm9 xmm11 xmm13 ymm8 ymm12 ymm18 ymm19 zmm0 zmm1 zmm2 zmm3 zmm4 zmm5 zmm10 zmm14 zmm17 zmm20 zmm21
..B5.28: # Preds ..B5.28 ..B5.27
# Execution count [5.23e+01]
vpcmpgtd %ymm19, %ymm18, %k5 #383.17
lea (%rdx,%r11,2), %eax #387.67
vpaddd %ymm12, %ymm19, %ymm19 #383.17
kmovw %k5, %r8d #383.17
vpmovm2q %k5, %zmm22 #386.34
vpbroadcastb %r8d, %zmm23 #385.32
vpmovd2m %zmm22, %k3 #386.34
vptestmb .L_2il0floatpacket.17(%rip), %zmm23, %k1 #385.32
kshiftrw $8, %k3, %k4 #386.34
kshiftrq $8, %k1, %k2 #385.32
movslq %eax, %rax #385.32
lea (%rbx,%r11,2), %r15d #387.37
movslq %r15d, %r15 #387.37
lea (%rcx,%r11,2), %r8d #387.67
movslq %r8d, %r8 #387.67
addl $8, %r11d #383.17
vmovaps %zmm2, %zmm22 #386.34
vmovupd 16(%rdi,%rax,8), %zmm27{%k3}{z} #386.34
vmovupd 80(%rdi,%rax,8), %zmm25{%k4}{z} #386.34
vmovupd 32(%rdi,%r15,8), %zmm23{%k1}{z} #387.37
vmovupd 96(%rdi,%r15,8), %zmm30{%k2}{z} #387.37
vmovupd 8(%rdi,%rax,8), %zmm15{%k3}{z} #386.64
vmovupd 72(%rdi,%rax,8), %zmm26{%k4}{z} #386.64
vmovupd 64(%rdi,%r8,8), %zmm31{%k2}{z} #387.67
vmovupd 16(%r10,%rax,8), %zmm16{%k1}{z} #385.32
vmovupd 80(%r10,%rax,8), %zmm24{%k2}{z} #385.32
vpermi2pd %zmm25, %zmm27, %zmm22 #386.34
vpermt2pd %zmm25, %zmm3, %zmm27 #386.34
vpermt2pd %zmm30, %zmm17, %zmm23 #387.37
vpermt2pd %zmm24, %zmm17, %zmm16 #385.32
vmovupd (%rdi,%r8,8), %zmm30{%k1}{z} #387.67
vfnmadd231pd %zmm22, %zmm4, %zmm27 #386.54
vpermt2pd %zmm31, %zmm17, %zmm30 #387.67
vmovaps %zmm2, %zmm28 #386.64
vpermi2pd %zmm26, %zmm15, %zmm28 #386.64
vpermt2pd %zmm26, %zmm3, %zmm15 #386.64
vfmsub213pd %zmm30, %zmm4, %zmm22 #387.82
vaddpd %zmm28, %zmm27, %zmm29 #386.64
vsubpd %zmm23, %zmm22, %zmm24 #387.82
vfmsub231pd %zmm1, %zmm29, %zmm16 #387.82
vfmsub213pd %zmm16, %zmm5, %zmm24 #387.82
vfnmadd231pd %zmm24, %zmm0, %zmm15 #389.21
vfmadd231pd %zmm24, %zmm24, %zmm10{%k5} #390.21
vpermpd %zmm15, %zmm20, %zmm16 #389.21
vpermpd %zmm15, %zmm21, %zmm15 #389.21
vmovupd %zmm16, 16(%rdi,%rax,8){%k1} #389.21
vmovupd %zmm15, 80(%rdi,%rax,8){%k2} #389.21
cmpl %esi, %r11d #383.17
jb ..B5.28 # Prob 82% #383.17
# LOE rdi r10 edx ecx ebx esi r9d r11d r12d r13d r14d xmm6 xmm7 xmm9 xmm11 xmm13 ymm8 ymm12 ymm18 ymm19 zmm0 zmm1 zmm2 zmm3 zmm4 zmm5 zmm10 zmm14 zmm17 zmm20 zmm21
..B5.30: # Preds ..B5.28 ..B5.26
# Execution count [9.42e+00]
vaddpd %zmm14, %zmm10, %zmm14 #390.21
vshuff32x4 $238, %zmm14, %zmm14, %zmm10 #390.21
vaddpd %zmm14, %zmm10, %zmm15 #390.21
vpermpd $78, %zmm15, %zmm16 #390.21
vaddpd %zmm16, %zmm15, %zmm17 #390.21
vpermpd $177, %zmm17, %zmm18 #390.21
vaddpd %zmm18, %zmm17, %zmm19 #390.21
vaddsd %xmm6, %xmm19, %xmm6 #390.21
jmp ..B5.35 # Prob 100% #390.21
# LOE rdi r10 r9d r12d r13d r14d xmm6 xmm7 xmm9 xmm11 xmm13 ymm8 ymm12 zmm0 zmm1 zmm2 zmm3 zmm4 zmm5
..B5.31: # Preds ..B5.17
# Execution count [0.00e+00]
movl 208(%rsp), %r8d #387.67[spill]
imull %r14d, %r8d #387.67
lea (%r8,%r9,2), %eax #387.37
movl %eax, 432(%rsp) #387.37[spill]
# LOE rdi r10 r11 ebx r8d r9d r12d r13d r14d xmm6 xmm7 xmm9 xmm11 xmm13 ymm8 ymm12 zmm0 zmm1 zmm2 zmm3 zmm4 zmm5
..B5.32: # Preds ..B5.21 ..B5.19 ..B5.31
# Execution count [9.42e+00]
shrl $1, %ebx #354.21
xorl %eax, %eax #383.17
lea (%r9,%r8), %edx #385.32
movl 432(%rsp), %r9d #385.32[spill]
lea (%rdi,%r11,8), %rcx #386.34
lea (%r10,%r11,8), %r11 #385.32
.align 16,0x90
# LOE rcx rdi r10 r11 eax edx ebx r8d r9d r12d r13d r14d xmm6 xmm7 xmm9 xmm11 xmm13 ymm8 ymm12 zmm0 zmm1 zmm2 zmm3 zmm4 zmm5
..B5.33: # Preds ..B5.33 ..B5.32
# Execution count [5.23e+01]
lea (%rdx,%rax,2), %esi #387.57
movslq %esi, %rsi #385.32
lea (%r9,%rax,2), %r15d #387.37
movslq %r15d, %r15 #387.37
vmovsd 16(%rcx,%rsi,8), %xmm19 #387.57
vmulsd %xmm19, %xmm9, %xmm16 #387.57
vmovsd 24(%rcx,%rsi,8), %xmm10 #386.34
vmovsd 32(%rcx,%r15,8), %xmm15 #387.37
lea (%r8,%rax,2), %r15d #387.67
movslq %r15d, %r15 #387.67
incl %eax #383.17
vsubsd %xmm16, %xmm10, %xmm14 #386.54
vsubsd %xmm16, %xmm15, %xmm17 #387.57
vaddsd 8(%rcx,%rsi,8), %xmm14, %xmm20 #386.64
vaddsd (%rcx,%r15,8), %xmm17, %xmm18 #387.67
vfnmadd213sd 16(%r11,%rsi,8), %xmm7, %xmm20 #387.82
vfnmadd231sd %xmm13, %xmm18, %xmm20 #387.82
vfnmadd231sd %xmm20, %xmm11, %xmm19 #389.21
vfmadd231sd %xmm20, %xmm20, %xmm6 #390.21
vmovsd %xmm19, 16(%rcx,%rsi,8) #389.21
cmpl %ebx, %eax #383.17
jb ..B5.33 # Prob 82% #383.17
# LOE rcx rdi r10 r11 eax edx ebx r8d r9d r12d r13d r14d xmm6 xmm7 xmm9 xmm11 xmm13 ymm8 ymm12 zmm0 zmm1 zmm2 zmm3 zmm4 zmm5
..B5.34: # Preds ..B5.33
# Execution count [9.42e+00]
movl 400(%rsp), %r9d #[spill]
# LOE rdi r10 r9d r12d r13d r14d xmm6 xmm7 xmm9 xmm11 xmm13 ymm8 ymm12 zmm0 zmm1 zmm2 zmm3 zmm4 zmm5
..B5.35: # Preds ..B5.34 ..B5.30 ..B5.15
# Execution count [2.09e+01]
negl %r12d #392.27
incl %r14d #382.13
addl $3, %r12d #392.27
cmpl %r13d, %r14d #382.13
jb ..B5.15 # Prob 82% #382.13
# LOE rdi r10 r9d r12d r13d r14d xmm6 xmm7 xmm9 xmm11 xmm13 ymm8 ymm12 zmm0 zmm1 zmm2 zmm3 zmm4 zmm5
..B5.36: # Preds ..B5.35
# Execution count [3.77e+00]
vmovsd %xmm6, 328(%rsp) #[spill]
movl 368(%rsp), %ebx #[spill]
# LOE ebx r13d
..B5.37: # Preds ..B5.36 ..B5.13
# Execution count [4.19e+00]
negl %ebx #394.25
addl $3, %ebx #394.25
cmpl $1, 384(%rsp) #395.28[spill]
jne ..B5.40 # Prob 50% #395.28
# LOE ebx r13d
..B5.38: # Preds ..B5.37
# Execution count [2.09e+00]
movb 392(%rsp), %al #395.41[spill]
testb %al, %al #395.41
jne ..B5.40 # Prob 50% #395.41
# LOE ebx r13d
..B5.39: # Preds ..B5.38
# Execution count [1.05e+00]
vmovsd 40(%rsp), %xmm0 #395.13[spill]
vmovsd %xmm0, 360(%rsp) #395.13[spill]
jmp ..B5.41 # Prob 100% #395.13
# LOE ebx r13d
..B5.40: # Preds ..B5.37 ..B5.38
# Execution count [3.14e+00]
vmovsd .L_2il0floatpacket.9(%rip), %xmm0 #395.13
vmovsd 360(%rsp), %xmm1 #395.13[spill]
vfnmadd132sd 312(%rsp), %xmm0, %xmm1 #395.13[spill]
vdivsd %xmm1, %xmm0, %xmm2 #396.77
vmovsd %xmm2, 360(%rsp) #396.77[spill]
# LOE ebx r13d
..B5.41: # Preds ..B5.40 ..B5.39
# Execution count [4.19e+00]
movb 392(%rsp), %al #378.9[spill]
incb %al #378.9
movb %al, 392(%rsp) #378.9[spill]
cmpb $2, %al #378.9
jb ..B5.5 # Prob 50% #378.9
# LOE ebx r13d
..B5.42: # Preds ..B5.41
# Execution count [2.09e+00]
movl 400(%rsp), %ecx #[spill]
vmovsd 328(%rsp), %xmm0 #[spill]
vmovsd 304(%rsp), %xmm3 #[spill]
vmovsd 296(%rsp), %xmm1 #[spill]
movq 248(%rsp), %rax #[spill]
movq 256(%rsp), %r14 #[spill]
movl 264(%rsp), %r10d #[spill]
movq 272(%rsp), %r9 #[spill]
movq 280(%rsp), %r8 #[spill]
movl 288(%rsp), %edi #[spill]
movl 384(%rsp), %r11d #[spill]
movq 408(%rsp), %rdx #[spill]
movl 240(%rsp), %r12d #[spill]
testl %ecx, %ecx #399.36
jle ..B5.49 # Prob 50% #399.36
# LOE rax rdx r8 r9 r14 ecx edi r10d r11d r12d r13d xmm0 xmm1 xmm3
..B5.43: # Preds ..B5.42
# Execution count [2.09e+00]
movl $1, %esi #399.9
xorl %ebx, %ebx #399.9
testl %r10d, %r10d #399.9
je ..B5.47 # Prob 10% #399.9
# LOE rax rdx r8 r9 r14 ecx ebx esi edi r10d r11d r12d r13d xmm0 xmm1 xmm3
..B5.44: # Preds ..B5.43
# Execution count [1.88e+00]
movl %r11d, 384(%rsp) #[spill]
movl %r12d, 240(%rsp) #[spill]
movl %r13d, 424(%rsp) #[spill]
movl %ecx, 400(%rsp) #[spill]
movl 192(%rsp), %r12d #[spill]
movl 200(%rsp), %r11d #[spill]
.align 16,0x90
# LOE rax rdx r8 r9 r14 ebx edi r10d r11d r12d xmm0 xmm1 xmm3
..B5.45: # Preds ..B5.45 ..B5.44
# Execution count [5.23e+00]
movslq %ebx, %r13 #400.35
lea (%r12,%rbx,2), %ecx #356.19
shlq $4, %r13 #400.35
lea (%r11,%rbx,2), %esi #356.19
movslq %ecx, %rcx #401.35
incl %ebx #399.9
movslq %esi, %rsi #401.13
movq 24(%r9,%r13), %r15 #400.35
movq %r15, 8(%rdx,%r13) #400.13
movq 8(%rdx,%rcx,8), %r15 #401.35
movq %r15, 24(%rdx,%rsi,8) #401.13
movq 32(%r9,%r13), %r15 #400.35
movq %r15, 16(%rdx,%r13) #400.13
movq 16(%rdx,%rcx,8), %r13 #401.35
movq %r13, 32(%rdx,%rsi,8) #401.13
cmpl %r10d, %ebx #399.9
jb ..B5.45 # Prob 64% #399.9
# LOE rax rdx r8 r9 r14 ebx edi r10d r11d r12d xmm0 xmm1 xmm3
..B5.46: # Preds ..B5.45
# Execution count [1.88e+00]
movl 384(%rsp), %r11d #[spill]
lea 1(%rbx,%rbx), %esi #400.13
movl 240(%rsp), %r12d #[spill]
movl 424(%rsp), %r13d #[spill]
movl 400(%rsp), %ecx #[spill]
# LOE rax rdx r8 r9 r14 ecx esi edi r10d r11d r12d r13d xmm0 xmm1 xmm3
..B5.47: # Preds ..B5.46 ..B5.43
# Execution count [2.09e+00]
lea -1(%rsi), %ebx #399.9
cmpl %ecx, %ebx #399.9
jae ..B5.49 # Prob 10% #399.9
# LOE rax rdx r8 r9 r14 ecx esi edi r10d r11d r12d r13d xmm0 xmm1 xmm3
..B5.48: # Preds ..B5.47
# Execution count [1.88e+00]
movslq %esi, %rsi #400.35
movq 16(%r9,%rsi,8), %rbx #400.35
movq %rbx, (%rdx,%rsi,8) #400.13
movq (%rax,%rsi,8), %r15 #401.35
movq %r15, 16(%r14,%rsi,8) #401.13
# LOE rax rdx r8 r9 r14 ecx edi r10d r11d r12d r13d xmm0 xmm1 xmm3
..B5.49: # Preds ..B5.42 ..B5.47 ..B5.48
# Execution count [2.09e+00]
testl %r13d, %r13d #404.41
jle ..B5.56 # Prob 50% #404.41
# LOE rax rdx r8 r9 r14 ecx edi r10d r11d r12d r13d xmm0 xmm1 xmm3
..B5.50: # Preds ..B5.49
# Execution count [2.09e+00]
xorl %ebx, %ebx #404.9
movl $1, %r15d #404.9
xorl %esi, %esi #405.13
testl %edi, %edi #404.9
je ..B5.54 # Prob 10% #404.9
# LOE rax rdx r8 r9 r14 ecx ebx esi edi r10d r11d r12d r13d r15d xmm0 xmm1 xmm3
..B5.51: # Preds ..B5.50
# Execution count [1.88e+00]
movl %r12d, 240(%rsp) #[spill]
movl %r13d, 424(%rsp) #[spill]
movl %r11d, 384(%rsp) #[spill]
movl 48(%rsp), %r13d #[spill]
movl 56(%rsp), %r12d #[spill]
.align 16,0x90
# LOE rax rdx r8 r9 r14 ecx ebx esi edi r10d r12d r13d xmm0 xmm1 xmm3
..B5.52: # Preds ..B5.52 ..B5.51
# Execution count [5.23e+00]
movl %esi, %r15d #405.30
incl %ebx #404.9
subl %ecx, %r15d #405.30
lea (%r15,%rcx,2), %r11d #405.30
movslq %r11d, %r11 #405.30
movq 24(%rdx,%r11,8), %r15 #405.30
movq %r15, 16(%rdx,%r11,8) #405.13
lea (%rsi,%rcx,2), %r11d #406.30
movslq %r11d, %r11 #406.30
movq 16(%rdx,%r11,8), %r15 #406.30
movq %r15, 24(%rdx,%r11,8) #406.13
movq 40(%rdx,%r11,8), %r15 #405.30
movq %r15, 32(%rdx,%r11,8) #405.13
lea (%r12,%rsi), %r11d #387.67
movslq %r11d, %r11 #406.30
addl %r13d, %esi #404.9
movq 32(%rdx,%r11,8), %r15 #406.30
movq %r15, 40(%rdx,%r11,8) #406.13
cmpl %edi, %ebx #404.9
jb ..B5.52 # Prob 64% #404.9
# LOE rax rdx r8 r9 r14 ecx ebx esi edi r10d r12d r13d xmm0 xmm1 xmm3
..B5.53: # Preds ..B5.52
# Execution count [1.88e+00]
movl 384(%rsp), %r11d #[spill]
lea 1(%rbx,%rbx), %r15d #405.13
movl 240(%rsp), %r12d #[spill]
movl 424(%rsp), %r13d #[spill]
# LOE rax rdx r8 r9 r14 ecx edi r10d r11d r12d r13d r15d xmm0 xmm1 xmm3
..B5.54: # Preds ..B5.53 ..B5.50
# Execution count [2.09e+00]
lea -1(%r15), %ebx #404.9
cmpl %r13d, %ebx #404.9
jae ..B5.56 # Prob 10% #404.9
# LOE rax rdx r8 r9 r14 ecx edi r10d r11d r12d r13d r15d xmm0 xmm1 xmm3
..B5.55: # Preds ..B5.54
# Execution count [1.88e+00]
imull 208(%rsp), %r15d #405.30[spill]
addl 216(%rsp), %r15d #405.30[spill]
movslq %r15d, %r15 #405.30
movq 8(%r9,%r15,8), %rbx #405.30
movq %rbx, (%r9,%r15,8) #405.13
movq (%r8,%r15,8), %rsi #406.30
movq %rsi, 8(%r8,%r15,8) #406.13
# LOE rax rdx r8 r9 r14 ecx edi r10d r11d r12d r13d xmm0 xmm1 xmm3
..B5.56: # Preds ..B5.49 ..B5.54 ..B5.55
# Execution count [2.09e+00]
vdivsd %xmm3, %xmm0, %xmm2 #409.37
movl %r11d, %ebx #413.9
vcomisd %xmm1, %xmm2 #374.20
jae ..B5.3 # Prob 82% #374.20
# LOE rax rdx r8 r9 r14 ecx ebx edi r10d r11d r12d r13d xmm1 xmm2 xmm3
..B5.57: # Preds ..B5.3 ..B5.56
# Execution count [3.77e-01]
vmovsd 360(%rsp), %xmm0 #[spill]
# LOE ebx xmm0 xmm2
..B5.58: # Preds ..B5.57 ..B5.1
# Execution count [1.00e+00]
movl $.L_2__STRING.7, %edi #416.5
movl $1, %eax #416.5
vzeroupper #416.5
vmovsd %xmm2, (%rsp) #416.5[spill]
..___tag_value_solveRBA.541:
# printf(const char *__restrict__, ...)
call printf #416.5
..___tag_value_solveRBA.542:
# LOE ebx
..B5.59: # Preds ..B5.58
# Execution count [1.00e+00]
vmovsd (%rsp), %xmm2 #[spill]
movl $.L_2__STRING.8, %edi #417.5
vsqrtsd %xmm2, %xmm2, %xmm2 #417.5
movl %ebx, %esi #417.5
movl $1, %eax #417.5
vmovapd %xmm2, %xmm0 #417.5
..___tag_value_solveRBA.545:
# printf(const char *__restrict__, ...)
call printf #417.5
..___tag_value_solveRBA.546:
# LOE
..B5.60: # Preds ..B5.59
# Execution count [1.00e+00]
addq $472, %rsp #418.1
.cfi_restore 3
popq %rbx #418.1
.cfi_restore 15
popq %r15 #418.1
.cfi_restore 14
popq %r14 #418.1
.cfi_restore 13
popq %r13 #418.1
.cfi_restore 12
popq %r12 #418.1
movq %rbp, %rsp #418.1
popq %rbp #418.1
.cfi_def_cfa 7, 8
.cfi_restore 6
ret #418.1
.cfi_def_cfa 6, 16
.cfi_escape 0x10, 0x03, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xd8, 0xff, 0xff, 0xff, 0x22
.cfi_offset 6, -16
.cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xf8, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xf0, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xe8, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0f, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x22
# LOE
..B5.61: # Preds ..B5.22
# Execution count [9.42e-01]: Infreq
xorl %eax, %eax #383.17
jmp ..B5.26 # Prob 100% #383.17
.align 16,0x90
# LOE rdi r10 eax esi r8d r9d r12d r13d r14d xmm6 xmm7 xmm9 xmm11 xmm13 ymm8 ymm12 zmm0 zmm1 zmm2 zmm3 zmm4 zmm5 zmm10 zmm14
.cfi_endproc
# mark_end;
.type solveRBA,@function
.size solveRBA,.-solveRBA
..LNsolveRBA.4:
.data
# -- End solveRBA
.text
.L_2__routine_start_writeResult_5:
# -- Begin writeResult
.text
# mark_begin;
.align 16,0x90
.globl writeResult
# --- writeResult(Solver *, double *, char *)
writeResult:
# parameter 1: %rdi
# parameter 2: %rsi
# parameter 3: %rdx
..B6.1: # Preds ..B6.0
# Execution count [1.00e+00]
.cfi_startproc
..___tag_value_writeResult.562:
..L563:
#421.1
pushq %r15 #421.1
.cfi_def_cfa_offset 16
.cfi_offset 15, -16
pushq %rbx #421.1
.cfi_def_cfa_offset 24
.cfi_offset 3, -24
pushq %rbp #421.1
.cfi_def_cfa_offset 32
.cfi_offset 6, -32
subq $64, %rsp #421.1
.cfi_def_cfa_offset 96
movq %rsi, %r15 #421.1
movl 24(%rdi), %ebp #422.17
movl $.L_2__STRING.9, %esi #427.10
movl 28(%rdi), %ebx #423.17
movq %rdx, %rdi #427.10
# fopen(const char *__restrict__, const char *__restrict__)
call fopen #427.10
# LOE rax r12 r13 r14 r15 ebx ebp
..B6.19: # Preds ..B6.1
# Execution count [1.00e+00]
movq %rax, %rdi #427.10
# LOE rdi r12 r13 r14 r15 ebx ebp
..B6.2: # Preds ..B6.19
# Execution count [1.00e+00]
testq %rdi, %rdi #429.15
je ..B6.15 # Prob 0% #429.15
# LOE rdi r12 r13 r14 r15 ebx ebp
..B6.3: # Preds ..B6.2
# Execution count [9.93e-01]
addl $2, %ebx #434.32
xorl %ecx, %ecx #434.16
testl %ebx, %ebx #434.32
jle ..B6.13 # Prob 10% #434.32
# LOE rdi r12 r13 r14 r15 ecx ebx ebp
..B6.4: # Preds ..B6.3
# Execution count [8.94e-01]
movslq %ebp, %rsi #435.9
lea 2(%rbp), %edx #435.36
addq $2, %rsi #435.36
movq %rsi, 48(%rsp) #434.16[spill]
movl %ebx, 56(%rsp) #434.16[spill]
movq %r15, %rbx #434.16
movq %r12, (%rsp) #434.16[spill]
.cfi_offset 12, -96
movl %ecx, %r12d #434.16
movq %r13, 8(%rsp) #434.16[spill]
.cfi_offset 13, -88
movl %ecx, %r15d #434.16
movq %r14, 16(%rsp) #434.16[spill]
.cfi_offset 14, -80
movl %edx, %r14d #434.16
movq %rdi, %r13 #434.16
# LOE rbx r13 ebp r12d r14d r15d
..B6.5: # Preds ..B6.11 ..B6.4
# Execution count [4.96e+00]
xorl %edx, %edx #435.20
testl %r14d, %r14d #435.36
jle ..B6.10 # Prob 10% #435.36
# LOE rdx rbx r13 ebp r12d r14d r15d
..B6.6: # Preds ..B6.5
# Execution count [4.47e+00]
movslq %r12d, %r12 #436.32
movl %r15d, 40(%rsp) #436.32[spill]
movl %ebp, 32(%rsp) #436.32[spill]
movq %rdx, %rbp #436.32
movq %rbx, 24(%rsp) #436.32[spill]
movq 48(%rsp), %r15 #436.32[spill]
lea (%rbx,%r12,8), %rax #436.32
movq %rax, %rbx #436.32
# LOE rbx rbp r13 r15 r12d r14d
..B6.7: # Preds ..B6.8 ..B6.6
# Execution count [2.48e+01]
vmovsd (%rbx,%rbp,8), %xmm0 #436.13
movq %r13, %rdi #436.13
movl $.L_2__STRING.11, %esi #436.13
movl $1, %eax #436.13
# fprintf(FILE *__restrict__, const char *__restrict__, ...)
call fprintf #436.13
# LOE rbx rbp r13 r15 r12d r14d
..B6.8: # Preds ..B6.7
# Execution count [2.48e+01]
incq %rbp #435.39
cmpq %r15, %rbp #435.36
jl ..B6.7 # Prob 82% #435.36
# LOE rbx rbp r13 r15 r12d r14d
..B6.9: # Preds ..B6.8
# Execution count [4.47e+00]
movl 40(%rsp), %r15d #[spill]
movl 32(%rsp), %ebp #[spill]
movq 24(%rsp), %rbx #[spill]
# LOE rbx r13 ebp r12d r14d r15d
..B6.10: # Preds ..B6.9 ..B6.5
# Execution count [4.96e+00]
movl $10, %edi #438.9
movq %r13, %rsi #438.9
call fputc #438.9
# LOE rbx r13 ebp r12d r14d r15d
..B6.11: # Preds ..B6.10
# Execution count [4.96e+00]
incl %r15d #434.35
lea 2(%r12,%rbp), %r12d #434.35
cmpl 56(%rsp), %r15d #434.32[spill]
jl ..B6.5 # Prob 82% #434.32
# LOE rbx r13 ebp r12d r14d r15d
..B6.12: # Preds ..B6.11
# Execution count [8.94e-01]
movq (%rsp), %r12 #[spill]
.cfi_restore 12
movq %r13, %rdi #
movq 8(%rsp), %r13 #[spill]
.cfi_restore 13
movq 16(%rsp), %r14 #[spill]
.cfi_restore 14
# LOE rdi r12 r13 r14
..B6.13: # Preds ..B6.12 ..B6.3
# Execution count [9.93e-01]
addq $64, %rsp #441.5
.cfi_def_cfa_offset 32
.cfi_restore 6
popq %rbp #441.5
.cfi_def_cfa_offset 24
.cfi_restore 3
popq %rbx #441.5
.cfi_def_cfa_offset 16
.cfi_restore 15
popq %r15 #441.5
.cfi_def_cfa_offset 8
# fclose(FILE *)
jmp fclose #441.5
.cfi_def_cfa_offset 96
.cfi_offset 3, -24
.cfi_offset 6, -32
.cfi_offset 15, -16
# LOE
..B6.15: # Preds ..B6.2
# Execution count [7.13e-03]: Infreq
movl $il0_peep_printf_format_1, %edi #430.9
call puts #430.9
# LOE
..B6.16: # Preds ..B6.15
# Execution count [7.13e-03]: Infreq
movl $1, %edi #431.9
# exit(int)
call exit #431.9
.align 16,0x90
# LOE
.cfi_endproc
# mark_end;
.type writeResult,@function
.size writeResult,.-writeResult
..LNwriteResult.5:
.section .rodata.str1.4, "aMS",@progbits,1
.space 1, 0x00 # pad
.align 4
il0_peep_printf_format_1:
.long 1869771333
.word 8562
.byte 0
.data
# -- End writeResult
.text
.L_2__routine_start_initSolver_6:
# -- Begin initSolver
.text
# mark_begin;
.align 16,0x90
.globl initSolver
# --- initSolver(Solver *, Parameter *, int)
initSolver:
# parameter 1: %rdi
# parameter 2: %rsi
# parameter 3: %edx
..B7.1: # Preds ..B7.0
# Execution count [1.00e+00]
.cfi_startproc
..___tag_value_initSolver.604:
..L605:
#110.1
pushq %rbp #110.1
.cfi_def_cfa_offset 16
movq %rsp, %rbp #110.1
.cfi_def_cfa 6, 16
.cfi_offset 6, -16
andq $-64, %rsp #110.1
pushq %r12 #110.1
pushq %r13 #110.1
pushq %r14 #110.1
pushq %r15 #110.1
pushq %rbx #110.1
subq $408, %rsp #110.1
.cfi_escape 0x10, 0x03, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xd8, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xf8, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xf0, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xe8, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0f, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x22
movq %rdi, %r15 #110.1
movq %rsi, %rbx #110.1
movl $1140850688, %edi #111.5
movl %edx, 24(%rsp) #110.1[spill]
lea 36(%r15), %rsi #111.5
..___tag_value_initSolver.615:
# MPI_Comm_rank(MPI_Comm, int *)
call MPI_Comm_rank #111.5
..___tag_value_initSolver.616:
# LOE rbx r15
..B7.2: # Preds ..B7.1
# Execution count [1.00e+00]
movl $1140850688, %edi #112.5
lea 40(%r15), %rsi #112.5
..___tag_value_initSolver.617:
# MPI_Comm_size(MPI_Comm, int *)
call MPI_Comm_size #112.5
..___tag_value_initSolver.618:
# LOE rbx r15
..B7.3: # Preds ..B7.2
# Execution count [1.00e+00]
movl 20(%rbx), %eax #114.25
xorl %ecx, %ecx #115.5
movl %eax, 28(%r15) #114.5
movl $.L_2__STRING.5, %edi #116.5
cltd #115.25
idivl 40(%r15) #115.25
movl 36(%r15), %esi #115.36
cmpl %esi, %edx #115.5
movl 16(%rbx), %r8d #113.25
movl %r8d, %edx #116.5
setg %cl #115.5
addl %eax, %ecx #115.25
xorl %eax, %eax #116.5
movl %ecx, 32(%r15) #115.5
movl %r8d, 24(%r15) #113.5
..___tag_value_initSolver.619:
# printf(const char *__restrict__, ...)
call printf #116.5
..___tag_value_initSolver.620:
# LOE rbx r15
..B7.4: # Preds ..B7.3
# Execution count [1.00e+00]
vxorpd %xmm4, %xmm4, %xmm4 #122.41
vxorpd %xmm1, %xmm1, %xmm1 #121.41
vcvtsi2sd 20(%rbx), %xmm4, %xmm4 #122.41
vcvtsi2sd 16(%rbx), %xmm1, %xmm1 #121.41
vmovsd 8(%rbx), %xmm3 #122.23
vxorpd %xmm6, %xmm6, %xmm6 #123.38
vdivsd %xmm4, %xmm3, %xmm5 #122.41
vmovsd (%rbx), %xmm0 #121.23
movl $64, %edi #131.21
vdivsd %xmm1, %xmm0, %xmm2 #121.41
movl 32(%r15), %r12d #123.38
movl 36(%r15), %r14d #123.23
imull %r12d, %r14d #123.38
vmovsd %xmm2, (%r15) #121.5
vmovsd %xmm5, 8(%r15) #122.5
movl 24(%r15), %edx #128.21
vcvtsi2sd %r14d, %xmm6, %xmm6 #123.38
movq 32(%rbx), %r13 #124.23
movq 40(%rbx), %rax #125.23
movl 24(%rbx), %ebx #126.23
movl %ebx, 88(%r15) #126.5
lea 2(%rdx), %ebx #131.42
movq %r13, 64(%r15) #124.5
lea 2(%r12), %r13d #131.60
movl %ebx, %esi #131.21
imull %r13d, %esi #131.21
vmulsd %xmm6, %xmm5, %xmm7 #123.58
movslq %esi, %rsi #131.21
shlq $3, %rsi #131.21
vmovsd %xmm7, 16(%r15) #123.5
movq %rax, 72(%r15) #125.5
movl %edx, 40(%rsp) #128.21[spill]
movl 28(%r15), %r14d #129.21
..___tag_value_initSolver.622:
# allocate(int, size_t)
call allocate #131.21
..___tag_value_initSolver.623:
# LOE rax r15 ebx r12d r13d r14d
..B7.5: # Preds ..B7.4
# Execution count [1.00e+00]
movq %rax, 48(%r15) #131.5
lea 2(%r14), %eax #132.55
movl %ebx, %esi #132.21
movl $64, %edi #132.21
imull %eax, %esi #132.21
movslq %esi, %rsi #132.21
shlq $3, %rsi #132.21
movl %eax, 32(%rsp) #132.55[spill]
..___tag_value_initSolver.624:
# allocate(int, size_t)
call allocate #132.21
..___tag_value_initSolver.625:
# LOE rax r15 ebx r12d r13d r14d
..B7.6: # Preds ..B7.5
# Execution count [1.00e+00]
movq %rax, 56(%r15) #132.5
xorl %r9d, %r9d #139.5
vmovsd (%r15), %xmm6 #134.19
vmovsd 8(%r15), %xmm5 #135.19
testl %r13d, %r13d #139.37
movq 48(%r15), %r13 #136.19
jle ..B7.27 # Prob 9% #139.37
# LOE rax r9 r13 r15 ebx r12d r14d xmm5 xmm6
..B7.7: # Preds ..B7.6
# Execution count [9.00e-01]
vmovsd .L_2il0floatpacket.18(%rip), %xmm1 #142.27
movl %ebx, %r8d #141.9
vmulsd %xmm1, %xmm6, %xmm0 #142.38
andl $-8, %r8d #141.9
vmovdqu .L_2il0floatpacket.7(%rip), %ymm3 #141.9
vmovdqu .L_2il0floatpacket.6(%rip), %ymm2 #142.23
vbroadcastsd %xmm0, %zmm0 #142.38
vmovsd %xmm5, 304(%rsp) #139.5[spill]
vmovsd %xmm6, 16(%rsp) #139.5[spill]
movslq 40(%rsp), %rdi #142.13[spill]
movq %r13, %rcx #139.5
movslq %r12d, %rdx #139.5
xorl %r12d, %r12d #139.5
addq $2, %rdx #139.37
vmovups %zmm0, 128(%rsp) #139.5[spill]
lea 16(,%rdi,8), %rsi #142.13
movq %rcx, 312(%rsp) #139.5[spill]
movq %rdx, 264(%rsp) #139.5[spill]
movq %rsi, 272(%rsp) #139.5[spill]
movq %rdi, 280(%rsp) #139.5[spill]
movl %r8d, (%rsp) #139.5[spill]
movq %r9, 288(%rsp) #139.5[spill]
movq %rax, 48(%rsp) #139.5[spill]
movl %r14d, 8(%rsp) #139.5[spill]
movq %r15, 296(%rsp) #139.5[spill]
vpxord %zmm4, %zmm4, %zmm4 #142.23
# LOE r12 r13 ebx
..B7.8: # Preds ..B7.25 ..B7.7
# Execution count [5.00e+00]
vxorpd %xmm1, %xmm1, %xmm1 #140.33
movq 288(%rsp), %rax #140.37[spill]
vcvtsi2sd %eax, %xmm1, %xmm1 #140.33
movq 296(%rsp), %rcx #140.37[spill]
vmovsd 304(%rsp), %xmm0 #140.37[spill]
vfmadd213sd 16(%rcx), %xmm0, %xmm1 #140.37
vmovsd %xmm1, 352(%rsp) #140.37[spill]
testl %ebx, %ebx #141.36
jle ..B7.25 # Prob 50% #141.36
# LOE r12 r13 ebx
..B7.9: # Preds ..B7.8
# Execution count [4.50e+00]
cmpl $8, %ebx #141.9
jl ..B7.77 # Prob 10% #141.9
# LOE r12 r13 ebx
..B7.10: # Preds ..B7.9
# Execution count [4.50e+00]
cmpl $1001, %ebx #141.9
jl ..B7.76 # Prob 10% #141.9
# LOE r12 r13 ebx
..B7.11: # Preds ..B7.10
# Execution count [4.50e+00]
movq 312(%rsp), %r15 #141.9[spill]
andq $63, %r15 #141.9
testl $7, %r15d #141.9
je ..B7.13 # Prob 50% #141.9
# LOE r12 r13 ebx r15d
..B7.12: # Preds ..B7.11
# Execution count [2.25e+00]
xorl %r15d, %r15d #141.9
jmp ..B7.15 # Prob 100% #141.9
# LOE r12 r13 ebx r15d
..B7.13: # Preds ..B7.11
# Execution count [2.25e+00]
testl %r15d, %r15d #141.9
je ..B7.15 # Prob 50% #141.9
# LOE r12 r13 ebx r15d
..B7.14: # Preds ..B7.13
# Execution count [2.50e+01]
negl %r15d #141.9
addl $64, %r15d #141.9
shrl $3, %r15d #141.9
cmpl %r15d, %ebx #141.9
cmovl %ebx, %r15d #141.9
# LOE r12 r13 ebx r15d
..B7.15: # Preds ..B7.12 ..B7.14 ..B7.13
# Execution count [5.00e+00]
movl %ebx, %r14d #141.9
subl %r15d, %r14d #141.9
andl $7, %r14d #141.9
negl %r14d #141.9
addl %ebx, %r14d #141.9
cmpl $1, %r15d #141.9
jb ..B7.19 # Prob 50% #141.9
# LOE r12 r13 ebx r14d r15d
..B7.16: # Preds ..B7.15
# Execution count [4.50e+00]
vmovdqu .L_2il0floatpacket.7(%rip), %ymm0 #141.9
vmovsd .L_2il0floatpacket.18(%rip), %xmm2 #142.48
vpbroadcastd %r15d, %ymm1 #141.9
vmovdqu %ymm0, 96(%rsp) #141.9[spill]
vmulsd 352(%rsp), %xmm2, %xmm0 #142.48[spill]
vmovdqu %ymm1, 64(%rsp) #141.9[spill]
movq $0, 256(%rsp) #141.9[spill]
vzeroupper #142.48
call sin #142.48
# LOE r12 r13 ebx r14d r15d xmm0
..B7.82: # Preds ..B7.16
# Execution count [4.50e+00]
movl %ebx, 56(%rsp) #141.9[spill]
vbroadcastsd %xmm0, %zmm16 #142.48
vmovdqu32 64(%rsp), %ymm18 #141.9[spill]
vmovdqu32 96(%rsp), %ymm17 #141.9[spill]
vmovups 128(%rsp), %zmm19 #141.9[spill]
vmovdqu32 .L_2il0floatpacket.6(%rip), %ymm20 #141.9
movslq %r15d, %rsi #141.9
movq 256(%rsp), %rbx #141.9[spill]
movq 312(%rsp), %rdi #141.9[spill]
vpxord %zmm21, %zmm21, %zmm21 #141.9
# LOE rbx rsi rdi r12 r13 r14d r15d ymm17 ymm18 ymm20 zmm16 zmm19 zmm21
..B7.17: # Preds ..B7.83 ..B7.82
# Execution count [2.50e+01]
vpcmpgtd %ymm17, %ymm18, %k4 #141.9
vmovaps %zmm21, %zmm0 #142.23
kmovw %k4, %k1 #142.23
vcvtdq2pd %ymm17, %zmm2 #142.38
vmulpd %zmm2, %zmm19, %zmm1 #142.42
..___tag_value_initSolver.655:
call *__svml_sin8_mask_z0@GOTPCREL(%rip) #142.23
..___tag_value_initSolver.656:
# LOE rbx rsi rdi r12 r13 r14d r15d ymm17 ymm18 ymm20 zmm0 zmm16 zmm19 zmm21 k4
..B7.83: # Preds ..B7.17
# Execution count [2.50e+01]
vpaddd %ymm20, %ymm17, %ymm17 #141.9
vaddpd %zmm0, %zmm16, %zmm1 #142.48
vmovupd %zmm1, (%rdi,%rbx,8){%k4} #142.13
addq $8, %rbx #141.9
cmpq %rsi, %rbx #141.9
jb ..B7.17 # Prob 82% #141.9
# LOE rbx rsi rdi r12 r13 r14d r15d ymm17 ymm18 ymm20 zmm16 zmm19 zmm21
..B7.18: # Preds ..B7.83
# Execution count [4.50e+00]
movl 56(%rsp), %ebx #[spill]
cmpl %r15d, %ebx #141.9
je ..B7.25 # Prob 10% #141.9
# LOE r12 r13 ebx r14d r15d
..B7.19: # Preds ..B7.18 ..B7.15 ..B7.76
# Execution count [2.50e+01]
vpbroadcastd %r15d, %ymm0 #142.23
lea 8(%r15), %eax #141.9
vpaddd .L_2il0floatpacket.20(%rip), %ymm0, %ymm1 #142.23
vmovdqu %ymm1, 320(%rsp) #142.23[spill]
cmpl %eax, %r14d #141.9
jl ..B7.23 # Prob 50% #141.9
# LOE r12 r13 ebx r14d r15d
..B7.20: # Preds ..B7.19
# Execution count [4.50e+00]
vmovsd .L_2il0floatpacket.18(%rip), %xmm0 #142.48
vmulsd 352(%rsp), %xmm0, %xmm0 #142.48[spill]
movslq %r15d, %r15 #141.9
vzeroupper #142.48
call sin #142.48
# LOE r12 r13 r15 ebx r14d r15d xmm0
..B7.84: # Preds ..B7.20
# Execution count [4.50e+00]
vbroadcastsd %xmm0, %zmm16 #142.48
vmovdqu32 320(%rsp), %ymm17 #142.48[spill]
vmovups 128(%rsp), %zmm18 #142.48[spill]
vmovdqu32 .L_2il0floatpacket.6(%rip), %ymm19 #142.48
movq %r15, %rsi #142.48
movq 312(%rsp), %rdi #142.48[spill]
# LOE rsi rdi r12 r13 ebx r14d r15d ymm17 ymm19 zmm16 zmm18
..B7.21: # Preds ..B7.85 ..B7.84
# Execution count [2.50e+01]
vcvtdq2pd %ymm17, %zmm1 #142.38
vmulpd %zmm1, %zmm18, %zmm0 #142.42
..___tag_value_initSolver.663:
call *__svml_sin8_z0@GOTPCREL(%rip) #142.23
..___tag_value_initSolver.664:
# LOE rsi rdi r12 r13 ebx r14d r15d ymm17 ymm19 zmm0 zmm16 zmm18
..B7.85: # Preds ..B7.21
# Execution count [2.50e+01]
addl $8, %r15d #141.9
vpaddd %ymm19, %ymm17, %ymm17 #142.23
vaddpd %zmm0, %zmm16, %zmm1 #142.48
vmovupd %zmm1, (%rdi,%rsi,8) #142.13
addq $8, %rsi #141.9
cmpl %r14d, %r15d #141.9
jb ..B7.21 # Prob 82% #141.9
# LOE rsi rdi r12 r13 ebx r14d r15d ymm17 ymm19 zmm16 zmm18
..B7.23: # Preds ..B7.85 ..B7.19 ..B7.77
# Execution count [5.00e+00]
lea 1(%r14), %eax #141.9
cmpl %ebx, %eax #141.9
ja ..B7.25 # Prob 50% #141.9
# LOE r12 r13 ebx r14d
..B7.24: # Preds ..B7.23
# Execution count [2.50e+01]
vpbroadcastd %r14d, %ymm1 #142.23
movl %ebx, %eax #141.9
vpaddd .L_2il0floatpacket.20(%rip), %ymm1, %ymm2 #142.23
subl %r14d, %eax #141.9
vpbroadcastd %eax, %ymm4 #141.9
vpcmpgtd .L_2il0floatpacket.7(%rip), %ymm4, %k1 #141.9
kmovw %k1, 360(%rsp) #141.9[spill]
vcvtdq2pd %ymm2, %zmm3 #142.38
vmulpd 128(%rsp), %zmm3, %zmm1 #142.42[spill]
vpxord %zmm0, %zmm0, %zmm0 #142.23
..___tag_value_initSolver.667:
call *__svml_sin8_mask_z0@GOTPCREL(%rip) #142.23
..___tag_value_initSolver.668:
# LOE r12 r13 ebx r14d zmm0
..B7.87: # Preds ..B7.24
# Execution count [2.50e+01]
vmovups %zmm0, 192(%rsp) #142.23[spill]
vmovsd .L_2il0floatpacket.18(%rip), %xmm0 #142.48
vmulsd 352(%rsp), %xmm0, %xmm0 #142.48[spill]
vzeroupper #142.48
call sin #142.48
# LOE r12 r13 ebx r14d xmm0
..B7.86: # Preds ..B7.87
# Execution count [2.50e+01]
vbroadcastsd %xmm0, %zmm1 #142.48
kmovw 360(%rsp), %k1 #142.13[spill]
vaddpd 192(%rsp), %zmm1, %zmm2 #142.48[spill]
movslq %r14d, %r14 #142.13
addq %r12, %r14 #142.13
vmovupd %zmm2, (%r13,%r14,8){%k1} #142.13
# LOE r12 r13 ebx
..B7.25: # Preds ..B7.18 ..B7.23 ..B7.8 ..B7.86
# Execution count [5.00e+00]
movq 280(%rsp), %rax #139.5[spill]
movq 312(%rsp), %rcx #139.5[spill]
movq 288(%rsp), %rsi #139.5[spill]
incq %rsi #139.5
lea 2(%r12,%rax), %r12 #139.5
addq 272(%rsp), %rcx #139.5[spill]
movq %rcx, 312(%rsp) #139.5[spill]
movq %rsi, 288(%rsp) #139.5[spill]
cmpq 264(%rsp), %rsi #139.5[spill]
jb ..B7.8 # Prob 82% #139.5
# LOE r12 r13 ebx
..B7.26: # Preds ..B7.25
# Execution count [9.00e-01]
vmovsd 16(%rsp), %xmm6 #[spill]
movq 48(%rsp), %rax #[spill]
movl 8(%rsp), %r14d #[spill]
# LOE rax ebx r14d xmm6
..B7.27: # Preds ..B7.26 ..B7.6
# Execution count [1.00e+00]
cmpl $2, 24(%rsp) #146.20[spill]
je ..B7.48 # Prob 16% #146.20
# LOE rax ebx r14d xmm6
..B7.28: # Preds ..B7.27
# Execution count [8.40e-01]
xorl %edx, %edx #153.9
cmpl $0, 32(%rsp) #153.36[spill]
jle ..B7.45 # Prob 9% #153.36
# LOE rax rdx ebx r14d
..B7.29: # Preds ..B7.28
# Execution count [7.56e-01]
movslq 40(%rsp), %rcx #155.17[spill]
movl %ebx, %esi #154.13
movslq %r14d, %r14 #153.9
andl $-8, %esi #154.13
movq %rdx, %rdi #153.9
addq $2, %r14 #153.36
vmovdqu .L_2il0floatpacket.7(%rip), %ymm0 #154.13
lea 16(,%rcx,8), %r12 #155.17
movslq %esi, %rcx #154.13
xorl %r13d, %r13d #153.9
movq %rax, %r15 #153.9
movl %ebx, %edx #153.9
movq %rdi, %rbx #153.9
vpxord %zmm1, %zmm1, %zmm1 #155.29
# LOE rax rcx rbx r12 r13 r14 r15 edx esi ymm0 zmm1
..B7.30: # Preds ..B7.43 ..B7.29
# Execution count [7.56e-01]
testl %edx, %edx #154.40
jle ..B7.43 # Prob 50% #154.40
# LOE rax rcx rbx r12 r13 r14 r15 edx esi ymm0 zmm1
..B7.31: # Preds ..B7.30
# Execution count [0.00e+00]
cmpl $12, %edx #154.13
jle ..B7.37 # Prob 0% #154.13
# LOE rax rcx rbx r12 r13 r14 r15 edx esi ymm0 zmm1
..B7.32: # Preds ..B7.31
# Execution count [2.48e+00]
movq %rcx, (%rsp) #[spill]
movl %esi, 8(%rsp) #[spill]
movq %rax, 48(%rsp) #[spill]
movl %edx, 56(%rsp) #[spill]
# LOE rbx r12 r13 r14 r15
..B7.33: # Preds ..B7.35 ..B7.32
# Execution count [4.20e+00]
movq %r15, %rdi #154.13
xorl %esi, %esi #154.13
movq %r12, %rdx #154.13
vzeroupper #154.13
call __intel_skx_avx512_memset #154.13
# LOE rbx r12 r13 r14 r15
..B7.34: # Preds ..B7.33
# Execution count [4.20e+00]
incq %rbx #153.9
addq %r12, %r15 #153.9
addq %r12, %r13 #153.9
cmpq %r14, %rbx #153.9
jae ..B7.45 # Prob 18% #153.9
# LOE rbx r12 r13 r14 r15
..B7.35: # Preds ..B7.34
# Execution count [1.72e+00]
cmpl $12, 56(%rsp) #154.13[spill]
jg ..B7.33 # Prob 100% #154.13
# LOE rbx r12 r13 r14 r15
..B7.36: # Preds ..B7.35
# Execution count [0.00e+00]
vmovdqu .L_2il0floatpacket.7(%rip), %ymm0 #
movq (%rsp), %rcx #[spill]
movl 8(%rsp), %esi #[spill]
movq 48(%rsp), %rax #[spill]
movl 56(%rsp), %edx #[spill]
vpxord %zmm1, %zmm1, %zmm1 #
# LOE rax rcx rbx r12 r13 r14 r15 edx esi ymm0 zmm1
..B7.37: # Preds ..B7.36 ..B7.31
# Execution count [4.20e+00]
cmpl $8, %edx #154.13
jl ..B7.46 # Prob 10% #154.13
# LOE rax rcx rbx r12 r13 r14 r15 edx esi ymm0 zmm1
..B7.38: # Preds ..B7.37
# Execution count [4.20e+00]
movl %esi, %r8d #154.13
xorl %r9d, %r9d #154.13
# LOE rax rcx rbx r9 r12 r13 r14 r15 edx esi r8d ymm0 zmm1
..B7.39: # Preds ..B7.39 ..B7.38
# Execution count [2.33e+01]
vmovupd %zmm1, (%r15,%r9,8) #155.17
addq $8, %r9 #154.13
cmpq %rcx, %r9 #154.13
jb ..B7.39 # Prob 81% #154.13
# LOE rax rcx rbx r9 r12 r13 r14 r15 edx esi r8d ymm0 zmm1
..B7.41: # Preds ..B7.39 ..B7.46
# Execution count [4.67e+00]
lea 1(%r8), %r9d #154.13
cmpl %edx, %r9d #154.13
ja ..B7.43 # Prob 50% #154.13
# LOE rax rcx rbx r12 r13 r14 r15 edx esi r8d ymm0 zmm1
..B7.42: # Preds ..B7.41
# Execution count [2.33e+01]
movl %edx, %r9d #154.13
subl %r8d, %r9d #154.13
vpbroadcastd %r9d, %ymm2 #154.13
vpcmpgtd %ymm0, %ymm2, %k1 #154.13
movslq %r8d, %r8 #155.17
lea (%rax,%r8,8), %r10 #132.21
vmovupd %zmm1, (%r10,%r13){%k1} #155.17
# LOE rax rcx rbx r12 r13 r14 r15 edx esi ymm0 zmm1
..B7.43: # Preds ..B7.30 ..B7.41 ..B7.42
# Execution count [0.00e+00]
incq %rbx #153.9
addq %r12, %r15 #153.9
addq %r12, %r13 #153.9
cmpq %r14, %rbx #153.9
jb ..B7.30 # Prob 82% #153.9
# LOE rax rcx rbx r12 r13 r14 r15 edx esi ymm0 zmm1
..B7.45: # Preds ..B7.74 ..B7.70 ..B7.34 ..B7.43 ..B7.28
# ..B7.48
# Execution count [1.00e+00]
vzeroupper #159.1
addq $408, %rsp #159.1
.cfi_restore 3
popq %rbx #159.1
.cfi_restore 15
popq %r15 #159.1
.cfi_restore 14
popq %r14 #159.1
.cfi_restore 13
popq %r13 #159.1
.cfi_restore 12
popq %r12 #159.1
movq %rbp, %rsp #159.1
popq %rbp #159.1
.cfi_def_cfa 7, 8
.cfi_restore 6
ret #159.1
.cfi_def_cfa 6, 16
.cfi_escape 0x10, 0x03, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xd8, 0xff, 0xff, 0xff, 0x22
.cfi_offset 6, -16
.cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xf8, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xf0, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xe8, 0xff, 0xff, 0xff, 0x22
.cfi_escape 0x10, 0x0f, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x22
# LOE
..B7.46: # Preds ..B7.37
# Execution count [4.20e-01]: Infreq
xorl %r8d, %r8d #154.13
jmp ..B7.41 # Prob 100% #154.13
# LOE rax rcx rbx r12 r13 r14 r15 edx esi r8d ymm0 zmm1
..B7.48: # Preds ..B7.27
# Execution count [1.60e-01]: Infreq
xorl %esi, %esi #147.9
cmpl $0, 32(%rsp) #147.36[spill]
jle ..B7.45 # Prob 10% #147.36
# LOE rax rsi ebx r14d xmm6
..B7.49: # Preds ..B7.48
# Execution count [1.44e-01]: Infreq
vmulsd .L_2il0floatpacket.19(%rip), %xmm6, %xmm1 #149.44
movl %ebx, %r8d #148.13
vmovdqu32 .L_2il0floatpacket.7(%rip), %ymm17 #148.13
vmovdqu32 .L_2il0floatpacket.6(%rip), %ymm19 #149.29
vbroadcastsd %xmm1, %zmm16 #149.44
movslq 40(%rsp), %r13 #149.17[spill]
andl $-8, %r8d #148.13
movslq %r14d, %r14 #147.9
xorl %r15d, %r15d #147.9
addq $2, %r14 #147.36
movq %rax, %r12 #147.9
movl %r8d, (%rsp) #147.9[spill]
lea 16(,%r13,8), %rdi #149.17
movq %rdi, 24(%rsp) #147.9[spill]
movq %r14, 32(%rsp) #147.9[spill]
movq %rax, %r14 #147.9
vpxord %zmm18, %zmm18, %zmm18 #149.29
# LOE rsi r12 r13 r14 r15 ebx ymm17 ymm19 zmm16 zmm18
..B7.50: # Preds ..B7.70 ..B7.49
# Execution count [7.41e-01]: Infreq
testl %ebx, %ebx #148.40
jle ..B7.70 # Prob 50% #148.40
# LOE rsi r12 r13 r14 r15 ebx ymm17 ymm19 zmm16 zmm18
..B7.51: # Preds ..B7.50
# Execution count [6.90e-01]: Infreq
cmpl $8, %ebx #148.13
jl ..B7.73 # Prob 10% #148.13
# LOE rsi r12 r13 r14 r15 ebx ymm17 ymm19 zmm16 zmm18
..B7.52: # Preds ..B7.51
# Execution count [6.61e-01]: Infreq
movq %r14, 48(%rsp) #[spill]
movl %ebx, %edx #
movq 32(%rsp), %r8 #[spill]
movq %rsi, %rax #
movq 24(%rsp), %rsi #[spill]
movq %r13, %rcx #
# LOE rax rcx rsi r8 r12 r15 edx ymm17 ymm19 zmm16 zmm18
..B7.53: # Preds ..B7.74 ..B7.52
# Execution count [7.20e-01]: Infreq
cmpl $430, %edx #148.13
jl ..B7.72 # Prob 10% #148.13
# LOE rax rcx rsi r8 r12 r15 edx ymm17 ymm19 zmm16 zmm18
..B7.54: # Preds ..B7.53
# Execution count [7.20e-01]: Infreq
movq %r12, %r13 #148.13
andq $63, %r13 #148.13
testl $7, %r13d #148.13
je ..B7.56 # Prob 50% #148.13
# LOE rax rcx rsi r8 r12 r15 edx r13d ymm17 ymm19 zmm16 zmm18
..B7.55: # Preds ..B7.54
# Execution count [3.60e-01]: Infreq
xorl %r13d, %r13d #148.13
jmp ..B7.58 # Prob 100% #148.13
# LOE rax rcx rsi r8 r12 r15 edx r13d ymm17 ymm19 zmm16 zmm18
..B7.56: # Preds ..B7.54
# Execution count [3.60e-01]: Infreq
testl %r13d, %r13d #148.13
je ..B7.58 # Prob 50% #148.13
# LOE rax rcx rsi r8 r12 r15 edx r13d ymm17 ymm19 zmm16 zmm18
..B7.57: # Preds ..B7.56
# Execution count [4.00e+00]: Infreq
negl %r13d #148.13
addl $64, %r13d #148.13
shrl $3, %r13d #148.13
cmpl %r13d, %edx #148.13
cmovl %edx, %r13d #148.13
# LOE rax rcx rsi r8 r12 r15 edx r13d ymm17 ymm19 zmm16 zmm18
..B7.58: # Preds ..B7.55 ..B7.57 ..B7.56
# Execution count [8.00e-01]: Infreq
movl %edx, %edi #148.13
subl %r13d, %edi #148.13
andl $7, %edi #148.13
negl %edi #148.13
addl %edx, %edi #148.13
cmpl $1, %r13d #148.13
jb ..B7.63 # Prob 50% #148.13
# LOE rax rcx rsi r8 r12 r15 edx edi r13d ymm17 ymm19 zmm16 zmm18
..B7.60: # Preds ..B7.58
# Execution count [7.20e-01]: Infreq
vmovdqa32 %ymm17, %ymm21 #148.13
xorl %ebx, %ebx #148.13
vpbroadcastd %r13d, %ymm20 #148.13
movslq %r13d, %r14 #148.13
movq %rcx, 16(%rsp) #148.13[spill]
movq %r8, 32(%rsp) #148.13[spill]
movq %rax, 8(%rsp) #148.13[spill]
movl %edx, 56(%rsp) #148.13[spill]
# LOE rbx rsi r12 r14 r15 edi r13d ymm17 ymm19 ymm20 ymm21 zmm16 zmm18
..B7.61: # Preds ..B7.88 ..B7.60
# Execution count [4.00e+00]: Infreq
vpcmpgtd %ymm21, %ymm20, %k4 #148.13
vmovaps %zmm18, %zmm0 #149.29
kmovw %k4, %k1 #149.29
vcvtdq2pd %ymm21, %zmm2 #149.44
vmulpd %zmm2, %zmm16, %zmm1 #149.48
..___tag_value_initSolver.721:
call *__svml_sin8_mask_z0@GOTPCREL(%rip) #149.29
..___tag_value_initSolver.722:
# LOE rbx rsi r12 r14 r15 edi r13d ymm17 ymm19 ymm20 ymm21 zmm0 zmm16 zmm18 k4
..B7.88: # Preds ..B7.61
# Execution count [4.00e+00]: Infreq
vpaddd %ymm19, %ymm21, %ymm21 #148.13
vmovupd %zmm0, (%r12,%rbx,8){%k4} #149.17
addq $8, %rbx #148.13
cmpq %r14, %rbx #148.13
jb ..B7.61 # Prob 82% #148.13
# LOE rbx rsi r12 r14 r15 edi r13d ymm17 ymm19 ymm20 ymm21 zmm16 zmm18
..B7.62: # Preds ..B7.88
# Execution count [7.20e-01]: Infreq
movl 56(%rsp), %edx #[spill]
movq 16(%rsp), %rcx #[spill]
movq 32(%rsp), %r8 #[spill]
movq 8(%rsp), %rax #[spill]
cmpl %r13d, %edx #148.13
je ..B7.74 # Prob 10% #148.13
# LOE rax rcx rsi r8 r12 r15 edx edi r13d ymm17 ymm19 zmm16 zmm18
..B7.63: # Preds ..B7.58 ..B7.62
# Execution count [6.48e-01]: Infreq
movq %rsi, 24(%rsp) #[spill]
movq %rax, %rsi #
movq %r8, 32(%rsp) #[spill]
movl %r13d, %eax #
movq 48(%rsp), %r14 #[spill]
movq %rcx, %r13 #
movl %edx, %ebx #
# LOE rsi r12 r13 r14 r15 eax ebx edi ymm17 ymm19 zmm16 zmm18
..B7.64: # Preds ..B7.63 ..B7.72
# Execution count [4.00e+00]: Infreq
vpbroadcastd %eax, %ymm1 #149.29
lea 8(%rax), %r8d #148.13
vpaddd .L_2il0floatpacket.20(%rip), %ymm1, %ymm20 #149.29
cmpl %r8d, %edi #148.13
jl ..B7.68 # Prob 50% #148.13
# LOE rsi r12 r13 r14 r15 eax ebx edi ymm17 ymm19 ymm20 zmm16 zmm18
..B7.65: # Preds ..B7.64
# Execution count [7.20e-01]: Infreq
movslq %eax, %rdx #148.13
movq %r14, 48(%rsp) #148.13[spill]
movl %eax, %r14d #148.13
movl %ebx, 56(%rsp) #148.13[spill]
movq %rdx, %rbx #148.13
# LOE rbx rsi r12 r13 r15 edi r14d ymm17 ymm19 ymm20 zmm16 zmm18
..B7.66: # Preds ..B7.89 ..B7.65
# Execution count [4.00e+00]: Infreq
vcvtdq2pd %ymm20, %zmm1 #149.44
vmulpd %zmm1, %zmm16, %zmm0 #149.48
..___tag_value_initSolver.732:
call *__svml_sin8_z0@GOTPCREL(%rip) #149.29
..___tag_value_initSolver.733:
# LOE rbx rsi r12 r13 r15 edi r14d ymm17 ymm19 ymm20 zmm0 zmm16 zmm18
..B7.89: # Preds ..B7.66
# Execution count [4.00e+00]: Infreq
addl $8, %r14d #148.13
vpaddd %ymm19, %ymm20, %ymm20 #149.29
vmovupd %zmm0, (%r12,%rbx,8) #149.17
addq $8, %rbx #148.13
cmpl %edi, %r14d #148.13
jb ..B7.66 # Prob 82% #148.13
# LOE rbx rsi r12 r13 r15 edi r14d ymm17 ymm19 ymm20 zmm16 zmm18
..B7.67: # Preds ..B7.89
# Execution count [7.20e-01]: Infreq
movq 48(%rsp), %r14 #[spill]
movl 56(%rsp), %ebx #[spill]
# LOE rsi r12 r13 r14 r15 ebx edi ymm17 ymm19 zmm16 zmm18
..B7.68: # Preds ..B7.67 ..B7.64 ..B7.73
# Execution count [8.00e-01]: Infreq
lea 1(%rdi), %r8d #148.13
cmpl %ebx, %r8d #148.13
ja ..B7.70 # Prob 50% #148.13
# LOE rsi r12 r13 r14 r15 ebx edi ymm17 ymm19 zmm16 zmm18
..B7.69: # Preds ..B7.68
# Execution count [4.00e+00]: Infreq
vpbroadcastd %edi, %ymm2 #149.29
movl %ebx, %r8d #148.13
vpaddd .L_2il0floatpacket.20(%rip), %ymm2, %ymm3 #149.29
subl %edi, %r8d #148.13
vpbroadcastd %r8d, %ymm5 #148.13
vpcmpgtd %ymm17, %ymm5, %k4 #148.13
vmovaps %zmm18, %zmm0 #149.29
kmovw %k4, %k1 #149.29
vcvtdq2pd %ymm3, %zmm4 #149.44
vmulpd %zmm4, %zmm16, %zmm1 #149.48
..___tag_value_initSolver.736:
call *__svml_sin8_mask_z0@GOTPCREL(%rip) #149.29
..___tag_value_initSolver.737:
# LOE rsi r12 r13 r14 r15 ebx edi ymm17 ymm19 zmm0 zmm16 zmm18 k4
..B7.90: # Preds ..B7.69
# Execution count [4.00e+00]: Infreq
movslq %edi, %rdi #149.17
addq %r15, %rdi #149.17
vmovupd %zmm0, (%r14,%rdi,8){%k4} #149.17
# LOE rsi r12 r13 r14 r15 ebx ymm17 ymm19 zmm16 zmm18
..B7.70: # Preds ..B7.68 ..B7.50 ..B7.90
# Execution count [7.28e-01]: Infreq
incq %rsi #147.9
lea 2(%r15,%r13), %r15 #147.9
addq 24(%rsp), %r12 #147.9[spill]
cmpq 32(%rsp), %rsi #147.9[spill]
jb ..B7.50 # Prob 82% #147.9
jmp ..B7.45 # Prob 100% #147.9
# LOE rsi r12 r13 r14 r15 ebx ymm17 ymm19 zmm16 zmm18
..B7.72: # Preds ..B7.53
# Execution count [7.20e-02]: Infreq
movq %rsi, 24(%rsp) #[spill]
movq %rax, %rsi #
movq %r8, 32(%rsp) #[spill]
movl %edx, %ebx #
movq 48(%rsp), %r14 #[spill]
movq %rcx, %r13 #
movl (%rsp), %edi #148.13[spill]
xorl %eax, %eax #148.13
jmp ..B7.64 # Prob 100% #148.13
# LOE rsi r12 r13 r14 r15 eax ebx edi ymm17 ymm19 zmm16 zmm18
..B7.73: # Preds ..B7.51
# Execution count [7.20e-02]: Infreq
xorl %edi, %edi #148.13
jmp ..B7.68 # Prob 100% #148.13
# LOE rsi r12 r13 r14 r15 ebx edi ymm17 ymm19 zmm16 zmm18
..B7.74: # Preds ..B7.62
# Execution count [7.20e-02]: Infreq
incq %rax #147.9
addq %rsi, %r12 #147.9
lea 2(%r15,%rcx), %r15 #147.9
cmpq %r8, %rax #147.9
jb ..B7.53 # Prob 82% #147.9
jmp ..B7.45 # Prob 100% #147.9
# LOE rax rcx rsi r8 r12 r15 edx ymm17 ymm19 zmm16 zmm18
..B7.76: # Preds ..B7.10
# Execution count [4.50e-01]: Infreq
movl (%rsp), %r14d #141.9[spill]
xorl %r15d, %r15d #141.9
jmp ..B7.19 # Prob 100% #141.9
# LOE r12 r13 ebx r14d r15d
..B7.77: # Preds ..B7.9
# Execution count [4.50e-01]: Infreq
xorl %r14d, %r14d #141.9
jmp ..B7.23 # Prob 100% #141.9
.align 16,0x90
# LOE r12 r13 ebx r14d
.cfi_endproc
# mark_end;
.type initSolver,@function
.size initSolver,.-initSolver
..LNinitSolver.6:
.data
# -- End initSolver
.section .rodata, "a"
.align 64
.align 64
.L_2il0floatpacket.0:
.long 0x00000010,0x00000010,0x00000010,0x00000010,0x00000010,0x00000010,0x00000010,0x00000010,0x00000010,0x00000010,0x00000010,0x00000010,0x00000010,0x00000010,0x00000010,0x00000010
.type .L_2il0floatpacket.0,@object
.size .L_2il0floatpacket.0,64
.align 64
.L_2il0floatpacket.1:
.long 0x00000001,0x00000001,0x00000001,0x00000001,0x00000001,0x00000001,0x00000001,0x00000001,0x00000001,0x00000001,0x00000001,0x00000001,0x00000001,0x00000001,0x00000001,0x00000001
.type .L_2il0floatpacket.1,@object
.size .L_2il0floatpacket.1,64
.align 64
.L_2il0floatpacket.2:
.long 0x00000000,0x00000001,0x00000002,0x00000003,0x00000004,0x00000005,0x00000006,0x00000007,0x00000008,0x00000009,0x0000000a,0x0000000b,0x0000000c,0x0000000d,0x0000000e,0x0000000f
.type .L_2il0floatpacket.2,@object
.size .L_2il0floatpacket.2,64
.align 64
.L_2il0floatpacket.3:
.long 0x00000001,0x00000002,0x00000003,0x00000004,0x00000005,0x00000006,0x00000007,0x00000008,0x00000009,0x0000000a,0x0000000b,0x0000000c,0x0000000d,0x0000000e,0x0000000f,0x00000010
.type .L_2il0floatpacket.3,@object
.size .L_2il0floatpacket.3,64
.align 64
.L_2il0floatpacket.4:
.long 0x00000000,0x00000001,0x00000002,0x00000003,0x00000004,0x00000005,0x00000006,0x00000007,0x00000008,0x00000009,0x0000000a,0x0000000b,0x0000000c,0x0000000d,0x0000000e,0x0000000f
.type .L_2il0floatpacket.4,@object
.size .L_2il0floatpacket.4,64
.align 64
.L_2il0floatpacket.11:
.long 0x00000000,0x00000000,0x00000002,0x00000000,0x00000004,0x00000000,0x00000006,0x00000000,0x00000008,0x00000000,0x0000000a,0x00000000,0x0000000c,0x00000000,0x0000000e,0x00000000
.type .L_2il0floatpacket.11,@object
.size .L_2il0floatpacket.11,64
.align 64
.L_2il0floatpacket.12:
.long 0x00000001,0x00000000,0x00000003,0x00000000,0x00000005,0x00000000,0x00000007,0x00000000,0x00000009,0x00000000,0x0000000b,0x00000000,0x0000000d,0x00000000,0x0000000f,0x00000000
.type .L_2il0floatpacket.12,@object
.size .L_2il0floatpacket.12,64
.align 64
.L_2il0floatpacket.13:
.long 0x00000000,0x40000000,0x00000000,0x40000000,0x00000000,0x40000000,0x00000000,0x40000000,0x00000000,0x40000000,0x00000000,0x40000000,0x00000000,0x40000000,0x00000000,0x40000000
.type .L_2il0floatpacket.13,@object
.size .L_2il0floatpacket.13,64
.align 64
.L_2il0floatpacket.14:
.long 0x00000000,0x00000000,0x00000002,0x00000000,0x00000004,0x00000000,0x00000006,0x00000000,0x00000008,0x00000000,0x0000000a,0x00000000,0x0000000c,0x00000000,0x0000000e,0x00000000
.type .L_2il0floatpacket.14,@object
.size .L_2il0floatpacket.14,64
.align 64
.L_2il0floatpacket.15:
.long 0x00000000,0x00000000,0x00000000,0x00000000,0x00000001,0x00000000,0x00000000,0x00000000,0x00000002,0x00000000,0x00000000,0x00000000,0x00000003,0x00000000,0x00000000,0x00000000
.type .L_2il0floatpacket.15,@object
.size .L_2il0floatpacket.15,64
.align 64
.L_2il0floatpacket.16:
.long 0x00000004,0x00000000,0x00000000,0x00000000,0x00000005,0x00000000,0x00000000,0x00000000,0x00000006,0x00000000,0x00000000,0x00000000,0x00000007,0x00000000,0x00000000,0x00000000
.type .L_2il0floatpacket.16,@object
.size .L_2il0floatpacket.16,64
.align 64
.L_2il0floatpacket.17:
.long 0x00020001,0x00080004,0x00200010,0x00800040,0x00000000,0x00000000,0x00000000,0x00000000,0x00000000,0x00000000,0x00000000,0x00000000,0x00000000,0x00000000,0x00000000,0x00000000
.type .L_2il0floatpacket.17,@object
.size .L_2il0floatpacket.17,64
.align 32
.L_2il0floatpacket.6:
.long 0x00000008,0x00000008,0x00000008,0x00000008,0x00000008,0x00000008,0x00000008,0x00000008
.type .L_2il0floatpacket.6,@object
.size .L_2il0floatpacket.6,32
.align 32
.L_2il0floatpacket.7:
.long 0x00000000,0x00000001,0x00000002,0x00000003,0x00000004,0x00000005,0x00000006,0x00000007
.type .L_2il0floatpacket.7,@object
.size .L_2il0floatpacket.7,32
.align 32
.L_2il0floatpacket.20:
.long 0x00000000,0x00000001,0x00000002,0x00000003,0x00000004,0x00000005,0x00000006,0x00000007
.type .L_2il0floatpacket.20,@object
.size .L_2il0floatpacket.20,32
.align 8
.L_2il0floatpacket.5:
.long 0x00000000,0x3fe00000
.type .L_2il0floatpacket.5,@object
.size .L_2il0floatpacket.5,8
.align 8
.L_2il0floatpacket.8:
.long 0x00000000,0x40000000
.type .L_2il0floatpacket.8,@object
.size .L_2il0floatpacket.8,8
.align 8
.L_2il0floatpacket.9:
.long 0x00000000,0x3ff00000
.type .L_2il0floatpacket.9,@object
.size .L_2il0floatpacket.9,8
.align 8
.L_2il0floatpacket.10:
.long 0x00000000,0x3fd00000
.type .L_2il0floatpacket.10,@object
.size .L_2il0floatpacket.10,8
.align 8
.L_2il0floatpacket.18:
.long 0x54442d18,0x402921fb
.type .L_2il0floatpacket.18,@object
.size .L_2il0floatpacket.18,8
.align 8
.L_2il0floatpacket.19:
.long 0x54442d18,0x401921fb
.type .L_2il0floatpacket.19,@object
.size .L_2il0floatpacket.19,8
.section .rodata.str1.4, "aMS",@progbits,1
.space 1, 0x00 # pad
.align 4
.L_2__STRING.4:
.long 1633955440
.word 116
.type .L_2__STRING.4,@object
.size .L_2__STRING.4,6
.space 2, 0x00 # pad
.align 4
.L_2__STRING.9:
.word 119
.type .L_2__STRING.9,@object
.size .L_2__STRING.9,2
.space 2, 0x00 # pad
.align 4
.L_2__STRING.11:
.long 2123301
.type .L_2__STRING.11,@object
.size .L_2__STRING.11,4
.align 4
.L_2__STRING.1:
.long 1681010725
.word 8250
.byte 0
.type .L_2__STRING.1,@object
.size .L_2__STRING.1,7
.space 1, 0x00 # pad
.align 4
.L_2__STRING.2:
.long 775041317
.long 538994232
.byte 0
.type .L_2__STRING.2,@object
.size .L_2__STRING.2,9
.space 3, 0x00 # pad
.align 4
.L_2__STRING.6:
.long 1986817875
.long 1948283493
.long 543911791
.long 1763730469
.long 1634887028
.long 1852795252
.word 2675
.byte 0
.type .L_2__STRING.6,@object
.size .L_2__STRING.6,27
.space 1, 0x00 # pad
.align 4
.L_2__STRING.7:
.long 1634625862
.long 1835999340
.long 979461989
.long 174466336
.byte 0
.type .L_2__STRING.7,@object
.size .L_2__STRING.7,17
.space 3, 0x00 # pad
.align 4
.L_2__STRING.8:
.long 1986817875
.long 1948283493
.long 543911791
.long 1763730469
.long 1634887028
.long 1852795252
.long 1869881459
.long 1634038304
.long 622880867
.word 2662
.byte 0
.type .L_2__STRING.8,@object
.size .L_2__STRING.8,39
.space 1, 0x00 # pad
.align 4
.L_2__STRING.5:
.long 1263419730
.long 979641632
.long 1634560288
.long 1668238456
.long 975203425
.long 744760608
.long 1634560544
.long 1668238456
.long 975203425
.long 174335264
.byte 0
.type .L_2__STRING.5,@object
.size .L_2__STRING.5,41
.section .rodata.str1.32, "aMS",@progbits,1
.align 32
.align 32
.L_2__STRING.0:
.long 539173667
.long 1263419730
.long 543434016
.long 589505315
.long 589505315
.long 589505315
.long 589505315
.long 589505315
.long 589505315
.long 589505315
.long 589505315
.long 589505315
.long 589505315
.long 589505315
.long 589505315
.long 589505315
.long 170074915
.byte 0
.type .L_2__STRING.0,@object
.size .L_2__STRING.0,69
.data
.section .note.GNU-stack, ""
# End
Reference in New Issue View Git Blame Copy Permalink