jump32divsteps_mod2.S
// bitslice functions
.p2align 2,,3
.syntax unified
.text
.global jump32divsteps_mod2
.type jump32divsteps_mod2, %function
//int jump32divsteps_mod2(int delta, int *M, int *f, int *g);
jump32divsteps_mod2:
push {r4-r11,lr}
vmov s2, r1 // save result matrix ptr
vmov s3, r0 // save delta
ldr r4, [r2]
ldr r5, [r2, #4]
ldr r6, [r2, #8]
ldr r7, [r2, #12]
mov.w r0, #0 // f
ubfx.w r8, r4, #0, #1
ubfx.w r9, r4, #4, #1
ubfx.w r10, r4, #8, #1
ubfx.w r11, r4, #12, #1
eor.w r0, r8, r0, LSL #1
eor.w r0, r9, r0, LSL #1
eor.w r0, r10, r0, LSL #1
eor.w r0, r11, r0, LSL #1
ubfx.w r8, r4, #16, #1
ubfx.w r9, r4, #20, #1
ubfx.w r10, r4, #24, #1
ubfx.w r11, r4, #28, #1
eor.w r0, r8, r0, LSL #1
eor.w r0, r9, r0, LSL #1
eor.w r0, r10, r0, LSL #1
eor.w r0, r11, r0, LSL #1
ubfx.w r8, r5, #0, #1
ubfx.w r9, r5, #4, #1
ubfx.w r10, r5, #8, #1
ubfx.w r11, r5, #12, #1
eor.w r0, r8, r0, LSL #1
eor.w r0, r9, r0, LSL #1
eor.w r0, r10, r0, LSL #1
eor.w r0, r11, r0, LSL #1
ubfx.w r8, r5, #16, #1
ubfx.w r9, r5, #20, #1
ubfx.w r10, r5, #24, #1
ubfx.w r11, r5, #28, #1
eor.w r0, r8, r0, LSL #1
eor.w r0, r9, r0, LSL #1
eor.w r0, r10, r0, LSL #1
eor.w r0, r11, r0, LSL #1
ubfx.w r8, r6, #0, #1
ubfx.w r9, r6, #4, #1
ubfx.w r10, r6, #8, #1
ubfx.w r11, r6, #12, #1
eor.w r0, r8, r0, LSL #1
eor.w r0, r9, r0, LSL #1
eor.w r0, r10, r0, LSL #1
eor.w r0, r11, r0, LSL #1
ubfx.w r8, r6, #16, #1
ubfx.w r9, r6, #20, #1
ubfx.w r10, r6, #24, #1
ubfx.w r11, r6, #28, #1
eor.w r0, r8, r0, LSL #1
eor.w r0, r9, r0, LSL #1
eor.w r0, r10, r0, LSL #1
eor.w r0, r11, r0, LSL #1
ubfx.w r8, r7, #0, #1
ubfx.w r9, r7, #4, #1
ubfx.w r10, r7, #8, #1
ubfx.w r11, r7, #12, #1
eor.w r0, r8, r0, LSL #1
eor.w r0, r9, r0, LSL #1
eor.w r0, r10, r0, LSL #1
eor.w r0, r11, r0, LSL #1
ubfx.w r8, r7, #16, #1
ubfx.w r9, r7, #20, #1
ubfx.w r10, r7, #24, #1
ubfx.w r11, r7, #28, #1
eor.w r0, r8, r0, LSL #1
eor.w r0, r9, r0, LSL #1
eor.w r0, r10, r0, LSL #1
eor.w r0, r11, r0, LSL #1
ldr r4, [r3]
ldr r5, [r3, #4]
ldr r6, [r3, #8]
ldr r7, [r3, #12]
mov.w r1, #0 // g
ubfx.w r8, r4, #0, #1
ubfx.w r9, r4, #4, #1
ubfx.w r10, r4, #8, #1
ubfx.w r11, r4, #12, #1
eor.w r1, r8, r1, LSL #1
eor.w r1, r9, r1, LSL #1
eor.w r1, r10, r1, LSL #1
eor.w r1, r11, r1, LSL #1
ubfx.w r8, r4, #16, #1
ubfx.w r9, r4, #20, #1
ubfx.w r10, r4, #24, #1
ubfx.w r11, r4, #28, #1
eor.w r1, r8, r1, LSL #1
eor.w r1, r9, r1, LSL #1
eor.w r1, r10, r1, LSL #1
eor.w r1, r11, r1, LSL #1
ubfx.w r8, r5, #0, #1
ubfx.w r9, r5, #4, #1
ubfx.w r10, r5, #8, #1
ubfx.w r11, r5, #12, #1
eor.w r1, r8, r1, LSL #1
eor.w r1, r9, r1, LSL #1
eor.w r1, r10, r1, LSL #1
eor.w r1, r11, r1, LSL #1
ubfx.w r8, r5, #16, #1
ubfx.w r9, r5, #20, #1
ubfx.w r10, r5, #24, #1
ubfx.w r11, r5, #28, #1
eor.w r1, r8, r1, LSL #1
eor.w r1, r9, r1, LSL #1
eor.w r1, r10, r1, LSL #1
eor.w r1, r11, r1, LSL #1
ubfx.w r8, r6, #0, #1
ubfx.w r9, r6, #4, #1
ubfx.w r10, r6, #8, #1
ubfx.w r11, r6, #12, #1
eor.w r1, r8, r1, LSL #1
eor.w r1, r9, r1, LSL #1
eor.w r1, r10, r1, LSL #1
eor.w r1, r11, r1, LSL #1
ubfx.w r8, r6, #16, #1
ubfx.w r9, r6, #20, #1
ubfx.w r10, r6, #24, #1
ubfx.w r11, r6, #28, #1
eor.w r1, r8, r1, LSL #1
eor.w r1, r9, r1, LSL #1
eor.w r1, r10, r1, LSL #1
eor.w r1, r11, r1, LSL #1
ubfx.w r8, r7, #0, #1
ubfx.w r9, r7, #4, #1
ubfx.w r10, r7, #8, #1
ubfx.w r11, r7, #12, #1
eor.w r1, r8, r1, LSL #1
eor.w r1, r9, r1, LSL #1
eor.w r1, r10, r1, LSL #1
eor.w r1, r11, r1, LSL #1
ubfx.w r8, r7, #16, #1
ubfx.w r9, r7, #20, #1
ubfx.w r10, r7, #24, #1
ubfx.w r11, r7, #28, #1
eor.w r1, r8, r1, LSL #1
eor.w r1, r9, r1, LSL #1
eor.w r1, r10, r1, LSL #1
eor.w r1, r11, r1, LSL #1
mov r6, #(1<<31)
mov r9, #(1<<31)
mov r7, #0
mov r8, #0
vmov.f32 s0, #31.0
vmov.f32 s1, #1.0 // float 1.0 #112
vcvt.f32.s32 s3, s3 // convert to float
bs2_jump32divsteps_0: // first half
vcmp.f32 s3, s1 // delta > 0?
vmrs APSR_nzcv, FPSCR // move carry
itttt cs
tstcs r0, r1, LSL #1 // set cs by g0[0], then if cs
movcs r5, r0 // f0<->g0
movcs r0, r1
movcs r1, r5
ittt cs
movcs r5, r6 // u0<->r0
movcs r6, r8
movcs r8, r5
itttt cs
movcs r5, r7 // v0<->s0
movcs r7, r9
movcs r9, r5
vnegcs.f32 s3, s3
bs2_jump32divsteps_1: // second half
vadd.f32 s3, s3, s1 // delta++
and r4, r6, r1, ASR #31 // d0 = a0 & b0
eors r8, r8, r4 // (a0^b0)
and r4, r7, r1, ASR #31 // d0 = a0 & b0
eors r9, r9, r4 // (a0^b0)
and r4, r0, r1, ASR #31 // d0 = a0 & b0
eors r1, r1, r4 // (a0^b0)
lsl r1, r1, #1 // g = g/x
vsub.f32 s0, s0, s1
vcmp.f32 s0, #0.0
vmrs APSR_nzcv, FPSCR // move c flag
ittt cs // u = xu, v = xv if ct >= 0
lsrcs r6, r6, #1
lsrcs r7, r7, #1
bcs bs2_jump32divsteps_0
bs2_jump32divsteps_2: // clean up
and.w r10, r0, #0x1
ubfx.w r3, r0, #1, #1
eor.w r10, r3, r10, LSL #4
ubfx.w r4, r0, #2, #1
eor.w r10, r4, r10, LSL #4
ubfx.w r5, r0, #3, #1
eor.w r10, r5, r10, LSL #4
ubfx.w r2, r0, #4, #1
eor.w r10, r2, r10, LSL #4
ubfx.w r3, r0, #5, #1
eor.w r10, r3, r10, LSL #4
ubfx.w r4, r0, #6, #1
eor.w r10, r4, r10, LSL #4
ubfx.w r5, r0, #7, #1
eor.w r10, r5, r10, LSL #4
ubfx.w r2, r0, #8, #1
mov.w r11, r2
ubfx.w r3, r0, #9, #1
eor.w r11, r3, r11, LSL #4
ubfx.w r4, r0, #10, #1
eor.w r11, r4, r11, LSL #4
ubfx.w r5, r0, #11, #1
eor.w r11, r5, r11, LSL #4
ubfx.w r2, r0, #12, #1
eor.w r11, r2, r11, LSL #4
ubfx.w r3, r0, #13, #1
eor.w r11, r3, r11, LSL #4
ubfx.w r4, r0, #14, #1
eor.w r11, r4, r11, LSL #4
ubfx.w r5, r0, #15, #1
eor.w r11, r5, r11, LSL #4
ubfx.w r2, r0, #16, #1
mov.w r12, r2
ubfx.w r3, r0, #17, #1
eor.w r12, r3, r12, LSL #4
ubfx.w r4, r0, #18, #1
eor.w r12, r4, r12, LSL #4
ubfx.w r5, r0, #19, #1
eor.w r12, r5, r12, LSL #4
ubfx.w r2, r0, #20, #1
eor.w r12, r2, r12, LSL #4
ubfx.w r3, r0, #21, #1
eor.w r12, r3, r12, LSL #4
ubfx.w r4, r0, #22, #1
eor.w r12, r4, r12, LSL #4
ubfx.w r5, r0, #23, #1
eor.w r12, r5, r12, LSL #4
ubfx.w r2, r0, #24, #1
mov.w lr, r2
ubfx.w r3, r0, #25, #1
eor.w lr, r3, lr, LSL #4
ubfx.w r4, r0, #26, #1
eor.w lr, r4, lr, LSL #4
ubfx.w r5, r0, #27, #1
eor.w lr, r5, lr, LSL #4
ubfx.w r2, r0, #28, #1
eor.w lr, r2, lr, LSL #4
ubfx.w r3, r0, #29, #1
eor.w lr, r3, lr, LSL #4
ubfx.w r4, r0, #30, #1
eor.w lr, r4, lr, LSL #4
ubfx.w r5, r0, #31, #1
eor.w lr, r5, lr, LSL #4
vmov r0, s2 // reload output ptr for results
str r12, [r0, #4]
str r11, [r0, #8]
str r10, [r0, #12]
str lr, [r0], #16
and.w r10, r1, #0x1
ubfx.w r3, r1, #1, #1
eor.w r10, r3, r10, LSL #4
ubfx.w r4, r1, #2, #1
eor.w r10, r4, r10, LSL #4
ubfx.w r5, r1, #3, #1
eor.w r10, r5, r10, LSL #4
ubfx.w r2, r1, #4, #1
eor.w r10, r2, r10, LSL #4
ubfx.w r3, r1, #5, #1
eor.w r10, r3, r10, LSL #4
ubfx.w r4, r1, #6, #1
eor.w r10, r4, r10, LSL #4
ubfx.w r5, r1, #7, #1
eor.w r10, r5, r10, LSL #4
ubfx.w r2, r1, #8, #1
mov.w r11, r2
ubfx.w r3, r1, #9, #1
eor.w r11, r3, r11, LSL #4
ubfx.w r4, r1, #10, #1
eor.w r11, r4, r11, LSL #4
ubfx.w r5, r1, #11, #1
eor.w r11, r5, r11, LSL #4
ubfx.w r2, r1, #12, #1
eor.w r11, r2, r11, LSL #4
ubfx.w r3, r1, #13, #1
eor.w r11, r3, r11, LSL #4
ubfx.w r4, r1, #14, #1
eor.w r11, r4, r11, LSL #4
ubfx.w r5, r1, #15, #1
eor.w r11, r5, r11, LSL #4
ubfx.w r2, r1, #16, #1
mov.w r12, r2
ubfx.w r3, r1, #17, #1
eor.w r12, r3, r12, LSL #4
ubfx.w r4, r1, #18, #1
eor.w r12, r4, r12, LSL #4
ubfx.w r5, r1, #19, #1
eor.w r12, r5, r12, LSL #4
ubfx.w r2, r1, #20, #1
eor.w r12, r2, r12, LSL #4
ubfx.w r3, r1, #21, #1
eor.w r12, r3, r12, LSL #4
ubfx.w r4, r1, #22, #1
eor.w r12, r4, r12, LSL #4
ubfx.w r5, r1, #23, #1
eor.w r12, r5, r12, LSL #4
ubfx.w r2, r1, #24, #1
mov.w lr, r2
ubfx.w r3, r1, #25, #1
eor.w lr, r3, lr, LSL #4
ubfx.w r4, r1, #26, #1
eor.w lr, r4, lr, LSL #4
ubfx.w r5, r1, #27, #1
eor.w lr, r5, lr, LSL #4
ubfx.w r2, r1, #28, #1
eor.w lr, r2, lr, LSL #4
ubfx.w r3, r1, #29, #1
eor.w lr, r3, lr, LSL #4
ubfx.w r4, r1, #30, #1
eor.w lr, r4, lr, LSL #4
ubfx.w r5, r1, #31, #1
eor.w lr, r5, lr, LSL #4
str r12, [r0, #4]
str r11, [r0, #8]
str r10, [r0, #12]
str lr, [r0], #16
and.w r10, r6, #0x1
ubfx.w r3, r6, #1, #1
eor.w r10, r3, r10, LSL #4
ubfx.w r4, r6, #2, #1
eor.w r10, r4, r10, LSL #4
ubfx.w r5, r6, #3, #1
eor.w r10, r5, r10, LSL #4
ubfx.w r2, r6, #4, #1
eor.w r10, r2, r10, LSL #4
ubfx.w r3, r6, #5, #1
eor.w r10, r3, r10, LSL #4
ubfx.w r4, r6, #6, #1
eor.w r10, r4, r10, LSL #4
ubfx.w r5, r6, #7, #1
eor.w r10, r5, r10, LSL #4
ubfx.w r2, r6, #8, #1
mov.w r11, r2
ubfx.w r3, r6, #9, #1
eor.w r11, r3, r11, LSL #4
ubfx.w r4, r6, #10, #1
eor.w r11, r4, r11, LSL #4
ubfx.w r5, r6, #11, #1
eor.w r11, r5, r11, LSL #4
ubfx.w r2, r6, #12, #1
eor.w r11, r2, r11, LSL #4
ubfx.w r3, r6, #13, #1
eor.w r11, r3, r11, LSL #4
ubfx.w r4, r6, #14, #1
eor.w r11, r4, r11, LSL #4
ubfx.w r5, r6, #15, #1
eor.w r11, r5, r11, LSL #4
ubfx.w r2, r6, #16, #1
mov.w r12, r2
ubfx.w r3, r6, #17, #1
eor.w r12, r3, r12, LSL #4
ubfx.w r4, r6, #18, #1
eor.w r12, r4, r12, LSL #4
ubfx.w r5, r6, #19, #1
eor.w r12, r5, r12, LSL #4
ubfx.w r2, r6, #20, #1
eor.w r12, r2, r12, LSL #4
ubfx.w r3, r6, #21, #1
eor.w r12, r3, r12, LSL #4
ubfx.w r4, r6, #22, #1
eor.w r12, r4, r12, LSL #4
ubfx.w r5, r6, #23, #1
eor.w r12, r5, r12, LSL #4
ubfx.w r2, r6, #24, #1
mov.w lr, r2
ubfx.w r3, r6, #25, #1
eor.w lr, r3, lr, LSL #4
ubfx.w r4, r6, #26, #1
eor.w lr, r4, lr, LSL #4
ubfx.w r5, r6, #27, #1
eor.w lr, r5, lr, LSL #4
ubfx.w r2, r6, #28, #1
eor.w lr, r2, lr, LSL #4
ubfx.w r3, r6, #29, #1
eor.w lr, r3, lr, LSL #4
ubfx.w r4, r6, #30, #1
eor.w lr, r4, lr, LSL #4
ubfx.w r5, r6, #31, #1
eor.w lr, r5, lr, LSL #4
str r12, [r0, #4]
str r11, [r0, #8]
str r10, [r0, #12]
str lr, [r0], #16
and.w r10, r7, #0x1
ubfx.w r3, r7, #1, #1
eor.w r10, r3, r10, LSL #4
ubfx.w r4, r7, #2, #1
eor.w r10, r4, r10, LSL #4
ubfx.w r5, r7, #3, #1
eor.w r10, r5, r10, LSL #4
ubfx.w r2, r7, #4, #1
eor.w r10, r2, r10, LSL #4
ubfx.w r3, r7, #5, #1
eor.w r10, r3, r10, LSL #4
ubfx.w r4, r7, #6, #1
eor.w r10, r4, r10, LSL #4
ubfx.w r5, r7, #7, #1
eor.w r10, r5, r10, LSL #4
ubfx.w r2, r7, #8, #1
mov.w r11, r2
ubfx.w r3, r7, #9, #1
eor.w r11, r3, r11, LSL #4
ubfx.w r4, r7, #10, #1
eor.w r11, r4, r11, LSL #4
ubfx.w r5, r7, #11, #1
eor.w r11, r5, r11, LSL #4
ubfx.w r2, r7, #12, #1
eor.w r11, r2, r11, LSL #4
ubfx.w r3, r7, #13, #1
eor.w r11, r3, r11, LSL #4
ubfx.w r4, r7, #14, #1
eor.w r11, r4, r11, LSL #4
ubfx.w r5, r7, #15, #1
eor.w r11, r5, r11, LSL #4
ubfx.w r2, r7, #16, #1
mov.w r12, r2
ubfx.w r3, r7, #17, #1
eor.w r12, r3, r12, LSL #4
ubfx.w r4, r7, #18, #1
eor.w r12, r4, r12, LSL #4
ubfx.w r5, r7, #19, #1
eor.w r12, r5, r12, LSL #4
ubfx.w r2, r7, #20, #1
eor.w r12, r2, r12, LSL #4
ubfx.w r3, r7, #21, #1
eor.w r12, r3, r12, LSL #4
ubfx.w r4, r7, #22, #1
eor.w r12, r4, r12, LSL #4
ubfx.w r5, r7, #23, #1
eor.w r12, r5, r12, LSL #4
ubfx.w r2, r7, #24, #1
mov.w lr, r2
ubfx.w r3, r7, #25, #1
eor.w lr, r3, lr, LSL #4
ubfx.w r4, r7, #26, #1
eor.w lr, r4, lr, LSL #4
ubfx.w r5, r7, #27, #1
eor.w lr, r5, lr, LSL #4
ubfx.w r2, r7, #28, #1
eor.w lr, r2, lr, LSL #4
ubfx.w r3, r7, #29, #1
eor.w lr, r3, lr, LSL #4
ubfx.w r4, r7, #30, #1
eor.w lr, r4, lr, LSL #4
ubfx.w r5, r7, #31, #1
eor.w lr, r5, lr, LSL #4
str r12, [r0, #4]
str r11, [r0, #8]
str r10, [r0, #12]
str lr, [r0], #16
and.w r10, r8, #0x1
ubfx.w r3, r8, #1, #1
eor.w r10, r3, r10, LSL #4
ubfx.w r4, r8, #2, #1
eor.w r10, r4, r10, LSL #4
ubfx.w r5, r8, #3, #1
eor.w r10, r5, r10, LSL #4
ubfx.w r2, r8, #4, #1
eor.w r10, r2, r10, LSL #4
ubfx.w r3, r8, #5, #1
eor.w r10, r3, r10, LSL #4
ubfx.w r4, r8, #6, #1
eor.w r10, r4, r10, LSL #4
ubfx.w r5, r8, #7, #1
eor.w r10, r5, r10, LSL #4
ubfx.w r2, r8, #8, #1
mov.w r11, r2
ubfx.w r3, r8, #9, #1
eor.w r11, r3, r11, LSL #4
ubfx.w r4, r8, #10, #1
eor.w r11, r4, r11, LSL #4
ubfx.w r5, r8, #11, #1
eor.w r11, r5, r11, LSL #4
ubfx.w r2, r8, #12, #1
eor.w r11, r2, r11, LSL #4
ubfx.w r3, r8, #13, #1
eor.w r11, r3, r11, LSL #4
ubfx.w r4, r8, #14, #1
eor.w r11, r4, r11, LSL #4
ubfx.w r5, r8, #15, #1
eor.w r11, r5, r11, LSL #4
ubfx.w r2, r8, #16, #1
mov.w r12, r2
ubfx.w r3, r8, #17, #1
eor.w r12, r3, r12, LSL #4
ubfx.w r4, r8, #18, #1
eor.w r12, r4, r12, LSL #4
ubfx.w r5, r8, #19, #1
eor.w r12, r5, r12, LSL #4
ubfx.w r2, r8, #20, #1
eor.w r12, r2, r12, LSL #4
ubfx.w r3, r8, #21, #1
eor.w r12, r3, r12, LSL #4
ubfx.w r4, r8, #22, #1
eor.w r12, r4, r12, LSL #4
ubfx.w r5, r8, #23, #1
eor.w r12, r5, r12, LSL #4
ubfx.w r2, r8, #24, #1
mov.w lr, r2
ubfx.w r3, r8, #25, #1
eor.w lr, r3, lr, LSL #4
ubfx.w r4, r8, #26, #1
eor.w lr, r4, lr, LSL #4
ubfx.w r5, r8, #27, #1
eor.w lr, r5, lr, LSL #4
ubfx.w r2, r8, #28, #1
eor.w lr, r2, lr, LSL #4
ubfx.w r3, r8, #29, #1
eor.w lr, r3, lr, LSL #4
ubfx.w r4, r8, #30, #1
eor.w lr, r4, lr, LSL #4
ubfx.w r5, r8, #31, #1
eor.w lr, r5, lr, LSL #4
str r12, [r0, #4]
str r11, [r0, #8]
str r10, [r0, #12]
str lr, [r0], #16
and.w r10, r9, #0x1
ubfx.w r3, r9, #1, #1
eor.w r10, r3, r10, LSL #4
ubfx.w r4, r9, #2, #1
eor.w r10, r4, r10, LSL #4
ubfx.w r5, r9, #3, #1
eor.w r10, r5, r10, LSL #4
ubfx.w r2, r9, #4, #1
eor.w r10, r2, r10, LSL #4
ubfx.w r3, r9, #5, #1
eor.w r10, r3, r10, LSL #4
ubfx.w r4, r9, #6, #1
eor.w r10, r4, r10, LSL #4
ubfx.w r5, r9, #7, #1
eor.w r10, r5, r10, LSL #4
ubfx.w r2, r9, #8, #1
mov.w r11, r2
ubfx.w r3, r9, #9, #1
eor.w r11, r3, r11, LSL #4
ubfx.w r4, r9, #10, #1
eor.w r11, r4, r11, LSL #4
ubfx.w r5, r9, #11, #1
eor.w r11, r5, r11, LSL #4
ubfx.w r2, r9, #12, #1
eor.w r11, r2, r11, LSL #4
ubfx.w r3, r9, #13, #1
eor.w r11, r3, r11, LSL #4
ubfx.w r4, r9, #14, #1
eor.w r11, r4, r11, LSL #4
ubfx.w r5, r9, #15, #1
eor.w r11, r5, r11, LSL #4
ubfx.w r2, r9, #16, #1
mov.w r12, r2
ubfx.w r3, r9, #17, #1
eor.w r12, r3, r12, LSL #4
ubfx.w r4, r9, #18, #1
eor.w r12, r4, r12, LSL #4
ubfx.w r5, r9, #19, #1
eor.w r12, r5, r12, LSL #4
ubfx.w r2, r9, #20, #1
eor.w r12, r2, r12, LSL #4
ubfx.w r3, r9, #21, #1
eor.w r12, r3, r12, LSL #4
ubfx.w r4, r9, #22, #1
eor.w r12, r4, r12, LSL #4
ubfx.w r5, r9, #23, #1
eor.w r12, r5, r12, LSL #4
ubfx.w r2, r9, #24, #1
mov.w lr, r2
ubfx.w r3, r9, #25, #1
eor.w lr, r3, lr, LSL #4
ubfx.w r4, r9, #26, #1
eor.w lr, r4, lr, LSL #4
ubfx.w r5, r9, #27, #1
eor.w lr, r5, lr, LSL #4
ubfx.w r2, r9, #28, #1
eor.w lr, r2, lr, LSL #4
ubfx.w r3, r9, #29, #1
eor.w lr, r3, lr, LSL #4
ubfx.w r4, r9, #30, #1
eor.w lr, r4, lr, LSL #4
ubfx.w r5, r9, #31, #1
eor.w lr, r5, lr, LSL #4
str r12, [r0, #4]
str r11, [r0, #8]
str r10, [r0, #12]
str lr, [r0], #16
vcvt.s32.f32 s3, s3 // back to int
vmov r0, s3 // restore delta
pop {r4-r11,pc}
