1 |
.file "idct.c" |
addreg1 = r14 |
2 |
.pred.safe_across_calls p1-p5,p16-p63 |
addreg2 = r15 |
3 |
.sbss |
c0 = f32 |
4 |
.align 8 |
c1 = f33 |
5 |
.type blk.0#,@object |
c2 = f34 |
6 |
.size blk.0#,8 |
c3 = f35 |
7 |
blk.0: |
c4 = f36 |
8 |
.skip 8 |
c5 = f37 |
9 |
.align 8 |
c6 = f38 |
10 |
.type i.1#,@object |
c7 = f39 |
11 |
.size i.1#,8 |
c8 = f40 |
12 |
i.1: |
c9 = f41 |
13 |
.skip 8 |
c10 = f42 |
14 |
.align 8 |
c11 = f43 |
15 |
.type X0.2#,@object |
c12 = f44 |
16 |
.size X0.2#,8 |
c13 = f45 |
17 |
X0.2: |
c14 = f46 |
18 |
.skip 8 |
c15 = f47 |
19 |
.align 8 |
|
20 |
.type X1.3#,@object |
.sdata |
|
.size X1.3#,8 |
|
|
X1.3: |
|
|
.skip 8 |
|
|
.align 8 |
|
|
.type X2.4#,@object |
|
|
.size X2.4#,8 |
|
|
X2.4: |
|
|
.skip 8 |
|
|
.align 8 |
|
|
.type X3.5#,@object |
|
|
.size X3.5#,8 |
|
|
X3.5: |
|
|
.skip 8 |
|
|
.align 8 |
|
|
.type X4.6#,@object |
|
|
.size X4.6#,8 |
|
|
X4.6: |
|
|
.skip 8 |
|
|
.align 8 |
|
|
.type X5.7#,@object |
|
|
.size X5.7#,8 |
|
|
X5.7: |
|
|
.skip 8 |
|
|
.align 8 |
|
|
.type X6.8#,@object |
|
|
.size X6.8#,8 |
|
|
X6.8: |
|
|
.skip 8 |
|
|
.align 8 |
|
|
.type X7.9#,@object |
|
|
.size X7.9#,8 |
|
|
X7.9: |
|
|
.skip 8 |
|
|
.align 8 |
|
|
.type X8.10#,@object |
|
|
.size X8.10#,8 |
|
|
X8.10: |
|
|
.skip 8 |
|
|
.text |
|
21 |
.align 16 |
.align 16 |
22 |
.global idct_ia64# |
.data_c0: |
23 |
.proc idct_ia64# |
.single 0.353553390593273730857504233427, 0.353553390593273730857504233427 |
24 |
idct_ia64: |
.data_c1: |
25 |
.prologue |
.single -2.414213562373094923430016933708, -2.414213562373094923430016933708 |
26 |
alloc r16 = ar.pfs, 1, 4, 8, 0 |
.align 16 |
27 |
.body |
.data_c2: |
28 |
addl r14 = @gprel(i.1#), gp |
.single -0.414213562373095034452319396223, -0.414213562373095034452319396223 |
29 |
addl r33 = @gprel(blk.0#), gp |
.data_c3: |
30 |
addl r11 = @gprel(X1.3#), gp |
.single 0.198912367379658006072418174881, 0.198912367379658006072418174881 |
31 |
;; |
.align 16 |
32 |
st8 [r14] = r0 |
.data_c4: |
33 |
mov r41 = r14 |
.single 5.027339492125848074977056967327, 5.027339492125848074977056967327 |
34 |
addl r10 = @gprel(X2.4#), gp |
.data_c5: |
35 |
addl r14 = 565, r0 |
.single 0.668178637919298878955487452913, 0.668178637919298878955487452913 |
36 |
addl r9 = @gprel(X3.5#), gp |
.align 16 |
37 |
addl r8 = @gprel(X4.6#), gp |
.data_c6: |
38 |
;; |
.single 1.496605762665489169904731170391, 1.496605762665489169904731170391 |
39 |
setf.sig f14 = r14 |
.data_c7: |
40 |
addl r44 = @gprel(X5.7#), gp |
.single 0.461939766255643369241568052530, 0.461939766255643369241568052530 |
41 |
addl r43 = @gprel(X6.8#), gp |
.align 16 |
42 |
addl r14 = 2276, r0 |
.data_c8: |
43 |
addl r42 = @gprel(X7.9#), gp |
.single 0.191341716182544890889616340246, 0.191341716182544890889616340246 |
44 |
addl r3 = @gprel(X0.2#), gp |
.data_c9: |
45 |
;; |
.single 0.847759065022573476966272210120, 0.847759065022573476966272210120 |
|
setf.sig f15 = r14 |
|
|
addl r2 = @gprel(X8.10#), gp |
|
|
addl r14 = 3406, r0 |
|
|
;; |
|
|
setf.sig f13 = r14 |
|
|
addl r14 = 2408, r0 |
|
|
;; |
|
|
setf.sig f12 = r14 |
|
|
addl r14 = 799, r0 |
|
|
;; |
|
|
setf.sig f11 = r14 |
|
|
addl r14 = 4017, r0 |
|
|
;; |
|
|
setf.sig f9 = r14 |
|
|
addl r14 = 1108, r0 |
|
|
;; |
|
|
setf.sig f10 = r14 |
|
|
addl r14 = 3784, r0 |
|
|
;; |
|
|
setf.sig f8 = r14 |
|
|
addl r14 = 181, r0 |
|
|
;; |
|
|
setf.sig f7 = r14 |
|
|
.L6: |
|
|
ld8 r14 = [r41] |
|
|
;; |
|
|
shladd r27 = r14, 4, r32 |
|
|
;; |
|
|
adds r40 = 12, r27 |
|
|
adds r39 = 8, r27 |
|
|
adds r31 = 2, r27 |
|
|
;; |
|
|
ld2 r15 = [r40] |
|
|
ld2 r14 = [r39] |
|
|
adds r38 = 4, r27 |
|
|
;; |
|
|
sxt2 r28 = r15 |
|
|
sxt2 r14 = r14 |
|
|
adds r37 = 14, r27 |
|
|
ld2 r15 = [r31] |
|
|
;; |
|
|
shl r25 = r14, 11 |
|
|
ld2 r16 = [r38] |
|
|
sxt2 r23 = r15 |
|
|
;; |
|
|
st8 [r11] = r25 |
|
|
sxt2 r20 = r16 |
|
|
ld2 r15 = [r37] |
|
|
or r14 = r28, r25 |
|
|
adds r30 = 10, r27 |
|
|
;; |
|
|
sxt2 r22 = r15 |
|
|
st8 [r10] = r28 |
|
|
or r14 = r20, r14 |
|
|
ld2 r15 = [r30] |
|
|
st8 [r9] = r20 |
|
|
;; |
|
|
or r14 = r23, r14 |
|
|
sxt2 r19 = r15 |
|
|
adds r29 = 6, r27 |
|
|
st8 [r8] = r23 |
|
|
;; |
|
|
or r14 = r22, r14 |
|
|
ld2 r15 = [r29] |
|
|
st8 [r44] = r22 |
|
|
;; |
|
|
sxt2 r18 = r15 |
|
|
or r14 = r19, r14 |
|
|
st8 [r43] = r19 |
|
|
;; |
|
|
or r14 = r18, r14 |
|
|
st8 [r33] = r27 |
|
|
st8 [r42] = r18 |
|
|
;; |
|
|
cmp.ne p6, p7 = 0, r14 |
|
|
(p6) br.cond.dptk .L7 |
|
|
ld2 r14 = [r27] |
|
|
;; |
|
|
shladd r14 = r14, 3, r0 |
|
|
;; |
|
|
st2 [r37] = r14 |
|
|
st2 [r40] = r14 |
|
|
st2 [r30] = r14 |
|
|
st2 [r39] = r14 |
|
|
st2 [r29] = r14 |
|
|
st2 [r38] = r14 |
|
|
st2 [r31] = r14 |
|
|
st2 [r27] = r14 |
|
|
br .L5 |
|
|
.L7: |
|
|
add r21 = r19, r18 |
|
|
add r14 = r23, r22 |
|
|
ld2 r17 = [r27] |
|
|
;; |
|
|
setf.sig f32 = r21 |
|
|
setf.sig f6 = r14 |
|
|
add r15 = r20, r28 |
|
|
;; |
|
|
xma.l f32 = f32, f12, f0 |
|
|
shladd r16 = r20, 1, r20 |
|
|
sxt2 r17 = r17 |
|
|
;; |
|
|
getf.sig r21 = f32 |
|
|
xma.l f6 = f6, f14, f0 |
|
|
shladd r16 = r16, 4, r20 |
|
|
setf.sig f32 = r22 |
|
|
;; |
|
|
getf.sig r14 = f6 |
|
|
dep.z r17 = r17, 11, 21 |
|
|
xma.l f32 = f32, f13, f0 |
|
|
;; |
|
|
adds r17 = 128, r17 |
|
|
shl r16 = r16, 5 |
|
|
getf.sig r22 = f32 |
|
|
;; |
|
|
sxt4 r17 = r17 |
|
|
setf.sig f32 = r19 |
|
|
sub r22 = r14, r22 |
|
|
;; |
|
|
add r24 = r17, r25 |
|
|
xma.l f32 = f32, f11, f0 |
|
|
sub r17 = r17, r25 |
|
|
;; |
|
|
getf.sig r19 = f32 |
|
|
setf.sig f32 = r18 |
|
|
;; |
|
|
sub r19 = r21, r19 |
|
|
xma.l f32 = f32, f9, f0 |
|
|
;; |
|
|
getf.sig r14 = f32 |
|
|
setf.sig f32 = r23 |
|
|
;; |
|
|
sub r21 = r21, r14 |
|
|
xma.l f32 = f32, f15, f6 |
|
|
;; |
|
|
sub r26 = r22, r21 |
|
|
getf.sig r23 = f32 |
|
|
setf.sig f6 = r15 |
|
|
add r22 = r22, r21 |
|
|
;; |
|
|
sub r18 = r23, r19 |
|
|
xma.l f6 = f6, f10, f0 |
|
|
setf.sig f32 = r28 |
|
|
;; |
|
|
getf.sig r15 = f6 |
|
|
add r20 = r18, r26 |
|
|
xma.l f32 = f32, f8, f0 |
|
|
;; |
|
|
setf.sig f6 = r20 |
|
|
add r16 = r15, r16 |
|
|
getf.sig r14 = f32 |
|
|
sub r18 = r18, r26 |
|
|
;; |
|
|
xma.l f6 = f6, f7, f0 |
|
|
add r23 = r23, r19 |
|
|
sub r15 = r15, r14 |
|
|
setf.sig f32 = r18 |
|
|
;; |
|
|
getf.sig r20 = f6 |
|
|
add r25 = r24, r16 |
|
|
add r19 = r17, r15 |
|
|
;; |
|
|
adds r20 = 128, r20 |
|
|
xma.l f32 = f32, f7, f0 |
|
|
sub r17 = r17, r15 |
|
|
;; |
|
|
shr r20 = r20, 8 |
|
|
getf.sig r18 = f32 |
|
|
add r14 = r25, r23 |
|
|
;; |
|
|
add r15 = r19, r20 |
|
|
shr r14 = r14, 8 |
|
|
adds r18 = 128, r18 |
|
|
st8 [r11] = r23 |
|
|
;; |
|
|
st2 [r27] = r14 |
|
|
shr r18 = r18, 8 |
|
|
shr r15 = r15, 8 |
|
|
;; |
|
|
st2 [r31] = r15 |
|
|
st8 [r43] = r22 |
|
|
sub r24 = r24, r16 |
|
|
add r14 = r17, r18 |
|
|
st8 [r44] = r26 |
|
|
;; |
|
|
add r15 = r24, r22 |
|
|
shr r14 = r14, 8 |
|
|
st8 [r42] = r25 |
|
|
;; |
|
|
shr r15 = r15, 8 |
|
|
st2 [r38] = r14 |
|
|
sub r16 = r24, r22 |
|
|
st8 [r2] = r24 |
|
|
;; |
|
|
st2 [r29] = r15 |
|
|
shr r16 = r16, 8 |
|
|
sub r14 = r17, r18 |
|
|
;; |
|
|
st2 [r39] = r16 |
|
|
st8 [r9] = r19 |
|
|
sub r15 = r19, r20 |
|
|
shr r14 = r14, 8 |
|
|
st8 [r3] = r17 |
|
|
sub r16 = r25, r23 |
|
|
;; |
|
|
st2 [r30] = r14 |
|
|
shr r15 = r15, 8 |
|
|
shr r16 = r16, 8 |
|
|
;; |
|
|
st2 [r40] = r15 |
|
|
st8 [r10] = r20 |
|
|
st2 [r37] = r16 |
|
|
st8 [r8] = r18 |
|
|
.L5: |
|
|
ld8 r14 = [r41] |
|
|
;; |
|
|
adds r14 = 1, r14 |
|
|
;; |
|
|
st8 [r41] = r14 |
|
|
cmp.ge p6, p7 = 7, r14 |
|
|
(p6) br.cond.dptk .L6 |
|
|
addl r14 = @gprel(i.1#), gp |
|
|
addl r36 = @gprel(X8.10#), gp |
|
|
addl r35 = @gprel(blk.0#), gp |
|
|
;; |
|
|
st8 [r14] = r0 |
|
|
mov r42 = r14 |
|
|
addl r33 = @gprel(X1.3#), gp |
|
|
addl r14 = 565, r0 |
|
|
addl r3 = @gprel(X2.4#), gp |
|
|
addl r11 = @gprel(X3.5#), gp |
|
|
;; |
|
|
setf.sig f14 = r14 |
|
|
addl r10 = @gprel(X4.6#), gp |
|
|
addl r9 = @gprel(X5.7#), gp |
|
|
addl r14 = 2276, r0 |
|
|
addl r8 = @gprel(X6.8#), gp |
|
|
addl r44 = @gprel(X7.9#), gp |
|
|
;; |
|
|
setf.sig f12 = r14 |
|
|
addl r43 = @gprel(iclp_ia64#), gp |
|
|
addl r34 = @gprel(X0.2#), gp |
|
|
addl r14 = 3406, r0 |
|
|
;; |
|
|
setf.sig f13 = r14 |
|
|
addl r14 = 2408, r0 |
|
|
;; |
|
|
setf.sig f10 = r14 |
|
|
addl r14 = 799, r0 |
|
|
;; |
|
|
setf.sig f11 = r14 |
|
|
addl r14 = 4017, r0 |
|
|
;; |
|
|
setf.sig f8 = r14 |
|
|
addl r14 = 1108, r0 |
|
|
;; |
|
|
setf.sig f9 = r14 |
|
|
addl r14 = 3784, r0 |
|
|
;; |
|
|
setf.sig f7 = r14 |
|
|
addl r14 = 181, r0 |
|
|
;; |
|
|
setf.sig f6 = r14 |
|
|
.L12: |
|
|
ld8 r14 = [r42] |
|
|
;; |
|
|
shladd r29 = r14, 1, r32 |
|
|
;; |
|
|
adds r41 = 96, r29 |
|
|
adds r39 = 64, r29 |
|
|
adds r31 = 16, r29 |
|
|
;; |
|
|
ld2 r15 = [r41] |
|
|
ld2 r14 = [r39] |
|
|
adds r37 = 32, r29 |
|
|
;; |
|
|
sxt2 r25 = r15 |
|
|
sxt2 r14 = r14 |
|
|
adds r40 = 112, r29 |
|
|
ld2 r15 = [r31] |
|
|
;; |
|
|
shl r26 = r14, 8 |
|
|
ld2 r16 = [r37] |
|
|
sxt2 r23 = r15 |
|
|
;; |
|
|
st8 [r33] = r26 |
|
|
sxt2 r16 = r16 |
|
|
ld2 r15 = [r40] |
|
|
or r14 = r25, r26 |
|
|
adds r38 = 80, r29 |
|
|
;; |
|
|
sxt2 r24 = r15 |
|
|
st8 [r3] = r25 |
|
|
or r14 = r16, r14 |
|
|
ld2 r15 = [r38] |
|
|
st8 [r11] = r16 |
|
|
;; |
|
|
or r14 = r23, r14 |
|
|
sxt2 r22 = r15 |
|
|
adds r30 = 48, r29 |
|
|
st8 [r10] = r23 |
|
|
;; |
|
|
or r14 = r24, r14 |
|
|
ld2 r15 = [r30] |
|
|
st8 [r9] = r24 |
|
|
;; |
|
|
sxt2 r20 = r15 |
|
|
or r14 = r22, r14 |
|
|
st8 [r8] = r22 |
|
|
;; |
|
|
or r14 = r20, r14 |
|
|
st8 [r35] = r29 |
|
|
st8 [r44] = r20 |
|
|
;; |
|
|
cmp.ne p6, p7 = 0, r14 |
|
|
(p6) br.cond.dptk .L13 |
|
|
ld2 r14 = [r29] |
|
|
ld8 r16 = [r43] |
|
|
;; |
|
|
sxt2 r14 = r14 |
|
|
;; |
|
|
adds r14 = 32, r14 |
|
|
;; |
|
|
extr r14 = r14, 6, 26 |
|
|
;; |
|
|
shladd r14 = r14, 1, r16 |
|
|
;; |
|
|
ld2 r15 = [r14] |
|
|
;; |
|
|
st2 [r40] = r15 |
|
|
st2 [r41] = r15 |
|
|
st2 [r38] = r15 |
|
|
st2 [r39] = r15 |
|
|
st2 [r30] = r15 |
|
|
st2 [r37] = r15 |
|
|
st2 [r31] = r15 |
|
|
st2 [r29] = r15 |
|
|
br .L11 |
|
|
.L13: |
|
|
add r19 = r23, r24 |
|
|
add r21 = r22, r20 |
|
|
add r17 = r16, r25 |
|
|
;; |
|
|
setf.sig f15 = r19 |
|
|
setf.sig f32 = r21 |
|
|
ld2 r2 = [r29] |
|
|
;; |
|
|
xma.l f15 = f15, f14, f0 |
|
|
shladd r18 = r16, 1, r16 |
|
|
sxt2 r2 = r2 |
|
|
;; |
|
|
getf.sig r19 = f15 |
|
|
xma.l f32 = f32, f10, f0 |
|
|
shladd r18 = r18, 4, r16 |
|
|
setf.sig f15 = r24 |
|
|
;; |
|
|
getf.sig r21 = f32 |
|
|
adds r19 = 4, r19 |
|
|
xma.l f15 = f15, f13, f0 |
|
|
setf.sig f32 = r22 |
|
|
;; |
|
|
adds r21 = 4, r21 |
|
|
getf.sig r24 = f15 |
|
|
xma.l f32 = f32, f11, f0 |
|
|
dep.z r2 = r2, 8, 24 |
|
|
setf.sig f15 = r20 |
|
|
;; |
|
|
getf.sig r15 = f32 |
|
|
sub r24 = r19, r24 |
|
|
xma.l f15 = f15, f8, f0 |
|
|
setf.sig f32 = r23 |
|
|
;; |
|
|
sub r15 = r21, r15 |
|
|
getf.sig r14 = f15 |
|
|
;; |
|
|
shr r15 = r15, 3 |
|
|
shr r24 = r24, 3 |
|
|
setf.sig f15 = r19 |
|
|
sub r21 = r21, r14 |
|
|
addl r2 = 8192, r2 |
|
|
;; |
|
|
xma.l f32 = f32, f12, f15 |
|
|
shr r21 = r21, 3 |
|
|
shl r18 = r18, 5 |
|
|
;; |
|
|
getf.sig r19 = f32 |
|
|
sub r28 = r24, r21 |
|
|
setf.sig f15 = r25 |
|
|
setf.sig f32 = r17 |
|
|
;; |
|
|
shr r19 = r19, 3 |
|
|
sxt4 r2 = r2 |
|
|
xma.l f32 = f32, f9, f0 |
|
|
;; |
|
|
sub r22 = r19, r15 |
|
|
add r25 = r2, r26 |
|
|
getf.sig r17 = f32 |
|
|
;; |
|
|
add r23 = r22, r28 |
|
|
xma.l f15 = f15, f7, f0 |
|
|
adds r17 = 4, r17 |
|
|
;; |
|
|
setf.sig f32 = r23 |
|
|
getf.sig r14 = f15 |
|
|
add r18 = r17, r18 |
|
|
sub r22 = r22, r28 |
|
|
;; |
|
|
xma.l f32 = f32, f6, f0 |
|
|
shr r18 = r18, 3 |
|
|
add r19 = r19, r15 |
|
|
sub r17 = r17, r14 |
|
|
;; |
|
|
add r27 = r25, r18 |
|
|
setf.sig f15 = r22 |
|
|
getf.sig r23 = f32 |
|
|
shr r17 = r17, 3 |
|
|
sub r2 = r2, r26 |
|
|
;; |
|
|
adds r23 = 128, r23 |
|
|
add r15 = r27, r19 |
|
|
xma.l f15 = f15, f6, f0 |
|
|
ld8 r20 = [r43] |
|
|
add r26 = r2, r17 |
|
|
;; |
|
|
shr r23 = r23, 8 |
|
|
getf.sig r22 = f15 |
|
|
shr r15 = r15, 14 |
|
|
;; |
|
|
adds r22 = 128, r22 |
|
|
add r14 = r26, r23 |
|
|
shladd r15 = r15, 1, r20 |
|
|
sub r2 = r2, r17 |
|
|
;; |
|
|
shr r22 = r22, 8 |
|
|
ld2 r16 = [r15] |
|
|
shr r14 = r14, 14 |
|
|
sub r25 = r25, r18 |
|
|
;; |
|
|
st2 [r29] = r16 |
|
|
shladd r14 = r14, 1, r20 |
|
|
add r15 = r2, r22 |
|
|
;; |
|
|
ld2 r16 = [r14] |
|
|
add r24 = r24, r21 |
|
|
shr r15 = r15, 14 |
|
|
;; |
|
|
st2 [r31] = r16 |
|
|
shladd r15 = r15, 1, r20 |
|
|
add r14 = r25, r24 |
|
|
;; |
|
|
ld2 r16 = [r15] |
|
|
shr r14 = r14, 14 |
|
|
st8 [r33] = r19 |
|
|
;; |
|
|
st2 [r37] = r16 |
|
|
shladd r14 = r14, 1, r20 |
|
|
sub r15 = r25, r24 |
|
|
;; |
|
|
ld2 r17 = [r14] |
|
|
shr r15 = r15, 14 |
|
|
sub r16 = r2, r22 |
|
|
;; |
|
|
st2 [r30] = r17 |
|
|
shladd r15 = r15, 1, r20 |
|
|
st8 [r8] = r24 |
|
|
;; |
|
|
ld2 r17 = [r15] |
|
|
shr r16 = r16, 14 |
|
|
st8 [r9] = r28 |
|
|
;; |
|
|
st2 [r39] = r17 |
|
|
sub r14 = r26, r23 |
|
|
shladd r16 = r16, 1, r20 |
|
|
st8 [r44] = r27 |
|
|
;; |
|
|
ld2 r15 = [r16] |
|
|
shr r14 = r14, 14 |
|
|
;; |
|
|
st2 [r38] = r15 |
|
|
st8 [r36] = r25 |
|
|
shladd r14 = r14, 1, r20 |
|
|
sub r18 = r27, r19 |
|
|
;; |
|
|
ld2 r15 = [r14] |
|
|
st8 [r11] = r26 |
|
|
shr r18 = r18, 14 |
|
|
;; |
|
|
st2 [r41] = r15 |
|
|
st8 [r34] = r2 |
|
|
shladd r18 = r18, 1, r20 |
|
|
st8 [r3] = r23 |
|
|
;; |
|
|
ld2 r14 = [r18] |
|
|
st8 [r10] = r22 |
|
|
;; |
|
|
st2 [r40] = r14 |
|
|
.L11: |
|
|
ld8 r14 = [r42] |
|
|
;; |
|
|
adds r14 = 1, r14 |
|
|
;; |
|
|
st8 [r42] = r14 |
|
|
cmp.ge p6, p7 = 7, r14 |
|
|
(p6) br.cond.dptk .L12 |
|
|
br.ret.sptk.many b0 |
|
|
.endp idct_ia64# |
|
46 |
.align 16 |
.align 16 |
47 |
.global idct_ia64_init# |
.data_c10: |
48 |
.proc idct_ia64_init# |
.single 2.847759065022573476966272210120, 2.847759065022573476966272210120 |
49 |
|
.data_c11: |
50 |
|
.single 5.027339492125848074977056967327, 5.027339492125848074977056967327 |
51 |
|
.align 16 |
52 |
|
.data_c12: |
53 |
|
.single 0.490392640201615215289621119155, 0.490392640201615215289621119155 |
54 |
|
.data_c13: |
55 |
|
.single 0.068974844820735750627882509889, 0.068974844820735750627882509889 |
56 |
|
.align 16 |
57 |
|
.data_c14: |
58 |
|
.single 0.097545161008064124041894160655, 0.097545161008064124041894160655 |
59 |
|
.data_c15: |
60 |
|
.single 1.000000000000000000000000000000, 1.000000000000000000000000000000 |
61 |
|
|
62 |
|
.text |
63 |
|
.global idct_ia64, idct_ia64_init |
64 |
|
|
65 |
|
.align 16 |
66 |
|
.proc idct_ia64_init |
67 |
idct_ia64_init: |
idct_ia64_init: |
68 |
.prologue |
br.ret.sptk.few b0 |
69 |
addl r14 = @ltoff(iclip_ia64#), gp |
.endp |
70 |
.save ar.lc, r2 |
|
71 |
mov r2 = ar.lc |
.align 16 |
72 |
.body |
.proc idct_ia64 |
73 |
addl r15 = @gprel(iclp_ia64#), gp |
idct_ia64: |
74 |
;; |
|
75 |
ld8 r14 = [r14] |
addreg3 = r20 |
76 |
addl r16 = 255, r0 |
addreg4 = r21 |
77 |
addl r17 = -512, r0 |
addreg5 = r22 |
78 |
;; |
addreg6 = r23 |
79 |
adds r14 = 1024, r14 |
|
80 |
addl r19 = -256, r0 |
one = f30 |
81 |
addl r18 = 255, r0 |
alloc r16 = ar.pfs, 1, 71, 0, 0 |
82 |
mov ar.lc = r16 |
addl addreg1 = @gprel(.data_c0#), gp |
83 |
;; |
addl addreg2 = @gprel(.data_c2#), gp |
84 |
st8 [r15] = r14 |
;; |
85 |
mov r20 = r14 |
add addreg3 = 32, addreg1 |
86 |
.L42: |
add addreg4 = 32, addreg2 |
87 |
sxt4 r14 = r17 |
add addreg5 = 64, addreg1 |
88 |
cmp4.gt p6, p7 = r19, r17 |
add addreg6 = 64, addreg2 |
89 |
;; |
;; |
90 |
shladd r15 = r14, 1, r20 |
ldfp8 c0, c1 = [addreg1] |
91 |
;; |
ldfp8 c2, c3 = [addreg2] |
92 |
(p6) st2 [r15] = r19 |
;; |
93 |
(p6) br.cond.dptk .L26 |
ldfp8 c4, c5 = [addreg3], 16 |
94 |
cmp4.le p6, p7 = r18, r17 |
ldfp8 c6, c7 = [addreg4], 16 |
95 |
;; |
add addreg1 = 96, addreg1 |
96 |
(p6) mov r14 = r18 |
add addreg2 = 96, addreg2 |
97 |
(p7) mov r14 = r17 |
;; |
98 |
;; |
ldfp8 c8, c9 = [addreg5], 16 |
99 |
st2 [r15] = r14 |
ldfp8 c10, c11 = [addreg6], 16 |
100 |
.L26: |
;; |
101 |
adds r15 = 1, r17 |
ldfp8 c12, c13 = [addreg1] |
102 |
;; |
ldfp8 c14, c15 = [addreg2] |
103 |
sxt4 r14 = r15 |
;; |
104 |
cmp4.gt p6, p7 = r19, r15 |
mov addreg1 = in0 |
105 |
;; |
fpack one = f1, f1 |
106 |
shladd r16 = r14, 1, r20 |
add addreg2 = 2, in0 |
107 |
;; |
;; |
108 |
(p6) st2 [r16] = r19 |
|
109 |
(p6) br.cond.dptk .L30 |
ld2 r33 = [addreg1], 4 |
110 |
cmp4.le p6, p7 = r18, r15 |
ld2 r34 = [addreg2], 4 |
111 |
;; |
;; |
112 |
(p6) mov r14 = r18 |
ld2 r35 = [addreg1], 4 |
113 |
(p7) mov r14 = r15 |
ld2 r36 = [addreg2], 4 |
114 |
;; |
;; |
115 |
st2 [r16] = r14 |
ld2 r37 = [addreg1], 4 |
116 |
.L30: |
ld2 r38 = [addreg2], 4 |
117 |
adds r15 = 2, r17 |
;; |
118 |
;; |
ld2 r39 = [addreg1], 4 |
119 |
sxt4 r14 = r15 |
ld2 r40 = [addreg2], 4 |
120 |
cmp4.gt p6, p7 = r19, r15 |
;; |
121 |
;; |
ld2 r41 = [addreg1], 4 |
122 |
shladd r16 = r14, 1, r20 |
ld2 r42 = [addreg2], 4 |
123 |
;; |
;; |
124 |
(p6) st2 [r16] = r19 |
ld2 r43 = [addreg1], 4 |
125 |
(p6) br.cond.dptk .L34 |
ld2 r44 = [addreg2], 4 |
126 |
cmp4.le p6, p7 = r18, r15 |
;; |
127 |
;; |
ld2 r45 = [addreg1], 4 |
128 |
(p6) mov r14 = r18 |
ld2 r46 = [addreg2], 4 |
129 |
(p7) mov r14 = r15 |
;; |
130 |
;; |
ld2 r47 = [addreg1], 4 |
131 |
st2 [r16] = r14 |
ld2 r48 = [addreg2], 4 |
132 |
.L34: |
;; |
133 |
adds r15 = 3, r17 |
ld2 r49 = [addreg1], 4 |
134 |
;; |
ld2 r50 = [addreg2], 4 |
135 |
sxt4 r14 = r15 |
;; |
136 |
cmp4.gt p6, p7 = r19, r15 |
ld2 r51 = [addreg1], 4 |
137 |
;; |
ld2 r52 = [addreg2], 4 |
138 |
shladd r16 = r14, 1, r20 |
;; |
139 |
;; |
ld2 r53 = [addreg1], 4 |
140 |
(p6) st2 [r16] = r19 |
ld2 r54 = [addreg2], 4 |
141 |
(p6) br.cond.dptk .L38 |
;; |
142 |
cmp4.le p6, p7 = r18, r15 |
ld2 r55 = [addreg1], 4 |
143 |
;; |
ld2 r56 = [addreg2], 4 |
144 |
(p6) mov r14 = r18 |
;; |
145 |
(p7) mov r14 = r15 |
ld2 r57 = [addreg1], 4 |
146 |
;; |
ld2 r58 = [addreg2], 4 |
147 |
st2 [r16] = r14 |
;; |
148 |
.L38: |
ld2 r59 = [addreg1], 4 |
149 |
adds r17 = 4, r17 |
ld2 r60 = [addreg2], 4 |
150 |
br.cloop.sptk.few .L42 |
;; |
151 |
;; |
ld2 r61 = [addreg1], 4 |
152 |
mov ar.lc = r2 |
ld2 r62 = [addreg2], 4 |
153 |
br.ret.sptk.many b0 |
;; |
154 |
.endp idct_ia64_init# |
ld2 r63 = [addreg1], 4 |
155 |
.common idct#,8,8 |
ld2 r64 = [addreg2], 4 |
156 |
.bss |
;; |
157 |
.align 2 |
ld2 r65 = [addreg1], 4 |
158 |
.type iclip_ia64#,@object |
ld2 r66 = [addreg2], 4 |
159 |
.size iclip_ia64#,2048 |
;; |
160 |
iclip_ia64: |
ld2 r67 = [addreg1], 4 |
161 |
.skip 2048 |
ld2 r68 = [addreg2], 4 |
162 |
.sbss |
;; |
163 |
.align 8 |
ld2 r69 = [addreg1], 4 |
164 |
.type iclp_ia64#,@object |
ld2 r70 = [addreg2], 4 |
165 |
.size iclp_ia64#,8 |
;; |
166 |
iclp_ia64: |
ld2 r71 = [addreg1], 4 |
167 |
.skip 8 |
ld2 r72 = [addreg2], 4 |
168 |
.ident "GCC: (GNU) 2.96 20000731 (Red Hat Linux 7.1 2.96-85)" |
;; |
169 |
|
ld2 r73 = [addreg1], 4 |
170 |
|
ld2 r74 = [addreg2], 4 |
171 |
|
;; |
172 |
|
ld2 r75 = [addreg1], 4 |
173 |
|
ld2 r76 = [addreg2], 4 |
174 |
|
;; |
175 |
|
ld2 r77 = [addreg1], 4 |
176 |
|
ld2 r78 = [addreg2], 4 |
177 |
|
;; |
178 |
|
ld2 r79 = [addreg1], 4 |
179 |
|
ld2 r80 = [addreg2], 4 |
180 |
|
;; |
181 |
|
ld2 r81 = [addreg1], 4 |
182 |
|
ld2 r82 = [addreg2], 4 |
183 |
|
;; |
184 |
|
ld2 r83 = [addreg1], 4 |
185 |
|
ld2 r84 = [addreg2], 4 |
186 |
|
;; |
187 |
|
ld2 r85 = [addreg1], 4 |
188 |
|
ld2 r86 = [addreg2], 4 |
189 |
|
;; |
190 |
|
ld2 r87 = [addreg1], 4 |
191 |
|
ld2 r88 = [addreg2], 4 |
192 |
|
;; |
193 |
|
ld2 r89 = [addreg1], 4 |
194 |
|
ld2 r90 = [addreg2], 4 |
195 |
|
;; |
196 |
|
ld2 r91 = [addreg1], 4 |
197 |
|
ld2 r92 = [addreg2], 4 |
198 |
|
;; |
199 |
|
ld2 r93 = [addreg1], 4 |
200 |
|
ld2 r94 = [addreg2], 4 |
201 |
|
;; |
202 |
|
ld2 r95 = [addreg1], 4 |
203 |
|
ld2 r96 = [addreg2], 4 |
204 |
|
;; |
205 |
|
sxt2 r33 = r33 |
206 |
|
sxt2 r34 = r34 |
207 |
|
sxt2 r35 = r35 |
208 |
|
sxt2 r36 = r36 |
209 |
|
sxt2 r37 = r37 |
210 |
|
sxt2 r38 = r38 |
211 |
|
sxt2 r39 = r39 |
212 |
|
sxt2 r40 = r40 |
213 |
|
sxt2 r41 = r41 |
214 |
|
sxt2 r42 = r42 |
215 |
|
sxt2 r43 = r43 |
216 |
|
sxt2 r44 = r44 |
217 |
|
sxt2 r45 = r45 |
218 |
|
sxt2 r46 = r46 |
219 |
|
sxt2 r47 = r47 |
220 |
|
sxt2 r48 = r48 |
221 |
|
sxt2 r49 = r49 |
222 |
|
sxt2 r50 = r50 |
223 |
|
sxt2 r51 = r51 |
224 |
|
sxt2 r52 = r52 |
225 |
|
sxt2 r53 = r53 |
226 |
|
sxt2 r54 = r54 |
227 |
|
sxt2 r55 = r55 |
228 |
|
sxt2 r56 = r56 |
229 |
|
sxt2 r57 = r57 |
230 |
|
sxt2 r58 = r58 |
231 |
|
sxt2 r59 = r59 |
232 |
|
sxt2 r60 = r60 |
233 |
|
sxt2 r61 = r61 |
234 |
|
sxt2 r62 = r62 |
235 |
|
sxt2 r63 = r63 |
236 |
|
sxt2 r64 = r64 |
237 |
|
sxt2 r65 = r65 |
238 |
|
sxt2 r66 = r66 |
239 |
|
sxt2 r67 = r67 |
240 |
|
sxt2 r68 = r68 |
241 |
|
sxt2 r69 = r69 |
242 |
|
sxt2 r70 = r70 |
243 |
|
sxt2 r71 = r71 |
244 |
|
sxt2 r72 = r72 |
245 |
|
sxt2 r73 = r73 |
246 |
|
sxt2 r74 = r74 |
247 |
|
sxt2 r75 = r75 |
248 |
|
sxt2 r76 = r76 |
249 |
|
sxt2 r77 = r77 |
250 |
|
sxt2 r78 = r78 |
251 |
|
sxt2 r79 = r79 |
252 |
|
sxt2 r80 = r80 |
253 |
|
sxt2 r81 = r81 |
254 |
|
sxt2 r82 = r82 |
255 |
|
sxt2 r83 = r83 |
256 |
|
sxt2 r84 = r84 |
257 |
|
sxt2 r85 = r85 |
258 |
|
sxt2 r86 = r86 |
259 |
|
sxt2 r87 = r87 |
260 |
|
sxt2 r88 = r88 |
261 |
|
sxt2 r89 = r89 |
262 |
|
sxt2 r90 = r90 |
263 |
|
sxt2 r91 = r91 |
264 |
|
sxt2 r92 = r92 |
265 |
|
sxt2 r93 = r93 |
266 |
|
sxt2 r94 = r94 |
267 |
|
sxt2 r95 = r95 |
268 |
|
sxt2 r96 = r96 |
269 |
|
;; |
270 |
|
setf.sig f48 = r33 |
271 |
|
setf.sig f49 = r34 |
272 |
|
setf.sig f50 = r35 |
273 |
|
setf.sig f51 = r36 |
274 |
|
setf.sig f52 = r37 |
275 |
|
setf.sig f53 = r38 |
276 |
|
setf.sig f54 = r39 |
277 |
|
setf.sig f55 = r40 |
278 |
|
setf.sig f56 = r41 |
279 |
|
setf.sig f57 = r42 |
280 |
|
setf.sig f58 = r43 |
281 |
|
setf.sig f59 = r44 |
282 |
|
setf.sig f60 = r45 |
283 |
|
setf.sig f61 = r46 |
284 |
|
setf.sig f62 = r47 |
285 |
|
setf.sig f63 = r48 |
286 |
|
setf.sig f64 = r49 |
287 |
|
setf.sig f65 = r50 |
288 |
|
setf.sig f66 = r51 |
289 |
|
setf.sig f67 = r52 |
290 |
|
setf.sig f68 = r53 |
291 |
|
setf.sig f69 = r54 |
292 |
|
setf.sig f70 = r55 |
293 |
|
setf.sig f71 = r56 |
294 |
|
setf.sig f72 = r57 |
295 |
|
setf.sig f73 = r58 |
296 |
|
setf.sig f74 = r59 |
297 |
|
setf.sig f75 = r60 |
298 |
|
setf.sig f76 = r61 |
299 |
|
setf.sig f77 = r62 |
300 |
|
setf.sig f78 = r63 |
301 |
|
setf.sig f79 = r64 |
302 |
|
setf.sig f80 = r65 |
303 |
|
setf.sig f81 = r66 |
304 |
|
setf.sig f82 = r67 |
305 |
|
setf.sig f83 = r68 |
306 |
|
setf.sig f84 = r69 |
307 |
|
setf.sig f85 = r70 |
308 |
|
setf.sig f86 = r71 |
309 |
|
setf.sig f87 = r72 |
310 |
|
setf.sig f88 = r73 |
311 |
|
setf.sig f89 = r74 |
312 |
|
setf.sig f90 = r75 |
313 |
|
setf.sig f91 = r76 |
314 |
|
setf.sig f92 = r77 |
315 |
|
setf.sig f93 = r78 |
316 |
|
setf.sig f94 = r79 |
317 |
|
setf.sig f95 = r80 |
318 |
|
setf.sig f96 = r81 |
319 |
|
setf.sig f97 = r82 |
320 |
|
setf.sig f98 = r83 |
321 |
|
setf.sig f99 = r84 |
322 |
|
setf.sig f100 = r85 |
323 |
|
setf.sig f101 = r86 |
324 |
|
setf.sig f102 = r87 |
325 |
|
setf.sig f103 = r88 |
326 |
|
setf.sig f104 = r89 |
327 |
|
setf.sig f105 = r90 |
328 |
|
setf.sig f106 = r91 |
329 |
|
setf.sig f107 = r92 |
330 |
|
setf.sig f108 = r93 |
331 |
|
setf.sig f109 = r94 |
332 |
|
setf.sig f110 = r95 |
333 |
|
setf.sig f111 = r96 |
334 |
|
;; |
335 |
|
fcvt.xf f48 = f48 |
336 |
|
fcvt.xf f49 = f49 |
337 |
|
fcvt.xf f50 = f50 |
338 |
|
fcvt.xf f51 = f51 |
339 |
|
fcvt.xf f52 = f52 |
340 |
|
fcvt.xf f53 = f53 |
341 |
|
fcvt.xf f54 = f54 |
342 |
|
fcvt.xf f55 = f55 |
343 |
|
fcvt.xf f56 = f56 |
344 |
|
fcvt.xf f57 = f57 |
345 |
|
fcvt.xf f58 = f58 |
346 |
|
fcvt.xf f59 = f59 |
347 |
|
fcvt.xf f60 = f60 |
348 |
|
fcvt.xf f61 = f61 |
349 |
|
fcvt.xf f62 = f62 |
350 |
|
fcvt.xf f63 = f63 |
351 |
|
fcvt.xf f64 = f64 |
352 |
|
fcvt.xf f65 = f65 |
353 |
|
fcvt.xf f66 = f66 |
354 |
|
fcvt.xf f67 = f67 |
355 |
|
fcvt.xf f68 = f68 |
356 |
|
fcvt.xf f69 = f69 |
357 |
|
fcvt.xf f70 = f70 |
358 |
|
fcvt.xf f71 = f71 |
359 |
|
fcvt.xf f72 = f72 |
360 |
|
fcvt.xf f73 = f73 |
361 |
|
fcvt.xf f74 = f74 |
362 |
|
fcvt.xf f75 = f75 |
363 |
|
fcvt.xf f76 = f76 |
364 |
|
fcvt.xf f77 = f77 |
365 |
|
fcvt.xf f78 = f78 |
366 |
|
fcvt.xf f79 = f79 |
367 |
|
fcvt.xf f80 = f80 |
368 |
|
fcvt.xf f81 = f81 |
369 |
|
fcvt.xf f82 = f82 |
370 |
|
fcvt.xf f83 = f83 |
371 |
|
fcvt.xf f84 = f84 |
372 |
|
fcvt.xf f85 = f85 |
373 |
|
fcvt.xf f86 = f86 |
374 |
|
fcvt.xf f87 = f87 |
375 |
|
fcvt.xf f88 = f88 |
376 |
|
fcvt.xf f89 = f89 |
377 |
|
fcvt.xf f90 = f90 |
378 |
|
fcvt.xf f91 = f91 |
379 |
|
fcvt.xf f92 = f92 |
380 |
|
fcvt.xf f93 = f93 |
381 |
|
fcvt.xf f94 = f94 |
382 |
|
fcvt.xf f95 = f95 |
383 |
|
fcvt.xf f96 = f96 |
384 |
|
fcvt.xf f97 = f97 |
385 |
|
fcvt.xf f98 = f98 |
386 |
|
fcvt.xf f99 = f99 |
387 |
|
fcvt.xf f100 = f100 |
388 |
|
fcvt.xf f101 = f101 |
389 |
|
fcvt.xf f102 = f102 |
390 |
|
fcvt.xf f103 = f103 |
391 |
|
fcvt.xf f104 = f104 |
392 |
|
fcvt.xf f105 = f105 |
393 |
|
fcvt.xf f106 = f106 |
394 |
|
fcvt.xf f107 = f107 |
395 |
|
fcvt.xf f108 = f108 |
396 |
|
fcvt.xf f109 = f109 |
397 |
|
fcvt.xf f110 = f110 |
398 |
|
fcvt.xf f111 = f111 |
399 |
|
;; |
400 |
|
fpack f48 = f48, f49 |
401 |
|
;; |
402 |
|
fpack f49 = f50, f51 |
403 |
|
;; |
404 |
|
fpack f50 = f52, f53 |
405 |
|
;; |
406 |
|
fpack f51 = f54, f55 |
407 |
|
;; |
408 |
|
fpack f52 = f56, f57 |
409 |
|
;; |
410 |
|
fpack f53 = f58, f59 |
411 |
|
;; |
412 |
|
fpack f54 = f60, f61 |
413 |
|
;; |
414 |
|
fpack f55 = f62, f63 |
415 |
|
;; |
416 |
|
fpack f56 = f64, f65 |
417 |
|
;; |
418 |
|
fpack f57 = f66, f67 |
419 |
|
;; |
420 |
|
fpack f58 = f68, f69 |
421 |
|
;; |
422 |
|
fpack f59 = f70, f71 |
423 |
|
;; |
424 |
|
fpack f60 = f72, f73 |
425 |
|
;; |
426 |
|
fpack f61 = f74, f75 |
427 |
|
;; |
428 |
|
fpack f62 = f76, f77 |
429 |
|
;; |
430 |
|
fpack f63 = f78, f79 |
431 |
|
;; |
432 |
|
fpack f64 = f80, f81 |
433 |
|
;; |
434 |
|
fpack f65 = f82, f83 |
435 |
|
;; |
436 |
|
fpack f66 = f84, f85 |
437 |
|
;; |
438 |
|
fpack f67 = f86, f87 |
439 |
|
;; |
440 |
|
fpack f68 = f88, f89 |
441 |
|
;; |
442 |
|
fpack f69 = f90, f91 |
443 |
|
;; |
444 |
|
fpack f70 = f92, f93 |
445 |
|
;; |
446 |
|
fpack f71 = f94, f95 |
447 |
|
;; |
448 |
|
fpack f72 = f96, f97 |
449 |
|
;; |
450 |
|
fpack f73 = f98, f99 |
451 |
|
;; |
452 |
|
fpack f74 = f100, f101 |
453 |
|
;; |
454 |
|
fpack f75 = f102, f103 |
455 |
|
;; |
456 |
|
fpack f76 = f104, f105 |
457 |
|
;; |
458 |
|
fpack f77 = f106, f107 |
459 |
|
;; |
460 |
|
fpack f78 = f108, f109 |
461 |
|
;; |
462 |
|
fpack f79 = f110, f111 |
463 |
|
;; |
464 |
|
fpma f48 = f48, c0, f0 |
465 |
|
fpma f49 = f49, c0, f0 |
466 |
|
fpma f50 = f50, c0, f0 |
467 |
|
fpma f51 = f51, c0, f0 |
468 |
|
;; |
469 |
|
|
470 |
|
// before pre shuffle |
471 |
|
// 48 49 50 51 |
472 |
|
// 52 53 54 55 |
473 |
|
// 56 57 58 59 |
474 |
|
// 60 61 62 63 |
475 |
|
// 64 65 66 67 |
476 |
|
// 68 69 70 71 |
477 |
|
// 72 73 74 75 |
478 |
|
// 76 77 78 79 |
479 |
|
|
480 |
|
// after pre shuffle |
481 |
|
// 48 49 50 51 |
482 |
|
// 64 53 54 55 |
483 |
|
// 56 57 58 59 |
484 |
|
// 72 61 62 63 |
485 |
|
// 52 65 66 67 |
486 |
|
// 76 69 70 71 |
487 |
|
// 60 73 74 75 |
488 |
|
// 68 77 78 79 |
489 |
|
// (f80, f64) = (f48, f64) $ (c0, c0), (line 0, 1) |
490 |
|
fpma f80 = f64, c0, f48 |
491 |
|
fpnma f64 = f64, c0, f48 |
492 |
|
;; |
493 |
|
// (f48, f72) = (f56, f72) $ (c1, c2), (line 2, 3) |
494 |
|
fpma f48 = f72, c1, f56 |
495 |
|
fpnma f72 = f72, c2, f56 |
496 |
|
;; |
497 |
|
// (f56, f76) = (f52, f76) $ (c3, c4), (line 4, 5) |
498 |
|
fpma f56 = f76, c3, f52 |
499 |
|
fpnma f76 = f76, c4, f52 |
500 |
|
;; |
501 |
|
// (f52, f68) = (f60, f68) $ (c5, c6), (line 6, 7) |
502 |
|
fpma f52 = f68, c5, f60 |
503 |
|
fpnma f68 = f68, c6, f60 |
504 |
|
;; |
505 |
|
;; |
506 |
|
// (f60, f72) = (f80, f72) $ (c7, c7), (line 0, 3) |
507 |
|
fpma f60 = f72, c7, f80 |
508 |
|
fpnma f72 = f72, c7, f80 |
509 |
|
;; |
510 |
|
// (f80, f48) = (f64, f48) $ (c8, c8), (line 1, 2) |
511 |
|
fpma f80 = f48, c8, f64 |
512 |
|
fpnma f48 = f48, c8, f64 |
513 |
|
;; |
514 |
|
// (f64, f52) = (f56, f52) $ (c9, c9), (line 4, 6) |
515 |
|
fpma f64 = f52, c9, f56 |
516 |
|
fpnma f52 = f52, c9, f56 |
517 |
|
;; |
518 |
|
// (f56, f68) = (f76, f68) $ (c10, c10), (line 5, 7) |
519 |
|
fpma f56 = f68, c10, f76 |
520 |
|
fpnma f68 = f68, c10, f76 |
521 |
|
;; |
522 |
|
;; |
523 |
|
// (f76, f52) = (f56, f52) $ (c11, c11), (line 5, 6) |
524 |
|
fpma f76 = f52, c11, f56 |
525 |
|
fpnma f52 = f52, c11, f56 |
526 |
|
;; |
527 |
|
// (f56, f64) = (f60, f64) $ (c12, c12), (line 0, 4) |
528 |
|
fpma f56 = f64, c12, f60 |
529 |
|
fpnma f64 = f64, c12, f60 |
530 |
|
;; |
531 |
|
// (f60, f68) = (f72, f68) $ (c14, c14), (line 3, 7) |
532 |
|
fpma f60 = f68, c14, f72 |
533 |
|
fpnma f68 = f68, c14, f72 |
534 |
|
;; |
535 |
|
;; |
536 |
|
// (f72, f76) = (f80, f76) $ (c13, c13), (line 1, 5) |
537 |
|
fpma f72 = f76, c13, f80 |
538 |
|
fpnma f76 = f76, c13, f80 |
539 |
|
;; |
540 |
|
// (f80, f52) = (f48, f52) $ (c13, c13), (line 2, 6) |
541 |
|
fpma f80 = f52, c13, f48 |
542 |
|
fpnma f52 = f52, c13, f48 |
543 |
|
;; |
544 |
|
|
545 |
|
// before post shuffle |
546 |
|
// 56 49 50 51 |
547 |
|
// 72 53 54 55 |
548 |
|
// 80 57 58 59 |
549 |
|
// 60 61 62 63 |
550 |
|
// 64 65 66 67 |
551 |
|
// 76 69 70 71 |
552 |
|
// 52 73 74 75 |
553 |
|
// 68 77 78 79 |
554 |
|
|
555 |
|
// after post shuffle |
556 |
|
// 56 49 50 51 |
557 |
|
// 72 53 54 55 |
558 |
|
// 52 57 58 59 |
559 |
|
// 60 61 62 63 |
560 |
|
// 68 65 66 67 |
561 |
|
// 80 69 70 71 |
562 |
|
// 76 73 74 75 |
563 |
|
// 64 77 78 79 |
564 |
|
|
565 |
|
// before pre shuffle |
566 |
|
// 56 49 50 51 |
567 |
|
// 72 53 54 55 |
568 |
|
// 52 57 58 59 |
569 |
|
// 60 61 62 63 |
570 |
|
// 68 65 66 67 |
571 |
|
// 80 69 70 71 |
572 |
|
// 76 73 74 75 |
573 |
|
// 64 77 78 79 |
574 |
|
|
575 |
|
// after pre shuffle |
576 |
|
// 56 49 50 51 |
577 |
|
// 72 65 54 55 |
578 |
|
// 52 57 58 59 |
579 |
|
// 60 73 62 63 |
580 |
|
// 68 53 66 67 |
581 |
|
// 80 77 70 71 |
582 |
|
// 76 61 74 75 |
583 |
|
// 64 69 78 79 |
584 |
|
// (f48, f65) = (f49, f65) $ (c0, c0), (line 0, 1) |
585 |
|
fpma f48 = f65, c0, f49 |
586 |
|
fpnma f65 = f65, c0, f49 |
587 |
|
;; |
588 |
|
// (f49, f73) = (f57, f73) $ (c1, c2), (line 2, 3) |
589 |
|
fpma f49 = f73, c1, f57 |
590 |
|
fpnma f73 = f73, c2, f57 |
591 |
|
;; |
592 |
|
// (f57, f77) = (f53, f77) $ (c3, c4), (line 4, 5) |
593 |
|
fpma f57 = f77, c3, f53 |
594 |
|
fpnma f77 = f77, c4, f53 |
595 |
|
;; |
596 |
|
// (f53, f69) = (f61, f69) $ (c5, c6), (line 6, 7) |
597 |
|
fpma f53 = f69, c5, f61 |
598 |
|
fpnma f69 = f69, c6, f61 |
599 |
|
;; |
600 |
|
;; |
601 |
|
// (f61, f73) = (f48, f73) $ (c7, c7), (line 0, 3) |
602 |
|
fpma f61 = f73, c7, f48 |
603 |
|
fpnma f73 = f73, c7, f48 |
604 |
|
;; |
605 |
|
// (f48, f49) = (f65, f49) $ (c8, c8), (line 1, 2) |
606 |
|
fpma f48 = f49, c8, f65 |
607 |
|
fpnma f49 = f49, c8, f65 |
608 |
|
;; |
609 |
|
// (f65, f53) = (f57, f53) $ (c9, c9), (line 4, 6) |
610 |
|
fpma f65 = f53, c9, f57 |
611 |
|
fpnma f53 = f53, c9, f57 |
612 |
|
;; |
613 |
|
// (f57, f69) = (f77, f69) $ (c10, c10), (line 5, 7) |
614 |
|
fpma f57 = f69, c10, f77 |
615 |
|
fpnma f69 = f69, c10, f77 |
616 |
|
;; |
617 |
|
;; |
618 |
|
// (f77, f53) = (f57, f53) $ (c11, c11), (line 5, 6) |
619 |
|
fpma f77 = f53, c11, f57 |
620 |
|
fpnma f53 = f53, c11, f57 |
621 |
|
;; |
622 |
|
// (f57, f65) = (f61, f65) $ (c12, c12), (line 0, 4) |
623 |
|
fpma f57 = f65, c12, f61 |
624 |
|
fpnma f65 = f65, c12, f61 |
625 |
|
;; |
626 |
|
// (f61, f69) = (f73, f69) $ (c14, c14), (line 3, 7) |
627 |
|
fpma f61 = f69, c14, f73 |
628 |
|
fpnma f69 = f69, c14, f73 |
629 |
|
;; |
630 |
|
;; |
631 |
|
// (f73, f77) = (f48, f77) $ (c13, c13), (line 1, 5) |
632 |
|
fpma f73 = f77, c13, f48 |
633 |
|
fpnma f77 = f77, c13, f48 |
634 |
|
;; |
635 |
|
// (f48, f53) = (f49, f53) $ (c13, c13), (line 2, 6) |
636 |
|
fpma f48 = f53, c13, f49 |
637 |
|
fpnma f53 = f53, c13, f49 |
638 |
|
;; |
639 |
|
|
640 |
|
// before post shuffle |
641 |
|
// 56 57 50 51 |
642 |
|
// 72 73 54 55 |
643 |
|
// 52 48 58 59 |
644 |
|
// 60 61 62 63 |
645 |
|
// 68 65 66 67 |
646 |
|
// 80 77 70 71 |
647 |
|
// 76 53 74 75 |
648 |
|
// 64 69 78 79 |
649 |
|
|
650 |
|
// after post shuffle |
651 |
|
// 56 57 50 51 |
652 |
|
// 72 73 54 55 |
653 |
|
// 52 53 58 59 |
654 |
|
// 60 61 62 63 |
655 |
|
// 68 69 66 67 |
656 |
|
// 80 48 70 71 |
657 |
|
// 76 77 74 75 |
658 |
|
// 64 65 78 79 |
659 |
|
|
660 |
|
// before pre shuffle |
661 |
|
// 56 57 50 51 |
662 |
|
// 72 73 54 55 |
663 |
|
// 52 53 58 59 |
664 |
|
// 60 61 62 63 |
665 |
|
// 68 69 66 67 |
666 |
|
// 80 48 70 71 |
667 |
|
// 76 77 74 75 |
668 |
|
// 64 65 78 79 |
669 |
|
|
670 |
|
// after pre shuffle |
671 |
|
// 56 57 50 51 |
672 |
|
// 72 73 66 55 |
673 |
|
// 52 53 58 59 |
674 |
|
// 60 61 74 63 |
675 |
|
// 68 69 54 67 |
676 |
|
// 80 48 78 71 |
677 |
|
// 76 77 62 75 |
678 |
|
// 64 65 70 79 |
679 |
|
// (f49, f66) = (f50, f66) $ (c0, c0), (line 0, 1) |
680 |
|
fpma f49 = f66, c0, f50 |
681 |
|
fpnma f66 = f66, c0, f50 |
682 |
|
;; |
683 |
|
// (f50, f74) = (f58, f74) $ (c1, c2), (line 2, 3) |
684 |
|
fpma f50 = f74, c1, f58 |
685 |
|
fpnma f74 = f74, c2, f58 |
686 |
|
;; |
687 |
|
// (f58, f78) = (f54, f78) $ (c3, c4), (line 4, 5) |
688 |
|
fpma f58 = f78, c3, f54 |
689 |
|
fpnma f78 = f78, c4, f54 |
690 |
|
;; |
691 |
|
// (f54, f70) = (f62, f70) $ (c5, c6), (line 6, 7) |
692 |
|
fpma f54 = f70, c5, f62 |
693 |
|
fpnma f70 = f70, c6, f62 |
694 |
|
;; |
695 |
|
;; |
696 |
|
// (f62, f74) = (f49, f74) $ (c7, c7), (line 0, 3) |
697 |
|
fpma f62 = f74, c7, f49 |
698 |
|
fpnma f74 = f74, c7, f49 |
699 |
|
;; |
700 |
|
// (f49, f50) = (f66, f50) $ (c8, c8), (line 1, 2) |
701 |
|
fpma f49 = f50, c8, f66 |
702 |
|
fpnma f50 = f50, c8, f66 |
703 |
|
;; |
704 |
|
// (f66, f54) = (f58, f54) $ (c9, c9), (line 4, 6) |
705 |
|
fpma f66 = f54, c9, f58 |
706 |
|
fpnma f54 = f54, c9, f58 |
707 |
|
;; |
708 |
|
// (f58, f70) = (f78, f70) $ (c10, c10), (line 5, 7) |
709 |
|
fpma f58 = f70, c10, f78 |
710 |
|
fpnma f70 = f70, c10, f78 |
711 |
|
;; |
712 |
|
;; |
713 |
|
// (f78, f54) = (f58, f54) $ (c11, c11), (line 5, 6) |
714 |
|
fpma f78 = f54, c11, f58 |
715 |
|
fpnma f54 = f54, c11, f58 |
716 |
|
;; |
717 |
|
// (f58, f66) = (f62, f66) $ (c12, c12), (line 0, 4) |
718 |
|
fpma f58 = f66, c12, f62 |
719 |
|
fpnma f66 = f66, c12, f62 |
720 |
|
;; |
721 |
|
// (f62, f70) = (f74, f70) $ (c14, c14), (line 3, 7) |
722 |
|
fpma f62 = f70, c14, f74 |
723 |
|
fpnma f70 = f70, c14, f74 |
724 |
|
;; |
725 |
|
;; |
726 |
|
// (f74, f78) = (f49, f78) $ (c13, c13), (line 1, 5) |
727 |
|
fpma f74 = f78, c13, f49 |
728 |
|
fpnma f78 = f78, c13, f49 |
729 |
|
;; |
730 |
|
// (f49, f54) = (f50, f54) $ (c13, c13), (line 2, 6) |
731 |
|
fpma f49 = f54, c13, f50 |
732 |
|
fpnma f54 = f54, c13, f50 |
733 |
|
;; |
734 |
|
|
735 |
|
// before post shuffle |
736 |
|
// 56 57 58 51 |
737 |
|
// 72 73 74 55 |
738 |
|
// 52 53 49 59 |
739 |
|
// 60 61 62 63 |
740 |
|
// 68 69 66 67 |
741 |
|
// 80 48 78 71 |
742 |
|
// 76 77 54 75 |
743 |
|
// 64 65 70 79 |
744 |
|
|
745 |
|
// after post shuffle |
746 |
|
// 56 57 58 51 |
747 |
|
// 72 73 74 55 |
748 |
|
// 52 53 54 59 |
749 |
|
// 60 61 62 63 |
750 |
|
// 68 69 70 67 |
751 |
|
// 80 48 49 71 |
752 |
|
// 76 77 78 75 |
753 |
|
// 64 65 66 79 |
754 |
|
|
755 |
|
// before pre shuffle |
756 |
|
// 56 57 58 51 |
757 |
|
// 72 73 74 55 |
758 |
|
// 52 53 54 59 |
759 |
|
// 60 61 62 63 |
760 |
|
// 68 69 70 67 |
761 |
|
// 80 48 49 71 |
762 |
|
// 76 77 78 75 |
763 |
|
// 64 65 66 79 |
764 |
|
|
765 |
|
// after pre shuffle |
766 |
|
// 56 57 58 51 |
767 |
|
// 72 73 74 67 |
768 |
|
// 52 53 54 59 |
769 |
|
// 60 61 62 75 |
770 |
|
// 68 69 70 55 |
771 |
|
// 80 48 49 79 |
772 |
|
// 76 77 78 63 |
773 |
|
// 64 65 66 71 |
774 |
|
// (f50, f67) = (f51, f67) $ (c0, c0), (line 0, 1) |
775 |
|
fpma f50 = f67, c0, f51 |
776 |
|
fpnma f67 = f67, c0, f51 |
777 |
|
;; |
778 |
|
// (f51, f75) = (f59, f75) $ (c1, c2), (line 2, 3) |
779 |
|
fpma f51 = f75, c1, f59 |
780 |
|
fpnma f75 = f75, c2, f59 |
781 |
|
;; |
782 |
|
// (f59, f79) = (f55, f79) $ (c3, c4), (line 4, 5) |
783 |
|
fpma f59 = f79, c3, f55 |
784 |
|
fpnma f79 = f79, c4, f55 |
785 |
|
;; |
786 |
|
// (f55, f71) = (f63, f71) $ (c5, c6), (line 6, 7) |
787 |
|
fpma f55 = f71, c5, f63 |
788 |
|
fpnma f71 = f71, c6, f63 |
789 |
|
;; |
790 |
|
;; |
791 |
|
// (f63, f75) = (f50, f75) $ (c7, c7), (line 0, 3) |
792 |
|
fpma f63 = f75, c7, f50 |
793 |
|
fpnma f75 = f75, c7, f50 |
794 |
|
;; |
795 |
|
// (f50, f51) = (f67, f51) $ (c8, c8), (line 1, 2) |
796 |
|
fpma f50 = f51, c8, f67 |
797 |
|
fpnma f51 = f51, c8, f67 |
798 |
|
;; |
799 |
|
// (f67, f55) = (f59, f55) $ (c9, c9), (line 4, 6) |
800 |
|
fpma f67 = f55, c9, f59 |
801 |
|
fpnma f55 = f55, c9, f59 |
802 |
|
;; |
803 |
|
// (f59, f71) = (f79, f71) $ (c10, c10), (line 5, 7) |
804 |
|
fpma f59 = f71, c10, f79 |
805 |
|
fpnma f71 = f71, c10, f79 |
806 |
|
;; |
807 |
|
;; |
808 |
|
// (f79, f55) = (f59, f55) $ (c11, c11), (line 5, 6) |
809 |
|
fpma f79 = f55, c11, f59 |
810 |
|
fpnma f55 = f55, c11, f59 |
811 |
|
;; |
812 |
|
// (f59, f67) = (f63, f67) $ (c12, c12), (line 0, 4) |
813 |
|
fpma f59 = f67, c12, f63 |
814 |
|
fpnma f67 = f67, c12, f63 |
815 |
|
;; |
816 |
|
// (f63, f71) = (f75, f71) $ (c14, c14), (line 3, 7) |
817 |
|
fpma f63 = f71, c14, f75 |
818 |
|
fpnma f71 = f71, c14, f75 |
819 |
|
;; |
820 |
|
;; |
821 |
|
// (f75, f79) = (f50, f79) $ (c13, c13), (line 1, 5) |
822 |
|
fpma f75 = f79, c13, f50 |
823 |
|
fpnma f79 = f79, c13, f50 |
824 |
|
;; |
825 |
|
// (f50, f55) = (f51, f55) $ (c13, c13), (line 2, 6) |
826 |
|
fpma f50 = f55, c13, f51 |
827 |
|
fpnma f55 = f55, c13, f51 |
828 |
|
;; |
829 |
|
|
830 |
|
// before post shuffle |
831 |
|
// 56 57 58 59 |
832 |
|
// 72 73 74 75 |
833 |
|
// 52 53 54 50 |
834 |
|
// 60 61 62 63 |
835 |
|
// 68 69 70 67 |
836 |
|
// 80 48 49 79 |
837 |
|
// 76 77 78 55 |
838 |
|
// 64 65 66 71 |
839 |
|
|
840 |
|
// after post shuffle |
841 |
|
// 56 57 58 59 |
842 |
|
// 72 73 74 75 |
843 |
|
// 52 53 54 55 |
844 |
|
// 60 61 62 63 |
845 |
|
// 68 69 70 71 |
846 |
|
// 80 48 49 50 |
847 |
|
// 76 77 78 79 |
848 |
|
// 64 65 66 67 |
849 |
|
;; |
850 |
|
fmix.r f51 = f56, f72 |
851 |
|
fmix.r f81 = f57, f73 |
852 |
|
fmix.r f82 = f58, f74 |
853 |
|
fmix.r f83 = f59, f75 |
854 |
|
fmix.r f84 = f52, f60 |
855 |
|
fmix.r f85 = f53, f61 |
856 |
|
fmix.r f86 = f54, f62 |
857 |
|
fmix.r f87 = f55, f63 |
858 |
|
fmix.r f88 = f68, f80 |
859 |
|
fmix.r f89 = f69, f48 |
860 |
|
fmix.r f90 = f70, f49 |
861 |
|
fmix.r f91 = f71, f50 |
862 |
|
fmix.r f92 = f76, f64 |
863 |
|
fmix.r f93 = f77, f65 |
864 |
|
fmix.r f94 = f78, f66 |
865 |
|
fmix.r f95 = f79, f67 |
866 |
|
;; |
867 |
|
fmix.l f56 = f56, f72 |
868 |
|
fmix.l f57 = f57, f73 |
869 |
|
fmix.l f58 = f58, f74 |
870 |
|
fmix.l f59 = f59, f75 |
871 |
|
fmix.l f52 = f52, f60 |
872 |
|
fmix.l f53 = f53, f61 |
873 |
|
fmix.l f54 = f54, f62 |
874 |
|
fmix.l f55 = f55, f63 |
875 |
|
fmix.l f68 = f68, f80 |
876 |
|
fmix.l f69 = f69, f48 |
877 |
|
fmix.l f70 = f70, f49 |
878 |
|
fmix.l f71 = f71, f50 |
879 |
|
fmix.l f76 = f76, f64 |
880 |
|
fmix.l f77 = f77, f65 |
881 |
|
fmix.l f78 = f78, f66 |
882 |
|
fmix.l f79 = f79, f67 |
883 |
|
;; |
884 |
|
fpma f56 = f56, c0, f0 |
885 |
|
fpma f52 = f52, c0, f0 |
886 |
|
fpma f68 = f68, c0, f0 |
887 |
|
fpma f76 = f76, c0, f0 |
888 |
|
;; |
889 |
|
|
890 |
|
// before pre shuffle |
891 |
|
// 56 52 68 76 |
892 |
|
// 51 84 88 92 |
893 |
|
// 57 53 69 77 |
894 |
|
// 81 85 89 93 |
895 |
|
// 58 54 70 78 |
896 |
|
// 82 86 90 94 |
897 |
|
// 59 55 71 79 |
898 |
|
// 83 87 91 95 |
899 |
|
|
900 |
|
// after pre shuffle |
901 |
|
// 56 52 68 76 |
902 |
|
// 58 84 88 92 |
903 |
|
// 57 53 69 77 |
904 |
|
// 59 85 89 93 |
905 |
|
// 51 54 70 78 |
906 |
|
// 83 86 90 94 |
907 |
|
// 81 55 71 79 |
908 |
|
// 82 87 91 95 |
909 |
|
// (f48, f58) = (f56, f58) $ (c0, c0), (line 0, 1) |
910 |
|
fpma f48 = f58, c0, f56 |
911 |
|
fpnma f58 = f58, c0, f56 |
912 |
|
;; |
913 |
|
// (f49, f59) = (f57, f59) $ (c1, c2), (line 2, 3) |
914 |
|
fpma f49 = f59, c1, f57 |
915 |
|
fpnma f59 = f59, c2, f57 |
916 |
|
;; |
917 |
|
// (f50, f83) = (f51, f83) $ (c3, c4), (line 4, 5) |
918 |
|
fpma f50 = f83, c3, f51 |
919 |
|
fpnma f83 = f83, c4, f51 |
920 |
|
;; |
921 |
|
// (f51, f82) = (f81, f82) $ (c5, c6), (line 6, 7) |
922 |
|
fpma f51 = f82, c5, f81 |
923 |
|
fpnma f82 = f82, c6, f81 |
924 |
|
;; |
925 |
|
;; |
926 |
|
// (f56, f59) = (f48, f59) $ (c7, c7), (line 0, 3) |
927 |
|
fpma f56 = f59, c7, f48 |
928 |
|
fpnma f59 = f59, c7, f48 |
929 |
|
;; |
930 |
|
// (f48, f49) = (f58, f49) $ (c8, c8), (line 1, 2) |
931 |
|
fpma f48 = f49, c8, f58 |
932 |
|
fpnma f49 = f49, c8, f58 |
933 |
|
;; |
934 |
|
// (f57, f51) = (f50, f51) $ (c9, c9), (line 4, 6) |
935 |
|
fpma f57 = f51, c9, f50 |
936 |
|
fpnma f51 = f51, c9, f50 |
937 |
|
;; |
938 |
|
// (f50, f82) = (f83, f82) $ (c10, c10), (line 5, 7) |
939 |
|
fpma f50 = f82, c10, f83 |
940 |
|
fpnma f82 = f82, c10, f83 |
941 |
|
;; |
942 |
|
;; |
943 |
|
// (f58, f51) = (f50, f51) $ (c11, c11), (line 5, 6) |
944 |
|
fpma f58 = f51, c11, f50 |
945 |
|
fpnma f51 = f51, c11, f50 |
946 |
|
;; |
947 |
|
// (f50, f57) = (f56, f57) $ (c12, c12), (line 0, 4) |
948 |
|
fpma f50 = f57, c12, f56 |
949 |
|
fpnma f57 = f57, c12, f56 |
950 |
|
;; |
951 |
|
// (f56, f82) = (f59, f82) $ (c14, c14), (line 3, 7) |
952 |
|
fpma f56 = f82, c14, f59 |
953 |
|
fpnma f82 = f82, c14, f59 |
954 |
|
;; |
955 |
|
;; |
956 |
|
// (f59, f58) = (f48, f58) $ (c13, c13), (line 1, 5) |
957 |
|
fpma f59 = f58, c13, f48 |
958 |
|
fpnma f58 = f58, c13, f48 |
959 |
|
;; |
960 |
|
// (f48, f51) = (f49, f51) $ (c13, c13), (line 2, 6) |
961 |
|
fpma f48 = f51, c13, f49 |
962 |
|
fpnma f51 = f51, c13, f49 |
963 |
|
;; |
964 |
|
|
965 |
|
// before post shuffle |
966 |
|
// 50 52 68 76 |
967 |
|
// 59 84 88 92 |
968 |
|
// 48 53 69 77 |
969 |
|
// 56 85 89 93 |
970 |
|
// 57 54 70 78 |
971 |
|
// 58 86 90 94 |
972 |
|
// 51 55 71 79 |
973 |
|
// 82 87 91 95 |
974 |
|
|
975 |
|
// after post shuffle |
976 |
|
// 50 52 68 76 |
977 |
|
// 59 84 88 92 |
978 |
|
// 51 53 69 77 |
979 |
|
// 56 85 89 93 |
980 |
|
// 82 54 70 78 |
981 |
|
// 48 86 90 94 |
982 |
|
// 58 55 71 79 |
983 |
|
// 57 87 91 95 |
984 |
|
|
985 |
|
// before pre shuffle |
986 |
|
// 50 52 68 76 |
987 |
|
// 59 84 88 92 |
988 |
|
// 51 53 69 77 |
989 |
|
// 56 85 89 93 |
990 |
|
// 82 54 70 78 |
991 |
|
// 48 86 90 94 |
992 |
|
// 58 55 71 79 |
993 |
|
// 57 87 91 95 |
994 |
|
|
995 |
|
// after pre shuffle |
996 |
|
// 50 52 68 76 |
997 |
|
// 59 54 88 92 |
998 |
|
// 51 53 69 77 |
999 |
|
// 56 55 89 93 |
1000 |
|
// 82 84 70 78 |
1001 |
|
// 48 87 90 94 |
1002 |
|
// 58 85 71 79 |
1003 |
|
// 57 86 91 95 |
1004 |
|
// (f49, f54) = (f52, f54) $ (c0, c0), (line 0, 1) |
1005 |
|
fpma f49 = f54, c0, f52 |
1006 |
|
fpnma f54 = f54, c0, f52 |
1007 |
|
;; |
1008 |
|
// (f52, f55) = (f53, f55) $ (c1, c2), (line 2, 3) |
1009 |
|
fpma f52 = f55, c1, f53 |
1010 |
|
fpnma f55 = f55, c2, f53 |
1011 |
|
;; |
1012 |
|
// (f53, f87) = (f84, f87) $ (c3, c4), (line 4, 5) |
1013 |
|
fpma f53 = f87, c3, f84 |
1014 |
|
fpnma f87 = f87, c4, f84 |
1015 |
|
;; |
1016 |
|
// (f60, f86) = (f85, f86) $ (c5, c6), (line 6, 7) |
1017 |
|
fpma f60 = f86, c5, f85 |
1018 |
|
fpnma f86 = f86, c6, f85 |
1019 |
|
;; |
1020 |
|
;; |
1021 |
|
// (f61, f55) = (f49, f55) $ (c7, c7), (line 0, 3) |
1022 |
|
fpma f61 = f55, c7, f49 |
1023 |
|
fpnma f55 = f55, c7, f49 |
1024 |
|
;; |
1025 |
|
// (f49, f52) = (f54, f52) $ (c8, c8), (line 1, 2) |
1026 |
|
fpma f49 = f52, c8, f54 |
1027 |
|
fpnma f52 = f52, c8, f54 |
1028 |
|
;; |
1029 |
|
// (f54, f60) = (f53, f60) $ (c9, c9), (line 4, 6) |
1030 |
|
fpma f54 = f60, c9, f53 |
1031 |
|
fpnma f60 = f60, c9, f53 |
1032 |
|
;; |
1033 |
|
// (f53, f86) = (f87, f86) $ (c10, c10), (line 5, 7) |
1034 |
|
fpma f53 = f86, c10, f87 |
1035 |
|
fpnma f86 = f86, c10, f87 |
1036 |
|
;; |
1037 |
|
;; |
1038 |
|
// (f62, f60) = (f53, f60) $ (c11, c11), (line 5, 6) |
1039 |
|
fpma f62 = f60, c11, f53 |
1040 |
|
fpnma f60 = f60, c11, f53 |
1041 |
|
;; |
1042 |
|
// (f53, f54) = (f61, f54) $ (c12, c12), (line 0, 4) |
1043 |
|
fpma f53 = f54, c12, f61 |
1044 |
|
fpnma f54 = f54, c12, f61 |
1045 |
|
;; |
1046 |
|
// (f61, f86) = (f55, f86) $ (c14, c14), (line 3, 7) |
1047 |
|
fpma f61 = f86, c14, f55 |
1048 |
|
fpnma f86 = f86, c14, f55 |
1049 |
|
;; |
1050 |
|
;; |
1051 |
|
// (f55, f62) = (f49, f62) $ (c13, c13), (line 1, 5) |
1052 |
|
fpma f55 = f62, c13, f49 |
1053 |
|
fpnma f62 = f62, c13, f49 |
1054 |
|
;; |
1055 |
|
// (f49, f60) = (f52, f60) $ (c13, c13), (line 2, 6) |
1056 |
|
fpma f49 = f60, c13, f52 |
1057 |
|
fpnma f60 = f60, c13, f52 |
1058 |
|
;; |
1059 |
|
|
1060 |
|
// before post shuffle |
1061 |
|
// 50 53 68 76 |
1062 |
|
// 59 55 88 92 |
1063 |
|
// 51 49 69 77 |
1064 |
|
// 56 61 89 93 |
1065 |
|
// 82 54 70 78 |
1066 |
|
// 48 62 90 94 |
1067 |
|
// 58 60 71 79 |
1068 |
|
// 57 86 91 95 |
1069 |
|
|
1070 |
|
// after post shuffle |
1071 |
|
// 50 53 68 76 |
1072 |
|
// 59 55 88 92 |
1073 |
|
// 51 60 69 77 |
1074 |
|
// 56 61 89 93 |
1075 |
|
// 82 86 70 78 |
1076 |
|
// 48 49 90 94 |
1077 |
|
// 58 62 71 79 |
1078 |
|
// 57 54 91 95 |
1079 |
|
|
1080 |
|
// before pre shuffle |
1081 |
|
// 50 53 68 76 |
1082 |
|
// 59 55 88 92 |
1083 |
|
// 51 60 69 77 |
1084 |
|
// 56 61 89 93 |
1085 |
|
// 82 86 70 78 |
1086 |
|
// 48 49 90 94 |
1087 |
|
// 58 62 71 79 |
1088 |
|
// 57 54 91 95 |
1089 |
|
|
1090 |
|
// after pre shuffle |
1091 |
|
// 50 53 68 76 |
1092 |
|
// 59 55 70 92 |
1093 |
|
// 51 60 69 77 |
1094 |
|
// 56 61 71 93 |
1095 |
|
// 82 86 88 78 |
1096 |
|
// 48 49 91 94 |
1097 |
|
// 58 62 89 79 |
1098 |
|
// 57 54 90 95 |
1099 |
|
// (f52, f70) = (f68, f70) $ (c0, c0), (line 0, 1) |
1100 |
|
fpma f52 = f70, c0, f68 |
1101 |
|
fpnma f70 = f70, c0, f68 |
1102 |
|
;; |
1103 |
|
// (f63, f71) = (f69, f71) $ (c1, c2), (line 2, 3) |
1104 |
|
fpma f63 = f71, c1, f69 |
1105 |
|
fpnma f71 = f71, c2, f69 |
1106 |
|
;; |
1107 |
|
// (f64, f91) = (f88, f91) $ (c3, c4), (line 4, 5) |
1108 |
|
fpma f64 = f91, c3, f88 |
1109 |
|
fpnma f91 = f91, c4, f88 |
1110 |
|
;; |
1111 |
|
// (f65, f90) = (f89, f90) $ (c5, c6), (line 6, 7) |
1112 |
|
fpma f65 = f90, c5, f89 |
1113 |
|
fpnma f90 = f90, c6, f89 |
1114 |
|
;; |
1115 |
|
;; |
1116 |
|
// (f66, f71) = (f52, f71) $ (c7, c7), (line 0, 3) |
1117 |
|
fpma f66 = f71, c7, f52 |
1118 |
|
fpnma f71 = f71, c7, f52 |
1119 |
|
;; |
1120 |
|
// (f52, f63) = (f70, f63) $ (c8, c8), (line 1, 2) |
1121 |
|
fpma f52 = f63, c8, f70 |
1122 |
|
fpnma f63 = f63, c8, f70 |
1123 |
|
;; |
1124 |
|
// (f67, f65) = (f64, f65) $ (c9, c9), (line 4, 6) |
1125 |
|
fpma f67 = f65, c9, f64 |
1126 |
|
fpnma f65 = f65, c9, f64 |
1127 |
|
;; |
1128 |
|
// (f64, f90) = (f91, f90) $ (c10, c10), (line 5, 7) |
1129 |
|
fpma f64 = f90, c10, f91 |
1130 |
|
fpnma f90 = f90, c10, f91 |
1131 |
|
;; |
1132 |
|
;; |
1133 |
|
// (f68, f65) = (f64, f65) $ (c11, c11), (line 5, 6) |
1134 |
|
fpma f68 = f65, c11, f64 |
1135 |
|
fpnma f65 = f65, c11, f64 |
1136 |
|
;; |
1137 |
|
// (f64, f67) = (f66, f67) $ (c12, c12), (line 0, 4) |
1138 |
|
fpma f64 = f67, c12, f66 |
1139 |
|
fpnma f67 = f67, c12, f66 |
1140 |
|
;; |
1141 |
|
// (f66, f90) = (f71, f90) $ (c14, c14), (line 3, 7) |
1142 |
|
fpma f66 = f90, c14, f71 |
1143 |
|
fpnma f90 = f90, c14, f71 |
1144 |
|
;; |
1145 |
|
;; |
1146 |
|
// (f69, f68) = (f52, f68) $ (c13, c13), (line 1, 5) |
1147 |
|
fpma f69 = f68, c13, f52 |
1148 |
|
fpnma f68 = f68, c13, f52 |
1149 |
|
;; |
1150 |
|
// (f52, f65) = (f63, f65) $ (c13, c13), (line 2, 6) |
1151 |
|
fpma f52 = f65, c13, f63 |
1152 |
|
fpnma f65 = f65, c13, f63 |
1153 |
|
;; |
1154 |
|
|
1155 |
|
// before post shuffle |
1156 |
|
// 50 53 64 76 |
1157 |
|
// 59 55 69 92 |
1158 |
|
// 51 60 52 77 |
1159 |
|
// 56 61 66 93 |
1160 |
|
// 82 86 67 78 |
1161 |
|
// 48 49 68 94 |
1162 |
|
// 58 62 65 79 |
1163 |
|
// 57 54 90 95 |
1164 |
|
|
1165 |
|
// after post shuffle |
1166 |
|
// 50 53 64 76 |
1167 |
|
// 59 55 69 92 |
1168 |
|
// 51 60 65 77 |
1169 |
|
// 56 61 66 93 |
1170 |
|
// 82 86 90 78 |
1171 |
|
// 48 49 52 94 |
1172 |
|
// 58 62 68 79 |
1173 |
|
// 57 54 67 95 |
1174 |
|
|
1175 |
|
// before pre shuffle |
1176 |
|
// 50 53 64 76 |
1177 |
|
// 59 55 69 92 |
1178 |
|
// 51 60 65 77 |
1179 |
|
// 56 61 66 93 |
1180 |
|
// 82 86 90 78 |
1181 |
|
// 48 49 52 94 |
1182 |
|
// 58 62 68 79 |
1183 |
|
// 57 54 67 95 |
1184 |
|
|
1185 |
|
// after pre shuffle |
1186 |
|
// 50 53 64 76 |
1187 |
|
// 59 55 69 78 |
1188 |
|
// 51 60 65 77 |
1189 |
|
// 56 61 66 79 |
1190 |
|
// 82 86 90 92 |
1191 |
|
// 48 49 52 95 |
1192 |
|
// 58 62 68 93 |
1193 |
|
// 57 54 67 94 |
1194 |
|
// (f63, f78) = (f76, f78) $ (c0, c0), (line 0, 1) |
1195 |
|
fpma f63 = f78, c0, f76 |
1196 |
|
fpnma f78 = f78, c0, f76 |
1197 |
|
;; |
1198 |
|
// (f70, f79) = (f77, f79) $ (c1, c2), (line 2, 3) |
1199 |
|
fpma f70 = f79, c1, f77 |
1200 |
|
fpnma f79 = f79, c2, f77 |
1201 |
|
;; |
1202 |
|
// (f71, f95) = (f92, f95) $ (c3, c4), (line 4, 5) |
1203 |
|
fpma f71 = f95, c3, f92 |
1204 |
|
fpnma f95 = f95, c4, f92 |
1205 |
|
;; |
1206 |
|
// (f72, f94) = (f93, f94) $ (c5, c6), (line 6, 7) |
1207 |
|
fpma f72 = f94, c5, f93 |
1208 |
|
fpnma f94 = f94, c6, f93 |
1209 |
|
;; |
1210 |
|
;; |
1211 |
|
// (f73, f79) = (f63, f79) $ (c7, c7), (line 0, 3) |
1212 |
|
fpma f73 = f79, c7, f63 |
1213 |
|
fpnma f79 = f79, c7, f63 |
1214 |
|
;; |
1215 |
|
// (f63, f70) = (f78, f70) $ (c8, c8), (line 1, 2) |
1216 |
|
fpma f63 = f70, c8, f78 |
1217 |
|
fpnma f70 = f70, c8, f78 |
1218 |
|
;; |
1219 |
|
// (f74, f72) = (f71, f72) $ (c9, c9), (line 4, 6) |
1220 |
|
fpma f74 = f72, c9, f71 |
1221 |
|
fpnma f72 = f72, c9, f71 |
1222 |
|
;; |
1223 |
|
// (f71, f94) = (f95, f94) $ (c10, c10), (line 5, 7) |
1224 |
|
fpma f71 = f94, c10, f95 |
1225 |
|
fpnma f94 = f94, c10, f95 |
1226 |
|
;; |
1227 |
|
;; |
1228 |
|
// (f75, f72) = (f71, f72) $ (c11, c11), (line 5, 6) |
1229 |
|
fpma f75 = f72, c11, f71 |
1230 |
|
fpnma f72 = f72, c11, f71 |
1231 |
|
;; |
1232 |
|
// (f71, f74) = (f73, f74) $ (c12, c12), (line 0, 4) |
1233 |
|
fpma f71 = f74, c12, f73 |
1234 |
|
fpnma f74 = f74, c12, f73 |
1235 |
|
;; |
1236 |
|
// (f73, f94) = (f79, f94) $ (c14, c14), (line 3, 7) |
1237 |
|
fpma f73 = f94, c14, f79 |
1238 |
|
fpnma f94 = f94, c14, f79 |
1239 |
|
;; |
1240 |
|
;; |
1241 |
|
// (f76, f75) = (f63, f75) $ (c13, c13), (line 1, 5) |
1242 |
|
fpma f76 = f75, c13, f63 |
1243 |
|
fpnma f75 = f75, c13, f63 |
1244 |
|
;; |
1245 |
|
// (f63, f72) = (f70, f72) $ (c13, c13), (line 2, 6) |
1246 |
|
fpma f63 = f72, c13, f70 |
1247 |
|
fpnma f72 = f72, c13, f70 |
1248 |
|
;; |
1249 |
|
|
1250 |
|
// before post shuffle |
1251 |
|
// 50 53 64 71 |
1252 |
|
// 59 55 69 76 |
1253 |
|
// 51 60 65 63 |
1254 |
|
// 56 61 66 73 |
1255 |
|
// 82 86 90 74 |
1256 |
|
// 48 49 52 75 |
1257 |
|
// 58 62 68 72 |
1258 |
|
// 57 54 67 94 |
1259 |
|
|
1260 |
|
// after post shuffle |
1261 |
|
// 50 53 64 71 |
1262 |
|
// 59 55 69 76 |
1263 |
|
// 51 60 65 72 |
1264 |
|
// 56 61 66 73 |
1265 |
|
// 82 86 90 94 |
1266 |
|
// 48 49 52 63 |
1267 |
|
// 58 62 68 75 |
1268 |
|
// 57 54 67 74 |
1269 |
|
;; |
1270 |
|
fmix.r f70 = f50, f59 |
1271 |
|
fmix.r f77 = f53, f55 |
1272 |
|
fmix.r f78 = f64, f69 |
1273 |
|
fmix.r f79 = f71, f76 |
1274 |
|
fmix.r f80 = f51, f56 |
1275 |
|
fmix.r f81 = f60, f61 |
1276 |
|
fmix.r f83 = f65, f66 |
1277 |
|
fmix.r f84 = f72, f73 |
1278 |
|
fmix.r f85 = f82, f48 |
1279 |
|
fmix.r f87 = f86, f49 |
1280 |
|
fmix.r f88 = f90, f52 |
1281 |
|
fmix.r f89 = f94, f63 |
1282 |
|
fmix.r f91 = f58, f57 |
1283 |
|
fmix.r f92 = f62, f54 |
1284 |
|
fmix.r f93 = f68, f67 |
1285 |
|
fmix.r f95 = f75, f74 |
1286 |
|
;; |
1287 |
|
fmix.l f50 = f50, f59 |
1288 |
|
fmix.l f53 = f53, f55 |
1289 |
|
fmix.l f64 = f64, f69 |
1290 |
|
fmix.l f71 = f71, f76 |
1291 |
|
fmix.l f51 = f51, f56 |
1292 |
|
fmix.l f60 = f60, f61 |
1293 |
|
fmix.l f65 = f65, f66 |
1294 |
|
fmix.l f72 = f72, f73 |
1295 |
|
fmix.l f82 = f82, f48 |
1296 |
|
fmix.l f86 = f86, f49 |
1297 |
|
fmix.l f90 = f90, f52 |
1298 |
|
fmix.l f94 = f94, f63 |
1299 |
|
fmix.l f58 = f58, f57 |
1300 |
|
fmix.l f62 = f62, f54 |
1301 |
|
fmix.l f68 = f68, f67 |
1302 |
|
fmix.l f75 = f75, f74 |
1303 |
|
;; |
1304 |
|
// 50 51 82 58 |
1305 |
|
// 70 80 85 91 |
1306 |
|
// 53 60 86 62 |
1307 |
|
// 77 81 87 92 |
1308 |
|
// 64 65 90 68 |
1309 |
|
// 78 83 88 93 |
1310 |
|
// 71 72 94 75 |
1311 |
|
// 79 84 89 95 |
1312 |
|
mov addreg1 = in0 |
1313 |
|
add addreg2 = 4, in0 |
1314 |
|
;; |
1315 |
|
fpcvt.fx f50 = f50 |
1316 |
|
fpcvt.fx f51 = f51 |
1317 |
|
fpcvt.fx f82 = f82 |
1318 |
|
fpcvt.fx f58 = f58 |
1319 |
|
fpcvt.fx f70 = f70 |
1320 |
|
fpcvt.fx f80 = f80 |
1321 |
|
fpcvt.fx f85 = f85 |
1322 |
|
fpcvt.fx f91 = f91 |
1323 |
|
fpcvt.fx f53 = f53 |
1324 |
|
fpcvt.fx f60 = f60 |
1325 |
|
fpcvt.fx f86 = f86 |
1326 |
|
fpcvt.fx f62 = f62 |
1327 |
|
fpcvt.fx f77 = f77 |
1328 |
|
fpcvt.fx f81 = f81 |
1329 |
|
fpcvt.fx f87 = f87 |
1330 |
|
fpcvt.fx f92 = f92 |
1331 |
|
fpcvt.fx f64 = f64 |
1332 |
|
fpcvt.fx f65 = f65 |
1333 |
|
fpcvt.fx f90 = f90 |
1334 |
|
fpcvt.fx f68 = f68 |
1335 |
|
fpcvt.fx f78 = f78 |
1336 |
|
fpcvt.fx f83 = f83 |
1337 |
|
fpcvt.fx f88 = f88 |
1338 |
|
fpcvt.fx f93 = f93 |
1339 |
|
fpcvt.fx f71 = f71 |
1340 |
|
fpcvt.fx f72 = f72 |
1341 |
|
fpcvt.fx f94 = f94 |
1342 |
|
fpcvt.fx f75 = f75 |
1343 |
|
fpcvt.fx f79 = f79 |
1344 |
|
fpcvt.fx f84 = f84 |
1345 |
|
fpcvt.fx f89 = f89 |
1346 |
|
fpcvt.fx f95 = f95 |
1347 |
|
;; |
1348 |
|
getf.sig r33 = f50 |
1349 |
|
getf.sig r34 = f51 |
1350 |
|
getf.sig r35 = f82 |
1351 |
|
getf.sig r36 = f58 |
1352 |
|
getf.sig r37 = f70 |
1353 |
|
getf.sig r38 = f80 |
1354 |
|
getf.sig r39 = f85 |
1355 |
|
getf.sig r40 = f91 |
1356 |
|
getf.sig r41 = f53 |
1357 |
|
getf.sig r42 = f60 |
1358 |
|
getf.sig r43 = f86 |
1359 |
|
getf.sig r44 = f62 |
1360 |
|
getf.sig r45 = f77 |
1361 |
|
getf.sig r46 = f81 |
1362 |
|
getf.sig r47 = f87 |
1363 |
|
getf.sig r48 = f92 |
1364 |
|
getf.sig r49 = f64 |
1365 |
|
getf.sig r50 = f65 |
1366 |
|
getf.sig r51 = f90 |
1367 |
|
getf.sig r52 = f68 |
1368 |
|
getf.sig r53 = f78 |
1369 |
|
getf.sig r54 = f83 |
1370 |
|
getf.sig r55 = f88 |
1371 |
|
getf.sig r56 = f93 |
1372 |
|
getf.sig r57 = f71 |
1373 |
|
getf.sig r58 = f72 |
1374 |
|
getf.sig r59 = f94 |
1375 |
|
getf.sig r60 = f75 |
1376 |
|
getf.sig r61 = f79 |
1377 |
|
getf.sig r62 = f84 |
1378 |
|
getf.sig r63 = f89 |
1379 |
|
getf.sig r64 = f95 |
1380 |
|
;; |
1381 |
|
shl r33 = r33, 7 |
1382 |
|
shl r34 = r34, 7 |
1383 |
|
shl r35 = r35, 7 |
1384 |
|
shl r36 = r36, 7 |
1385 |
|
shl r37 = r37, 7 |
1386 |
|
shl r38 = r38, 7 |
1387 |
|
shl r39 = r39, 7 |
1388 |
|
shl r40 = r40, 7 |
1389 |
|
shl r41 = r41, 7 |
1390 |
|
shl r42 = r42, 7 |
1391 |
|
shl r43 = r43, 7 |
1392 |
|
shl r44 = r44, 7 |
1393 |
|
shl r45 = r45, 7 |
1394 |
|
shl r46 = r46, 7 |
1395 |
|
shl r47 = r47, 7 |
1396 |
|
shl r48 = r48, 7 |
1397 |
|
shl r49 = r49, 7 |
1398 |
|
shl r50 = r50, 7 |
1399 |
|
shl r51 = r51, 7 |
1400 |
|
shl r52 = r52, 7 |
1401 |
|
shl r53 = r53, 7 |
1402 |
|
shl r54 = r54, 7 |
1403 |
|
shl r55 = r55, 7 |
1404 |
|
shl r56 = r56, 7 |
1405 |
|
shl r57 = r57, 7 |
1406 |
|
shl r58 = r58, 7 |
1407 |
|
shl r59 = r59, 7 |
1408 |
|
shl r60 = r60, 7 |
1409 |
|
shl r61 = r61, 7 |
1410 |
|
shl r62 = r62, 7 |
1411 |
|
shl r63 = r63, 7 |
1412 |
|
shl r64 = r64, 7 |
1413 |
|
;; |
1414 |
|
pack4.sss r33 = r33, r0 |
1415 |
|
pack4.sss r34 = r34, r0 |
1416 |
|
pack4.sss r35 = r35, r0 |
1417 |
|
pack4.sss r36 = r36, r0 |
1418 |
|
pack4.sss r37 = r37, r0 |
1419 |
|
pack4.sss r38 = r38, r0 |
1420 |
|
pack4.sss r39 = r39, r0 |
1421 |
|
pack4.sss r40 = r40, r0 |
1422 |
|
pack4.sss r41 = r41, r0 |
1423 |
|
pack4.sss r42 = r42, r0 |
1424 |
|
pack4.sss r43 = r43, r0 |
1425 |
|
pack4.sss r44 = r44, r0 |
1426 |
|
pack4.sss r45 = r45, r0 |
1427 |
|
pack4.sss r46 = r46, r0 |
1428 |
|
pack4.sss r47 = r47, r0 |
1429 |
|
pack4.sss r48 = r48, r0 |
1430 |
|
pack4.sss r49 = r49, r0 |
1431 |
|
pack4.sss r50 = r50, r0 |
1432 |
|
pack4.sss r51 = r51, r0 |
1433 |
|
pack4.sss r52 = r52, r0 |
1434 |
|
pack4.sss r53 = r53, r0 |
1435 |
|
pack4.sss r54 = r54, r0 |
1436 |
|
pack4.sss r55 = r55, r0 |
1437 |
|
pack4.sss r56 = r56, r0 |
1438 |
|
pack4.sss r57 = r57, r0 |
1439 |
|
pack4.sss r58 = r58, r0 |
1440 |
|
pack4.sss r59 = r59, r0 |
1441 |
|
pack4.sss r60 = r60, r0 |
1442 |
|
pack4.sss r61 = r61, r0 |
1443 |
|
pack4.sss r62 = r62, r0 |
1444 |
|
pack4.sss r63 = r63, r0 |
1445 |
|
pack4.sss r64 = r64, r0 |
1446 |
|
;; |
1447 |
|
pshr2 r33 = r33, 7 |
1448 |
|
pshr2 r34 = r34, 7 |
1449 |
|
pshr2 r35 = r35, 7 |
1450 |
|
pshr2 r36 = r36, 7 |
1451 |
|
pshr2 r37 = r37, 7 |
1452 |
|
pshr2 r38 = r38, 7 |
1453 |
|
pshr2 r39 = r39, 7 |
1454 |
|
pshr2 r40 = r40, 7 |
1455 |
|
pshr2 r41 = r41, 7 |
1456 |
|
pshr2 r42 = r42, 7 |
1457 |
|
pshr2 r43 = r43, 7 |
1458 |
|
pshr2 r44 = r44, 7 |
1459 |
|
pshr2 r45 = r45, 7 |
1460 |
|
pshr2 r46 = r46, 7 |
1461 |
|
pshr2 r47 = r47, 7 |
1462 |
|
pshr2 r48 = r48, 7 |
1463 |
|
pshr2 r49 = r49, 7 |
1464 |
|
pshr2 r50 = r50, 7 |
1465 |
|
pshr2 r51 = r51, 7 |
1466 |
|
pshr2 r52 = r52, 7 |
1467 |
|
pshr2 r53 = r53, 7 |
1468 |
|
pshr2 r54 = r54, 7 |
1469 |
|
pshr2 r55 = r55, 7 |
1470 |
|
pshr2 r56 = r56, 7 |
1471 |
|
pshr2 r57 = r57, 7 |
1472 |
|
pshr2 r58 = r58, 7 |
1473 |
|
pshr2 r59 = r59, 7 |
1474 |
|
pshr2 r60 = r60, 7 |
1475 |
|
pshr2 r61 = r61, 7 |
1476 |
|
pshr2 r62 = r62, 7 |
1477 |
|
pshr2 r63 = r63, 7 |
1478 |
|
pshr2 r64 = r64, 7 |
1479 |
|
;; |
1480 |
|
mux2 r33 = r33, 0xe1 |
1481 |
|
mux2 r34 = r34, 0xe1 |
1482 |
|
mux2 r35 = r35, 0xe1 |
1483 |
|
mux2 r36 = r36, 0xe1 |
1484 |
|
mux2 r37 = r37, 0xe1 |
1485 |
|
mux2 r38 = r38, 0xe1 |
1486 |
|
mux2 r39 = r39, 0xe1 |
1487 |
|
mux2 r40 = r40, 0xe1 |
1488 |
|
mux2 r41 = r41, 0xe1 |
1489 |
|
mux2 r42 = r42, 0xe1 |
1490 |
|
mux2 r43 = r43, 0xe1 |
1491 |
|
mux2 r44 = r44, 0xe1 |
1492 |
|
mux2 r45 = r45, 0xe1 |
1493 |
|
mux2 r46 = r46, 0xe1 |
1494 |
|
mux2 r47 = r47, 0xe1 |
1495 |
|
mux2 r48 = r48, 0xe1 |
1496 |
|
mux2 r49 = r49, 0xe1 |
1497 |
|
mux2 r50 = r50, 0xe1 |
1498 |
|
mux2 r51 = r51, 0xe1 |
1499 |
|
mux2 r52 = r52, 0xe1 |
1500 |
|
mux2 r53 = r53, 0xe1 |
1501 |
|
mux2 r54 = r54, 0xe1 |
1502 |
|
mux2 r55 = r55, 0xe1 |
1503 |
|
mux2 r56 = r56, 0xe1 |
1504 |
|
mux2 r57 = r57, 0xe1 |
1505 |
|
mux2 r58 = r58, 0xe1 |
1506 |
|
mux2 r59 = r59, 0xe1 |
1507 |
|
mux2 r60 = r60, 0xe1 |
1508 |
|
mux2 r61 = r61, 0xe1 |
1509 |
|
mux2 r62 = r62, 0xe1 |
1510 |
|
mux2 r63 = r63, 0xe1 |
1511 |
|
mux2 r64 = r64, 0xe1 |
1512 |
|
;; |
1513 |
|
st4 [addreg1] = r33, 8 |
1514 |
|
st4 [addreg2] = r34, 8 |
1515 |
|
st4 [addreg1] = r35, 8 |
1516 |
|
st4 [addreg2] = r36, 8 |
1517 |
|
st4 [addreg1] = r37, 8 |
1518 |
|
st4 [addreg2] = r38, 8 |
1519 |
|
st4 [addreg1] = r39, 8 |
1520 |
|
st4 [addreg2] = r40, 8 |
1521 |
|
st4 [addreg1] = r41, 8 |
1522 |
|
st4 [addreg2] = r42, 8 |
1523 |
|
st4 [addreg1] = r43, 8 |
1524 |
|
st4 [addreg2] = r44, 8 |
1525 |
|
st4 [addreg1] = r45, 8 |
1526 |
|
st4 [addreg2] = r46, 8 |
1527 |
|
st4 [addreg1] = r47, 8 |
1528 |
|
st4 [addreg2] = r48, 8 |
1529 |
|
st4 [addreg1] = r49, 8 |
1530 |
|
st4 [addreg2] = r50, 8 |
1531 |
|
st4 [addreg1] = r51, 8 |
1532 |
|
st4 [addreg2] = r52, 8 |
1533 |
|
st4 [addreg1] = r53, 8 |
1534 |
|
st4 [addreg2] = r54, 8 |
1535 |
|
st4 [addreg1] = r55, 8 |
1536 |
|
st4 [addreg2] = r56, 8 |
1537 |
|
st4 [addreg1] = r57, 8 |
1538 |
|
st4 [addreg2] = r58, 8 |
1539 |
|
st4 [addreg1] = r59, 8 |
1540 |
|
st4 [addreg2] = r60, 8 |
1541 |
|
st4 [addreg1] = r61, 8 |
1542 |
|
st4 [addreg2] = r62, 8 |
1543 |
|
st4 [addreg1] = r63, 8 |
1544 |
|
st4 [addreg2] = r64, 8 |
1545 |
|
|
1546 |
|
mov ar.pfs = r16 |
1547 |
|
br.ret.sptk.few b0 |
1548 |
|
|
1549 |
|
.endp |