tjh.dev
Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
1.align 2
2.global ___muldi3;
3.type ___muldi3, STT_FUNC;
4
5#ifdef CONFIG_ARITHMETIC_OPS_L1
6.section .l1.text
7#else
8.text
9#endif
10
11/*
12 R1:R0 * R3:R2
13 = R1.h:R1.l:R0.h:R0.l * R3.h:R3.l:R2.h:R2.l
14[X] = (R1.h * R3.h) * 2^96
15[X] + (R1.h * R3.l + R1.l * R3.h) * 2^80
16[X] + (R1.h * R2.h + R1.l * R3.l + R3.h * R0.h) * 2^64
17[T1] + (R1.h * R2.l + R3.h * R0.l + R1.l * R2.h + R3.l * R0.h) * 2^48
18[T2] + (R1.l * R2.l + R3.l * R0.l + R0.h * R2.h) * 2^32
19[T3] + (R0.l * R2.h + R2.l * R0.h) * 2^16
20[T4] + (R0.l * R2.l)
21
22 We can discard the first three lines marked "X" since we produce
23 only a 64 bit result. So, we need ten 16-bit multiplies.
24
25 Individual mul-acc results:
26[E1] = R1.h * R2.l + R3.h * R0.l + R1.l * R2.h + R3.l * R0.h
27[E2] = R1.l * R2.l + R3.l * R0.l + R0.h * R2.h
28[E3] = R0.l * R2.h + R2.l * R0.h
29[E4] = R0.l * R2.l
30
31 We also need to add high parts from lower-level results to higher ones:
32 E[n]c = E[n] + (E[n+1]c >> 16), where E4c := E4
33
34 One interesting property is that all parts of the result that depend
35 on the sign of the multiplication are discarded. Those would be the
36 multiplications involving R1.h and R3.h, but only the top 16 bit of
37 the 32 bit result depend on the sign, and since R1.h and R3.h only
38 occur in E1, the top half of these results is cut off.
39 So, we can just use FU mode for all of the 16-bit multiplies, and
40 ignore questions of when to use mixed mode. */
41
42___muldi3:
43 /* [SP] technically is part of the caller's frame, but we can
44 use it as scratch space. */
45 A0 = R2.H * R1.L, A1 = R2.L * R1.H (FU) || R3 = [SP + 12]; /* E1 */
46 A0 += R3.H * R0.L, A1 += R3.L * R0.H (FU) || [SP] = R4; /* E1 */
47 A0 += A1; /* E1 */
48 R4 = A0.w;
49 A0 = R0.l * R3.l (FU); /* E2 */
50 A0 += R2.l * R1.l (FU); /* E2 */
51
52 A1 = R2.L * R0.L (FU); /* E4 */
53 R3 = A1.w;
54 A1 = A1 >> 16; /* E3c */
55 A0 += R2.H * R0.H, A1 += R2.L * R0.H (FU); /* E2, E3c */
56 A1 += R0.L * R2.H (FU); /* E3c */
57 R0 = A1.w;
58 A1 = A1 >> 16; /* E2c */
59 A0 += A1; /* E2c */
60 R1 = A0.w;
61
62 /* low(result) = low(E3c):low(E4) */
63 R0 = PACK (R0.l, R3.l);
64 /* high(result) = E2c + (E1 << 16) */
65 R1.h = R1.h + R4.l (NS) || R4 = [SP];
66 RTS;
67
68.size ___muldi3, .-___muldi3