websnark/src/build_tomcook.js.old

1437 lines
43 KiB
JavaScript

module.exports = function buildTomCook(module, _prefix) {
const prefix = _prefix || "tomcook";
function buildMul1() {
const f = module.addFunction(prefix+"_mul1");
f.addParam("x", "i32");
f.addParam("y", "i32");
f.addParam("r", "i32");
f.addLocal("m0", "i64");
f.addLocal("m1", "i64");
f.addLocal("m2", "i64");
f.addLocal("m02", "i64");
f.addLocal("p0", "i64");
f.addLocal("p1", "i64");
f.addLocal("pn1", "i64");
f.addLocal("pn2", "i64");
f.addLocal("pi", "i64");
f.addLocal("q0", "i64");
f.addLocal("q1", "i64");
f.addLocal("qn1", "i64");
f.addLocal("qn2", "i64");
f.addLocal("qi", "i64");
f.addLocal("r0", "i64");
f.addLocal("r1", "i64");
f.addLocal("rn1", "i64");
f.addLocal("rn2", "i64");
f.addLocal("ri", "i64");
f.addLocal("s0", "i64");
f.addLocal("s1", "i64");
f.addLocal("s2", "i64");
f.addLocal("s3", "i64");
f.addLocal("s4", "i64");
f.addLocal("c", "i64");
const c = f.getCodeBuilder();
f.addCode(
// calculate p
c.setLocal("m0", c.i64_load32_u( c.getLocal("x"), 0)),
c.setLocal("m1", c.i64_load32_u( c.getLocal("x"), 4)),
c.setLocal("m2", c.i64_load32_u( c.getLocal("x"), 8)),
c.setLocal("m02", c.i64_add( c.getLocal("m0"), c.getLocal("m2"))),
c.setLocal("p0", c.getLocal("m0")),
c.setLocal("p1" , c.i64_add( c.getLocal("m02"),c.getLocal("m1"))),
c.setLocal("pn1", c.i64_sub( c.getLocal("m02"),c.getLocal("m1"))),
c.setLocal("pn2", c.i64_sub(
c.i64_shl(
c.i64_add( c.getLocal("pn1"), c.getLocal("m2")),
c.i64_const(1)
),
c.getLocal("m0")
)),
c.setLocal("pi", c.getLocal("m2")),
// calculate q and intercalate r
c.setLocal("m0", c.i64_load32_u( c.getLocal("y"), 0)),
c.setLocal("m1", c.i64_load32_u( c.getLocal("y"), 4)),
c.setLocal("m2", c.i64_load32_u( c.getLocal("y"), 8)),
c.setLocal("m02", c.i64_add( c.getLocal("m0"), c.getLocal("m2"))),
c.setLocal("q0", c.getLocal("m0")),
c.setLocal("r0", c.i64_mul(c.getLocal("p0"), c.getLocal("q0"))),
c.setLocal("q1" , c.i64_add( c.getLocal("m02"),c.getLocal("m1"))),
c.setLocal("r1", c.i64_mul(c.getLocal("p1"), c.getLocal("q1"))),
c.setLocal("qn1", c.i64_sub( c.getLocal("m02"),c.getLocal("m1"))),
c.setLocal("rn1", c.i64_mul(c.getLocal("pn1"), c.getLocal("qn1"))),
c.setLocal("qn2", c.i64_sub(
c.i64_shl(
c.i64_add( c.getLocal("qn1"), c.getLocal("m2") ),
c.i64_const(1)
),
c.getLocal("m0")
)),
c.setLocal("rn2", c.i64_mul(c.getLocal("pn2"), c.getLocal("qn2"))),
c.setLocal("qi", c.getLocal("m2")),
c.setLocal("ri", c.i64_mul(c.getLocal("pi"), c.getLocal("qi"))),
// Interpolation
c.setLocal("s0", c.getLocal("r0")),
c.setLocal("s4", c.getLocal("ri")),
c.setLocal("s3", c.i64_div_s(
c.i64_sub(
c.getLocal("rn2"),
c.getLocal("r1")
),
c.i64_const(3)
)),
c.setLocal("s1", c.i64_shr_s(
c.i64_sub(
c.getLocal("r1"),
c.getLocal("rn1")
),
c.i64_const(1)
)),
c.setLocal("s2", c.i64_sub(c.getLocal("rn1"), c.getLocal("r0"))),
c.setLocal("s3", c.i64_add(
c.i64_shr_s(
c.i64_sub(
c.getLocal("s2"),
c.getLocal("s3")
),
c.i64_const(1)
),
c.i64_shl(
c.getLocal("ri"),
c.i64_const(1)
)
)),
c.setLocal("s2", c.i64_add(
c.i64_sub(
c.getLocal("s2"),
c.getLocal("s4")
),
c.getLocal("s1"),
)),
c.setLocal("s1", c.i64_sub(
c.getLocal("s1"),
c.getLocal("s3")
)),
// Recomposition
c.setLocal("c", c.getLocal("s0")),
c.i64_store32(
c.getLocal("r"),
0,
c.i64_and(
c.getLocal("c"),
c.i64_const(0x1FFFFFFF)
)
),
c.setLocal("c",
c.i64_add(
c.i64_shr_u(
c.getLocal("c"),
c.i64_const(29)
),
c.getLocal("s1")
)
),
c.i64_store32(
c.getLocal("r"),
4,
c.i64_and(
c.getLocal("c"),
c.i64_const(0x1FFFFFFF)
)
),
c.setLocal("c",
c.i64_add(
c.i64_shr_u(
c.getLocal("c"),
c.i64_const(29)
),
c.getLocal("s2")
)
),
c.i64_store32(
c.getLocal("r"),
8,
c.i64_and(
c.getLocal("c"),
c.i64_const(0x1FFFFFFF)
)
),
c.setLocal("c",
c.i64_add(
c.i64_shr_u(
c.getLocal("c"),
c.i64_const(29)
),
c.getLocal("s3")
)
),
c.i64_store32(
c.getLocal("r"),
12,
c.i64_and(
c.getLocal("c"),
c.i64_const(0x1FFFFFFF)
)
),
c.setLocal("c",
c.i64_add(
c.i64_shr_u(
c.getLocal("c"),
c.i64_const(29)
),
c.getLocal("s4")
)
),
c.i64_store32(
c.getLocal("r"),
16,
c.i64_and(
c.getLocal("c"),
c.i64_const(0x1FFFFFFF)
)
),
c.setLocal("c",
c.i64_shr_u(
c.getLocal("c"),
c.i64_const(29)
),
),
c.i64_store32(
c.getLocal("r"),
20,
c.i64_and(
c.getLocal("c"),
c.i64_const(0x1FFFFFFF)
)
)
);
}
function buildMul(n) {
const f = module.addFunction(prefix+"_mul"+n);
f.addParam("x", "i32");
f.addParam("y", "i32");
f.addParam("r", "i32");
f.addLocal("s", "i32");
const c = f.getCodeBuilder();
f.addCode(
c.if(
c.i32_and(
c.i32_load( c.getLocal("x"), (n-1)*4 ),
c.i32_const(0x80000000)
),
c.if(
c.i32_and(
c.i32_load( c.getLocal("y"), (n-1)*4 ),
c.i32_const(0x80000000)
),
[
...c.call(prefix + "_mulnn" + n, c.getLocal("x"), c.getLocal("y"), c.getLocal("r")),
],
[
...c.call(prefix + "_mulnp" + n, c.getLocal("x"), c.getLocal("y"), c.getLocal("r")),
]
),
c.if(
c.i32_and(
c.i32_load( c.getLocal("y"), (n-1)*4 ),
c.i32_const(0x80000000)
),
[
...c.call(prefix + "_mulnp" + n, c.getLocal("y"), c.getLocal("x"), c.getLocal("r")),
],
[
...c.call(prefix + "_mulu" + n, c.getLocal("x"), c.getLocal("y"), c.getLocal("r")),
]
),
)
);
}
function buildMulU3() {
const f = module.addFunction(prefix+"_mulu3");
f.addParam("x", "i32");
f.addParam("y", "i32");
f.addParam("r", "i32");
f.addLocal("a0", "i64");
f.addLocal("a1", "i64");
f.addLocal("a2", "i64");
f.addLocal("b0", "i64");
f.addLocal("b1", "i64");
f.addLocal("b2", "i64");
f.addLocal("c", "i64");
const c = f.getCodeBuilder();
f.addCode(
// calculate p
c.setLocal("a0", c.i64_load32_u( c.getLocal("x"), 0)),
c.setLocal("b0", c.i64_load32_u( c.getLocal("y"), 0)),
c.setLocal("c", c.i64_mul( c.getLocal("a0"), c.getLocal("b0") )),
c.i64_store32(
c.getLocal("r"),
0,
c.i64_and(
c.getLocal("c"),
c.i64_const(0x1FFFFFFF)
)
),
c.setLocal("a1", c.i64_load32_u( c.getLocal("x"), 4)),
c.setLocal("b1", c.i64_load32_u( c.getLocal("y"), 4)),
c.setLocal("c", c.i64_add(
c.i64_shr_u( c.getLocal("c"), c.i64_const(29)),
c.i64_add(
c.i64_mul( c.getLocal("a0"), c.getLocal("b1") ),
c.i64_mul( c.getLocal("a1"), c.getLocal("b0") )
)
)),
c.i64_store32(
c.getLocal("r"),
4,
c.i64_and(
c.getLocal("c"),
c.i64_const(0x1FFFFFFF)
)
),
c.setLocal("a2", c.i64_load32_u( c.getLocal("x"), 8)),
c.setLocal("b2", c.i64_load32_u( c.getLocal("y"), 8)),
c.setLocal("c", c.i64_add(
c.i64_add(
c.i64_mul( c.getLocal("a0"), c.getLocal("b2") ),
c.i64_mul( c.getLocal("a2"), c.getLocal("b0") )
),
c.i64_add(
c.i64_shr_u( c.getLocal("c"), c.i64_const(29)),
c.i64_mul( c.getLocal("a1"), c.getLocal("b1") ),
)
)),
c.i64_store32(
c.getLocal("r"),
8,
c.i64_and(
c.getLocal("c"),
c.i64_const(0x1FFFFFFF)
)
),
c.setLocal("c", c.i64_add(
c.i64_add(
c.i64_shr_u( c.getLocal("c"), c.i64_const(29)),
c.i64_mul( c.getLocal("a1"), c.getLocal("b2"))
),
c.i64_mul( c.getLocal("a2"), c.getLocal("b1") ),
)),
c.i64_store32(
c.getLocal("r"),
12,
c.i64_and(
c.getLocal("c"),
c.i64_const(0x1FFFFFFF)
)
),
c.setLocal("c", c.i64_add(
c.i64_mul( c.getLocal("a2"), c.getLocal("b2") ),
c.i64_shr_u( c.getLocal("c"), c.i64_const(29)),
)),
c.i64_store32(
c.getLocal("r"),
16,
c.i64_and(
c.getLocal("c"),
c.i64_const(0x1FFFFFFF)
)
),
c.i64_store32(
c.getLocal("r"),
20,
c.i64_shr_u( c.getLocal("c"), c.i64_const(29)),
),
);
}
function buildMulNP3() {
const f = module.addFunction(prefix+"_mulnp3");
f.addParam("x", "i32");
f.addParam("y", "i32");
f.addParam("r", "i32");
f.addLocal("a0", "i64");
f.addLocal("a1", "i64");
f.addLocal("a2", "i64");
f.addLocal("b0", "i64");
f.addLocal("b1", "i64");
f.addLocal("b2", "i64");
f.addLocal("c", "i64");
const c = f.getCodeBuilder();
f.addCode(
// calculate p
c.setLocal("a0",
c.i64_add(
c.i64_xor(
c.i64_load32_u( c.getLocal("x"), 0),
c.i64_const(0x1FFFFFFF)
),
c.i64_const(1)
)
),
c.setLocal("b0", c.i64_load32_u( c.getLocal("y"), 0)),
c.setLocal("c", c.i64_sub(
c.i64_mul( c.getLocal("a0"), c.getLocal("b0") ),
c.i64_const(1)
)),
c.i64_store32(
c.getLocal("r"),
0,
c.i64_xor(
c.i64_and(
c.getLocal("c"),
c.i64_const(0x1FFFFFFF)
),
c.i64_const(0x1FFFFFFF)
)
),
c.setLocal("a1",
c.i64_xor(
c.i64_load32_u( c.getLocal("x"), 4),
c.i64_const(0x1FFFFFFF)
)
),
c.setLocal("b1", c.i64_load32_u( c.getLocal("y"), 4)),
c.setLocal("c", c.i64_add(
c.i64_shr_s( c.getLocal("c"), c.i64_const(29)),
c.i64_add(
c.i64_mul( c.getLocal("a0"), c.getLocal("b1") ),
c.i64_mul( c.getLocal("a1"), c.getLocal("b0") )
)
)),
c.i64_store32(
c.getLocal("r"),
4,
c.i64_xor(
c.i64_and(
c.getLocal("c"),
c.i64_const(0x1FFFFFFF)
),
c.i64_const(0x1FFFFFFF)
)
),
c.setLocal("a2",
c.i64_xor(
c.i64_load32_u( c.getLocal("x"), 8),
c.i64_const(0xFFFFFFFF)
)
),
c.setLocal("b2", c.i64_load32_u( c.getLocal("y"), 8)),
c.setLocal("c", c.i64_add(
c.i64_add(
c.i64_mul( c.getLocal("a0"), c.getLocal("b2") ),
c.i64_mul( c.getLocal("a2"), c.getLocal("b0") )
),
c.i64_add(
c.i64_shr_s( c.getLocal("c"), c.i64_const(29)),
c.i64_mul( c.getLocal("a1"), c.getLocal("b1") ),
)
)),
c.i64_store32(
c.getLocal("r"),
8,
c.i64_xor(
c.i64_and(
c.getLocal("c"),
c.i64_const(0x1FFFFFFF)
),
c.i64_const(0x1FFFFFFF)
)
),
c.setLocal("c", c.i64_add(
c.i64_add(
c.i64_shr_s( c.getLocal("c"), c.i64_const(29)),
c.i64_mul( c.getLocal("a1"), c.getLocal("b2"))
),
c.i64_mul( c.getLocal("a2"), c.getLocal("b1") ),
)),
c.i64_store32(
c.getLocal("r"),
12,
c.i64_xor(
c.i64_and(
c.getLocal("c"),
c.i64_const(0x1FFFFFFF)
),
c.i64_const(0x1FFFFFFF)
)
),
c.setLocal("c", c.i64_add(
c.i64_mul( c.getLocal("a2"), c.getLocal("b2") ),
c.i64_shr_s( c.getLocal("c"), c.i64_const(29)),
)),
c.i64_store32(
c.getLocal("r"),
16,
c.i64_xor(
c.i64_and(
c.getLocal("c"),
c.i64_const(0x1FFFFFFF)
),
c.i64_const(0x1FFFFFFF)
)
),
c.i64_store32(
c.getLocal("r"),
20,
c.i64_xor(
c.i64_shr_s( c.getLocal("c"), c.i64_const(29)),
c.i64_const(0xFFFFFFFF)
)
)
);
}
function buildMulNN3() {
const f = module.addFunction(prefix+"_mulnn3");
f.addParam("x", "i32");
f.addParam("y", "i32");
f.addParam("r", "i32");
f.addLocal("a0", "i64");
f.addLocal("a1", "i64");
f.addLocal("a2", "i64");
f.addLocal("b0", "i64");
f.addLocal("b1", "i64");
f.addLocal("b2", "i64");
f.addLocal("c", "i64");
const c = f.getCodeBuilder();
f.addCode(
// calculate p
c.setLocal("a0",
c.i64_add(
c.i64_xor(
c.i64_load32_u( c.getLocal("x"), 0),
c.i64_const(0x1FFFFFFF)
),
c.i64_const(1)
)
),
c.setLocal("b0",
c.i64_add(
c.i64_xor(
c.i64_load32_u( c.getLocal("y"), 0),
c.i64_const(0x1FFFFFFF)
),
c.i64_const(1)
)
),
c.setLocal("c", c.i64_mul( c.getLocal("a0"), c.getLocal("b0") )),
c.i64_store32(
c.getLocal("r"),
0,
c.i64_and(
c.getLocal("c"),
c.i64_const(0x1FFFFFFF)
),
),
c.setLocal("a1",
c.i64_xor(
c.i64_load32_u( c.getLocal("x"), 4),
c.i64_const(0x1FFFFFFF)
)
),
c.setLocal("b1",
c.i64_xor(
c.i64_load32_u( c.getLocal("y"), 4),
c.i64_const(0x1FFFFFFF)
)
),
c.setLocal("c", c.i64_add(
c.i64_shr_s( c.getLocal("c"), c.i64_const(29)),
c.i64_add(
c.i64_mul( c.getLocal("a0"), c.getLocal("b1") ),
c.i64_mul( c.getLocal("a1"), c.getLocal("b0") )
)
)),
c.i64_store32(
c.getLocal("r"),
4,
c.i64_and(
c.getLocal("c"),
c.i64_const(0x1FFFFFFF)
),
),
c.setLocal("a2",
c.i64_xor(
c.i64_load32_u( c.getLocal("x"), 8),
c.i64_const(0xFFFFFFFF)
)
),
c.setLocal("b2",
c.i64_xor(
c.i64_load32_u( c.getLocal("y"), 8),
c.i64_const(0xFFFFFFFF)
)
),
c.setLocal("c", c.i64_add(
c.i64_add(
c.i64_mul( c.getLocal("a0"), c.getLocal("b2") ),
c.i64_mul( c.getLocal("a2"), c.getLocal("b0") )
),
c.i64_add(
c.i64_shr_s( c.getLocal("c"), c.i64_const(29)),
c.i64_mul( c.getLocal("a1"), c.getLocal("b1") ),
)
)),
c.i64_store32(
c.getLocal("r"),
8,
c.i64_and(
c.getLocal("c"),
c.i64_const(0x1FFFFFFF)
)
),
c.setLocal("c", c.i64_add(
c.i64_add(
c.i64_shr_s( c.getLocal("c"), c.i64_const(29)),
c.i64_mul( c.getLocal("a1"), c.getLocal("b2"))
),
c.i64_mul( c.getLocal("a2"), c.getLocal("b1") ),
)),
c.i64_store32(
c.getLocal("r"),
12,
c.i64_and(
c.getLocal("c"),
c.i64_const(0x1FFFFFFF)
)
),
c.setLocal("c", c.i64_add(
c.i64_mul( c.getLocal("a2"), c.getLocal("b2") ),
c.i64_shr_s( c.getLocal("c"), c.i64_const(29)),
)),
c.i64_store32(
c.getLocal("r"),
16,
c.i64_and(
c.getLocal("c"),
c.i64_const(0x1FFFFFFF)
),
c.i64_const(0x1FFFFFFF)
),
c.i64_store32(
c.getLocal("r"),
20,
c.i64_shr_s( c.getLocal("c"), c.i64_const(29))
)
);
}
function buildNeg(n) {
const f = module.addFunction(prefix+"_neg"+n);
f.addParam("x", "i32");
f.addParam("r", "i32");
f.addLocal("c", "i64");
const c = f.getCodeBuilder();
for (let i=0; i<n-1; i++) {
f.addCode(
c.setLocal(
"c",
c.i64_add(
(i==0)
? c.i64_const(1)
: c.i64_shr_u(c.getLocal("c"), c.i64_const(29)),
c.i64_xor(
c.i64_load32_u( c.getLocal("x"), i*4),
c.i64_const(0x1FFFFFFF)
)
)
)
);
f.addCode(
c.i64_store32(
c.getLocal("r"),
i*4,
c.i64_and(
c.getLocal("c"),
c.i64_const(0x1FFFFFFF)
)
)
);
}
f.addCode(
c.setLocal(
"c",
c.i64_add(
c.i64_shr_u(c.getLocal("c"), c.i64_const(29)),
c.i64_xor(
c.i64_load32_u( c.getLocal("x"), (n-1)*4),
c.i64_const(0xFFFFFFFF)
) )
)
);
f.addCode(
c.i64_store32(
c.getLocal("r"),
(n-1)*4,
c.getLocal("c")
)
);
}
function buildAddOne(n) {
const f = module.addFunction(prefix+"_addOne"+n);
f.addParam("x", "i32");
f.addLocal("c", "i64");
const c = f.getCodeBuilder();
for (let i=0; i<n; i++) {
f.addCode(
c.setLocal(
"c",
c.i64_add(
c.i64_const(1),
c.i64_load32_u( c.getLocal("x"), i*4),
)
)
);
f.addCode(
c.i64_store32(
c.getLocal("x"),
i*4,
c.i64_and(
c.getLocal("c"),
c.i64_const((i == n-1) ? 0xFFFFFFFF : 0x1FFFFFFF)
)
)
);
if (i<n-1) {
f.addCode(
c.if(
c.i64_eqz(
c.i64_and(
c.getLocal("c"),
c.i64_const("0x20000000")
)
),
c.ret([])
)
);
}
}
}
function buildAdd(n) {
const f = module.addFunction(prefix+"_add"+n);
f.addParam("x", "i32");
f.addParam("y", "i32");
f.addParam("r", "i32");
f.addLocal("c", "i64");
const c = f.getCodeBuilder();
for (let i=0; i<n; i++) {
if ( i==0) {
f.addCode(
c.setLocal(
"c",
c.i64_add(
c.i64_load32_s( c.getLocal("x"), i*4),
c.i64_load32_s( c.getLocal("y"), i*4),
)
)
);
} else {
f.addCode(
c.setLocal(
"c",
c.i64_add(
c.i64_shr_u(c.getLocal("c"), c.i64_const(29)),
c.i64_add(
c.i64_load32_s( c.getLocal("x"), i*4),
c.i64_load32_s( c.getLocal("y"), i*4),
)
)
)
);
}
f.addCode(
c.i64_store32(
c.getLocal("r"),
i*4,
c.i64_and(
c.getLocal("c"),
c.i64_const(i==n-1 ? 0xFFFFFFFF : 0x1FFFFFFF)
)
)
);
}
}
function buildSub(n) {
const f = module.addFunction(prefix+"_sub"+n);
f.addParam("x", "i32");
f.addParam("y", "i32");
f.addParam("r", "i32");
f.addLocal("c", "i64");
const c = f.getCodeBuilder();
f.addCode(c.setLocal("c", c.i64_const(1 << 29)));
for (let i=0; i<n; i++) {
f.addCode(
c.setLocal(
"c",
c.i64_add(
c.i64_shr_u(c.getLocal("c"), c.i64_const(29)),
c.i64_add(
c.i64_load32_s( c.getLocal("x"), i*4),
c.i64_xor(
c.i64_load32_s( c.getLocal("y"), i*4),
c.i64_const(i==n-1 ? 0xFFFFFFFF : 0x1FFFFFFF)
)
)
)
)
);
f.addCode(
c.i64_store32(
c.getLocal("r"),
i*4,
c.i64_and(
c.getLocal("c"),
c.i64_const(i==n-1 ? 0xFFFFFFFF : 0x1FFFFFFF)
)
)
);
}
}
function buildDouble(n) {
const f = module.addFunction(prefix+"_double"+n);
f.addParam("x", "i32");
f.addParam("r", "i32");
f.addLocal("c", "i64");
const c = f.getCodeBuilder();
for (let i=0; i<n; i++) {
if ( i==0) {
f.addCode(
c.setLocal(
"c",
c.i64_shl(
c.i64_load32_s( c.getLocal("x"), i*4),
c.i64_const(1)
)
)
);
} else {
f.addCode(
c.setLocal(
"c",
c.i64_add(
c.i64_shr_u(c.getLocal("c"), c.i64_const(29)),
c.i64_shl(
c.i64_load32_s( c.getLocal("x"), i*4),
c.i64_const(1)
)
)
)
);
}
f.addCode(
c.i64_store32(
c.getLocal("r"),
i*4,
c.i64_and(
c.getLocal("c"),
c.i64_const(i==n-1 ? 0xFFFFFFFF : 0x1FFFFFFF)
)
)
);
}
}
function buildHalve(n) {
const f = module.addFunction(prefix+"_halve"+n);
f.addParam("x", "i32");
f.addParam("r", "i32");
f.addLocal("c", "i64");
f.addLocal("sign", "i64");
const c = f.getCodeBuilder();
for (let i=n-1; i>=0; i--) {
if ( i==n-1) {
f.addCode(
c.setLocal(
"c",
c.i64_load32_s( c.getLocal("x"), i*4)
),
c.setLocal(
"sign",
c.i64_shr_u(
c.getLocal("c"),
c.i64_const(63)
)
)
);
} else {
f.addCode(
c.setLocal(
"c",
c.i64_or(
c.i64_shl(c.getLocal("c"), c.i64_const(29)),
c.i64_load32_u( c.getLocal("x"), i*4),
)
)
);
}
f.addCode(
c.i64_store32(
c.getLocal("r"),
i*4,
c.i64_and(
c.i64_shr_s(
c.getLocal("c"),
c.i64_const(1)
),
c.i64_const((i==n-1) ? 0xFFFFFFFF : 0x1FFFFFFF)
)
)
);
}
f.addCode(
c.if(
c.i32_wrap_i64(
c.i64_and(
c.getLocal("c"),
c.getLocal("sign")
)
),
c.call(prefix + "_addOne" + n , c.getLocal("r"))
)
);
}
function buildDivShort(n) {
buildDivShortRaw(n, "pp");
buildDivShortRaw(n, "pn");
buildDivShortRaw(n, "np");
buildDivShortRaw(n, "nn");
let c;
const fxp = module.addFunction(prefix+"_divshortxp"+n);
c = fxp.getCodeBuilder();
fxp.addParam("x", "i32");
fxp.addParam("s", "i32");
fxp.addParam("r", "i32");
fxp.addCode(
c.if(
c.i32_and(
c.i32_load( c.getLocal("x"), (n-1)*4 ),
c.i32_const(0x80000000)
),
[
...c.call(prefix + "_divshortnp" + n, c.getLocal("x"), c.getLocal("s"), c.getLocal("r")),
],
[
...c.call(prefix + "_divshortpp" + n, c.getLocal("x"), c.getLocal("s"), c.getLocal("r")),
]
),
);
const fxn = module.addFunction(prefix+"_divshortxn"+n);
c = fxn.getCodeBuilder();
fxn.addParam("x", "i32");
fxn.addParam("s", "i32");
fxn.addParam("r", "i32");
fxn.addCode(
c.if(
c.i32_and(
c.i32_load( c.getLocal("x"), (n-1)*4 ),
c.i32_const(0x80000000)
),
[
...c.call(prefix + "_divshortnn" + n, c.getLocal("x"), c.getLocal("s"), c.getLocal("r")),
],
[
...c.call(prefix + "_divshortpn" + n, c.getLocal("x"), c.getLocal("s"), c.getLocal("r")),
]
),
);
const f = module.addFunction(prefix+"_divshort"+n);
c = f.getCodeBuilder();
f.addParam("x", "i32");
f.addParam("s", "i32");
f.addParam("r", "i32");
f.addCode(
c.if(
c.i32_lt_s(
c.getLocal("s"),
c.i32_const(0)
),
[
...c.call(prefix + "_divshortxn" + n, c.getLocal("x"), c.getLocal("s"), c.getLocal("r")),
],
[
...c.call(prefix + "_divshortxp" + n, c.getLocal("x"), c.getLocal("s"), c.getLocal("r")),
]
),
);
// signs = "pp", "np", "pn", "nn"
function buildDivShortRaw(n, signs) {
const f = module.addFunction(prefix+"_divshort"+signs+n);
f.addParam("x", "i32");
f.addParam("s32", "i32");
f.addParam("r", "i32");
f.addLocal("c", "i64");
f.addLocal("s", "i64");
const c = f.getCodeBuilder();
if (signs[1] == "n") {
f.addCode(c.setLocal("s", c.i64_sub(c.i64_const(0), c.getLocal("s"))));
}
f.addCode(
c.setLocal("s", c.i64_extend_i32_u(c.getLocal("s32"))),
c.setLocal(
"c",
signs[0] == "p"
? c.i64_load32_u( c.getLocal("x"), (n-1)*4)
: c.i64_xor(
c.i64_load32_u( c.getLocal("x"), (n-1)*4),
c.i64_const(0xFFFFFFFF)
)
),
c.i64_store32(
c.getLocal("r"),
(n-1)*4,
(signs == "pp") || (signs == "nn")
? c.i64_div_u(
c.getLocal("c"),
c.getLocal("s")
)
: c.i64_xor(
c.i64_div_u(
c.getLocal("c"),
c.getLocal("s")
),
c.i64_const(0xFFFFFFFF)
)
)
);
for (let i=n-2; i>=0; i--) {
f.addCode(
c.setLocal(
"c",
c.i64_or(
c.i64_shl(
c.i64_rem_u(
c.getLocal("c"),
c.getLocal("s")
),
c.i64_const(29)
),
signs[0] == "p"
? c.i64_load32_u( c.getLocal("x"), i*4)
: c.i64_xor(
c.i64_load32_u( c.getLocal("x"), i*4),
c.i64_const(0x1FFFFFFF)
)
)
)
);
if ((i==0) && (signs[0] == "n")) {
f.addCode(
c.setLocal(
"c",
c.i64_add(
c.getLocal("c"),
c.i64_const(1)
)
)
);
}
f.addCode(
c.i64_store32(
c.getLocal("r"),
i*4,
(signs == "pp") || (signs == "nn")
? c.i64_div_u(
c.getLocal("c"),
c.getLocal("s")
)
: c.i64_xor(
c.i64_div_u(
c.getLocal("c"),
c.getLocal("s")
),
c.i64_const(0x1FFFFFFF)
)
)
);
}
if ((signs == "pn") || (signs == "np")) {
f.addCode(
c.call(prefix + "_addOne" + n , c.getLocal("r"))
);
}
}
}
function buildRecompose(n) {
const sn = n/3;
const f = module.addFunction(prefix+"_recompose"+n);
f.addParam("s", "i32");
f.addParam("r", "i32");
f.addLocal("c", "i64");
const c = f.getCodeBuilder();
/*
0 sn sn*2 sn*3 sn*4 sn*5
0 sn
sn*2 sn*3
sn*4 sn*5
sn+6 sn*7
sn*8 sn*9
b= Math.floor(i/sn)
b*2*sn + (i-b*sn)
b*2*sn + (i-b*sn) - sn
*/
for (let i=0; i<sn; i++) {
f.addCode(
c.i64_store32(
c.getLocal("r"),
i*4,
c.i64_load32_s( c.getLocal("s"), i*4),
)
);
}
for (let i=sn; i<sn*5; i++) {
const b= Math.floor(i/sn);
const i1 = b*2*sn + (i-b*sn);
const i2 = i1 - sn;
if ( i==sn) {
f.addCode(
c.setLocal(
"c",
c.i64_add(
c.i64_load32_s( c.getLocal("s"), i1*4),
c.i64_load32_s( c.getLocal("s"), i2*4),
)
)
);
} else {
f.addCode(
c.setLocal(
"c",
c.i64_add(
c.i64_shr_u(c.getLocal("c"), c.i64_const(29)),
c.i64_add(
c.i64_load32_s( c.getLocal("s"), i1*4),
c.i64_load32_s( c.getLocal("s"), i2*4),
)
)
)
);
}
f.addCode(
c.i64_store32(
c.getLocal("r"),
i*4,
c.i64_and(
c.getLocal("c"),
c.i64_const(0x1FFFFFFF)
)
)
);
}
for (let i=sn*5; i<n; i++) {
const b= Math.floor(i/sn);
const i1 = b*2*sn + (i-b*sn);
const i2 = i1 - sn;
f.addCode(
c.setLocal(
"c",
c.i64_add(
c.i64_shr_u(c.getLocal("c"), c.i64_const(29)),
c.i64_load32_s( c.getLocal("s"), i2*4)
)
)
);
f.addCode(
c.i64_store32(
c.getLocal("r"),
i*4,
c.i64_and(
c.getLocal("c"),
c.i64_const(i==n-1 ? 0xFFFFFFFF : 0x1FFFFFFF)
)
)
);
}
}
function buildMulU(n) {
const sn = n/3;
const sn2 = sn*2;
const f = module.addFunction(prefix+"_mulu"+n);
f.addParam("x", "i32");
f.addParam("y", "i32");
f.addParam("r", "i32");
f.addLocal("m1", "i32");
f.addLocal("m2", "i32");
f.addLocal("n1", "i32");
f.addLocal("n2", "i32");
const c = f.getCodeBuilder();
const m0 = c.getLocal("x");
f.addCode(c.setLocal("m1", c.i32_add( c.getLocal("x"), c.i32_const(sn*4) )));
const m1 = c.getLocal("m1");
f.addCode(c.setLocal("m2", c.i32_add( c.getLocal("x"), c.i32_const((sn*2)*4 ))));
const m2 = c.getLocal("m2");
const n0 = c.getLocal("y");
f.addCode(c.setLocal("n1", c.i32_add( c.getLocal("y"), c.i32_const(sn*4) )));
const n1 = c.getLocal("n1");
f.addCode(c.setLocal("n2", c.i32_add( c.getLocal("y"), c.i32_const((sn*2)*4 ))));
const n2 = c.getLocal("n2");
const po = c.i32_const(module.alloc(sn*4));
f.addCode(c.call(prefix + "_add" + sn, m0, m2, po));
const p0 = m0;
const p1 = c.i32_const(module.alloc(sn*4));
f.addCode(c.call(prefix + "_add" + sn, po, m1, p1));
const pn1 = c.i32_const(module.alloc(sn*4));
f.addCode(c.call(prefix + "_sub" + sn, po, m1, pn1));
const pn2 = c.i32_const(module.alloc(sn*4));
f.addCode(
c.call(prefix + "_add" + sn, pn1, m2, pn2),
c.call(prefix + "_double" + sn, pn2, pn2),
c.call(prefix + "_sub" + sn, pn2, m0, pn2)
);
const pi = m2;
const qo = c.i32_const(module.alloc(sn*4));
f.addCode(c.call(prefix + "_add" + sn, n0, n2, qo));
const q0 = n0;
const q1 = c.i32_const(module.alloc(sn*4));
f.addCode(c.call(prefix + "_add" + sn, qo, n1, q1));
const qn1 = c.i32_const(module.alloc(sn*4));
f.addCode(c.call(prefix + "_sub" + sn, qo, n1, qn1));
const qn2 = c.i32_const(module.alloc(sn*4));
f.addCode(
c.call(prefix + "_add" + sn, qn1, n2, qn2),
c.call(prefix + "_double" + sn, qn2, qn2),
c.call(prefix + "_sub" + sn, qn2, n0, qn2)
);
const qi = n2;
const ps0 = module.alloc(sn*2*5*4);
const s0 = c.i32_const(ps0);
const s1 = c.i32_const(ps0 + (2*sn)*4);
const s2 = c.i32_const(ps0 + (4*sn)*4);
const s3 = c.i32_const(ps0 + (6*sn)*4);
const s4 = c.i32_const(ps0 + (8*sn)*4);
const r0 = s0;
const r1 = c.i32_const(module.alloc(sn*2*4));
const rn1 = c.i32_const(module.alloc(sn*2*4));
const rn2 = c.i32_const(module.alloc(sn*2*4));
const ri = s4;
f.addCode(c.call(prefix + "_mulu" + sn, p0, q0, r0));
f.addCode(c.call(prefix + "_mulu" + sn, p1, q1, r1));
f.addCode(c.call(prefix + "_mul" + sn, pn1, qn1, rn1));
f.addCode(c.call(prefix + "_mul" + sn, pn2, qn2, rn2));
f.addCode(c.call(prefix + "_mulu" + sn, pi, qi, ri));
const aux = c.i32_const(module.alloc(sn*4));
f.addCode(
// s3 = (r(-2) - r(1))/3
c.call(prefix + "_sub" + sn2, rn2, r1, s3),
c.call(prefix + "_divshort" + sn2, s3, c.i32_const(3), s3),
// s1 = (r(1) - r(-1))/2
c.call(prefix + "_sub" + sn2, r1, rn1, s1),
c.call(prefix + "_halve" + sn2, s1, s1),
// s2 = r(-1) - r(0)
c.call(prefix + "_sub" + sn2, rn1, r0, s2),
// s3 = (s2-s3)/2 + r(inf)*2
c.call(prefix + "_sub" + sn2, s2, s3, s3),
c.call(prefix + "_halve" + sn2, s3, s3),
c.call(prefix + "_double" + sn2, ri, aux),
c.call(prefix + "_add" + sn2, s3, aux, s3),
// s2 = s2 + s1 - s4
c.call(prefix + "_add" + sn2, s2, s1, s2),
c.call(prefix + "_sub" + sn2, s2, s4, s2),
// s1 = s1 - s3
c.call(prefix + "_sub" + sn2, s1, s3, s1),
);
f.addCode(c.call(prefix + "_recompose" + n, s0, c.getLocal("r")));
}
buildAddOne(6);
buildDivShort(6);
buildMul1();
buildMulU3();
buildMulNP3();
buildMulNN3();
buildNeg(3);
buildMul(3);
buildNeg(6);
buildAdd(3);
buildAdd(6);
buildSub(3);
buildSub(6);
buildDouble(3);
buildDouble(6);
buildHalve(6);
buildRecompose(9);
buildMulU(9);
module.exportFunction(prefix+"_mul1");
module.exportFunction(prefix+"_mulu3");
module.exportFunction(prefix+"_divshort6");
module.exportFunction(prefix+"_double6");
module.exportFunction(prefix+"_halve6");
module.exportFunction(prefix+"_mul3");
module.exportFunction(prefix+"_neg3");
module.exportFunction(prefix+"_neg6");
module.exportFunction(prefix+"_mulu9");
};