stan/jdk-24
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

8079565: aarch64: Add vectorization support for aarch64

Browse Source 
Add vectorization support

Reviewed-by: roland
This commit is contained in:
Ed Nevill 2015-06-04 12:04:18 +00:00
parent 5e3d542763
commit 10741a86f6
5 changed files with 1461 additions and 132 deletions

1391
hotspot/src/cpu/aarch64/vm/aarch64.ad

File diff suppressed because it is too large Load Diff

156
hotspot/src/cpu/aarch64/vm/assembler_aarch64.hpp

@ -466,6 +466,11 @@ class Address VALUE_OBJ_CLASS_SPEC {
case base_plus_offset:
{
unsigned size = i->get(31, 30);
if (i->get(26, 26) && i->get(23, 23)) {
// SIMD Q Type - Size = 128 bits
assert(size == 0, "bad size");
size = 0b100;
}
unsigned mask = (1 << size) - 1;
if (_offset < 0 || _offset & mask)
{
@ -1888,9 +1893,18 @@ public:
};
enum SIMD_RegVariant {
S32, D64, Q128
B, H, S, D, Q
};
#define INSN(NAME, op) \
void NAME(FloatRegister Rt, SIMD_RegVariant T, const Address &adr) { \
ld_st2((Register)Rt, adr, (int)T & 3, op + ((T==Q) ? 0b10:0b00), 1); \
} \
INSN(ldr, 1);
INSN(str, 0);
#undef INSN
private:
@ -1997,27 +2011,87 @@ public:
rf(Vm, 16), f(0b000111, 15, 10), rf(Vn, 5), rf(Vd, 0); \
}
INSN(eor, 0b101110001);
INSN(orr, 0b001110101);
INSN(eor, 0b101110001);
INSN(orr, 0b001110101);
INSN(andr, 0b001110001);
INSN(bic, 0b001110011);
INSN(bif, 0b101110111);
INSN(bit, 0b101110101);
INSN(bsl, 0b101110011);
INSN(orn, 0b001110111);
INSN(bic, 0b001110011);
INSN(bif, 0b101110111);
INSN(bit, 0b101110101);
INSN(bsl, 0b101110011);
INSN(orn, 0b001110111);
#undef INSN
#define INSN(NAME, opc) \
#define INSN(NAME, opc, opc2) \
void NAME(FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn, FloatRegister Vm) { \
starti; \
f(0, 31), f((int)T & 1, 30), f(opc, 29), f(0b01110, 28, 24); \
f((int)T >> 1, 23, 22), f(1, 21), rf(Vm, 16), f(0b100001, 15, 10); \
f((int)T >> 1, 23, 22), f(1, 21), rf(Vm, 16), f(opc2, 15, 10); \
rf(Vn, 5), rf(Vd, 0); \
}
INSN(addv, 0);
INSN(subv, 1);
INSN(addv, 0, 0b100001);
INSN(subv, 1, 0b100001);
INSN(mulv, 0, 0b100111);
INSN(sshl, 0, 0b010001);
INSN(ushl, 1, 0b010001);
#undef INSN
#define INSN(NAME, opc, opc2) \
void NAME(FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn) { \
starti; \
f(0, 31), f((int)T & 1, 30), f(opc, 29), f(0b01110, 28, 24); \
f((int)T >> 1, 23, 22), f(opc2, 21, 10); \
rf(Vn, 5), rf(Vd, 0); \
}
INSN(absr, 0, 0b100000101110);
INSN(negr, 1, 0b100000101110);
INSN(notr, 1, 0b100000010110);
INSN(addv, 0, 0b110001101110);
#undef INSN
#define INSN(NAME, op0, cmode0) \
void NAME(FloatRegister Vd, SIMD_Arrangement T, unsigned imm8, unsigned lsl = 0) { \
unsigned cmode = cmode0; \
unsigned op = op0; \
starti; \
assert(lsl == 0 || \
((T == T4H || T == T8H) && lsl == 8) || \
((T == T2S || T == T4S) && ((lsl >> 3) < 4)), "invalid shift"); \
cmode |= lsl >> 2; \
if (T == T4H || T == T8H) cmode |= 0b1000; \
if (!(T == T4H || T == T8H || T == T2S || T == T4S)) { \
assert(op == 0 && cmode0 == 0, "must be MOVI"); \
cmode = 0b1110; \
if (T == T1D || T == T2D) op = 1; \
} \
f(0, 31), f((int)T & 1, 30), f(op, 29), f(0b0111100000, 28, 19); \
f(imm8 >> 5, 18, 16), f(cmode, 15, 12), f(0x01, 11, 10), f(imm8 & 0b11111, 9, 5); \
rf(Vd, 0); \
}
INSN(movi, 0, 0);
INSN(orri, 0, 1);
INSN(mvni, 1, 0);
INSN(bici, 1, 1);
#undef INSN
#define INSN(NAME, op1, op2, op3) \
void NAME(FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn, FloatRegister Vm) { \
starti; \
assert(T == T2S || T == T4S || T == T2D, "invalid arrangement"); \
f(0, 31), f((int)T & 1, 30), f(op1, 29), f(0b01110, 28, 24), f(op2, 23); \
f(T==T2D ? 1:0, 22); f(1, 21), rf(Vm, 16), f(op3, 15, 10), rf(Vn, 5), rf(Vd, 0); \
}
INSN(fadd, 0, 0, 0b110101);
INSN(fdiv, 1, 0, 0b111111);
INSN(fmul, 1, 0, 0b110111);
INSN(fsub, 0, 1, 0b110101);
#undef INSN
@ -2064,19 +2138,40 @@ public:
#undef INSN
void shl(FloatRegister Vd, FloatRegister Vn, SIMD_Arrangement T, int shift){
void ins(FloatRegister Vd, SIMD_RegVariant T, FloatRegister Vn, int didx, int sidx) {
starti;
/* The encodings for the immh:immb fields (bits 22:16) are
* 0001 xxx 8B/16B, shift = xxx
* 001x xxx 4H/8H, shift = xxxx
* 01xx xxx 2S/4S, shift = xxxxx
* 1xxx xxx 1D/2D, shift = xxxxxx (1D is RESERVED)
*/
assert((1 << ((T>>1)+3)) > shift, "Invalid Shift value");
f(0, 31), f(T & 1, 30), f(0b0011110, 29, 23), f((1 << ((T>>1)+3))|shift, 22, 16);
f(0b010101, 15, 10), rf(Vn, 5), rf(Vd, 0);
assert(T != Q, "invalid register variant");
f(0b01101110000, 31, 21), f(((didx<<1)|1)<<(int)T, 20, 16), f(0, 15);
f(sidx<<(int)T, 14, 11), f(1, 10), rf(Vn, 5), rf(Vd, 0);
}
void umov(Register Rd, FloatRegister Vn, SIMD_RegVariant T, int idx) {
starti;
f(0, 31), f(T==D ? 1:0, 30), f(0b001110000, 29, 21);
f(((idx<<1)|1)<<(int)T, 20, 16), f(0b001111, 15, 10);
rf(Vn, 5), rf(Rd, 0);
}
#define INSN(NAME, opc, opc2) \
void NAME(FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn, int shift){ \
starti; \
/* The encodings for the immh:immb fields (bits 22:16) are \
* 0001 xxx 8B/16B, shift = xxx \
* 001x xxx 4H/8H, shift = xxxx \
* 01xx xxx 2S/4S, shift = xxxxx \
* 1xxx xxx 1D/2D, shift = xxxxxx (1D is RESERVED) \
*/ \
assert((1 << ((T>>1)+3)) > shift, "Invalid Shift value"); \
f(0, 31), f(T & 1, 30), f(opc, 29), f(0b011110, 28, 23), \
f((1 << ((T>>1)+3))|shift, 22, 16); f(opc2, 15, 10), rf(Vn, 5), rf(Vd, 0); \
}
INSN(shl, 0, 0b010101);
INSN(sshr, 0, 0b000001);
INSN(ushr, 1, 0b000001);
#undef INSN
void ushll(FloatRegister Vd, SIMD_Arrangement Ta, FloatRegister Vn, SIMD_Arrangement Tb, int shift) {
starti;
/* The encodings for the immh:immb fields (bits 22:16) are
@ -2149,6 +2244,23 @@ public:
rf(Vn, 5), rf(Vd, 0);
}
void dup(FloatRegister Vd, SIMD_Arrangement T, Register Xs)
{
starti;
assert(T != T1D, "reserved encoding");
f(0,31), f((int)T & 1, 30), f(0b001110000, 29, 21);
f((1 << (T >> 1)), 20, 16), f(0b000011, 15, 10), rf(Xs, 5), rf(Vd, 0);
}
void dup(FloatRegister Vd, SIMD_Arrangement T, FloatRegister Vn, int index = 0)
{
starti;
assert(T != T1D, "reserved encoding");
f(0, 31), f((int)T & 1, 30), f(0b001110000, 29, 21);
f(((1 << (T >> 1)) | (index << ((T >> 1) + 1))), 20, 16);
f(0b000001, 15, 10), rf(Vn, 5), rf(Vd, 0);
}
// CRC32 instructions
#define INSN(NAME, sf, sz) \
void NAME(Register Rd, Register Rn, Register Rm) { \

4
hotspot/src/cpu/aarch64/vm/macroAssembler_aarch64.cpp

@ -2802,8 +2802,8 @@ void MacroAssembler::kernel_crc32(Register crc, Register buf, Register len,
uzp2(v21, v20, v16, T2D);
eor(v20, T16B, v17, v21);
shl(v16, v28, T2D, 1);
shl(v17, v20, T2D, 1);
shl(v16, T2D, v28, 1);
shl(v17, T2D, v20, 1);
eor(v0, T16B, v0, v16);
eor(v1, T16B, v1, v17);

40
hotspot/src/cpu/aarch64/vm/macroAssembler_aarch64.hpp

@ -37,6 +37,7 @@ class MacroAssembler: public Assembler {
friend class LIR_Assembler;
using Assembler::mov;
using Assembler::movi;
protected:
@ -464,6 +465,45 @@ public:
void movptr(Register r, uintptr_t imm64);
// Macro to mov replicated immediate to vector register.
// Where imm32 == hex abcdefgh, Vd will get the following values
// for different arrangements in T
// T8B: Vd = ghghghghghghghgh
// T16B: Vd = ghghghghghghghghghghghghghghghgh
// T4H: Vd = efghefghefghefgh
// T8H: Vd = efghefghefghefghefghefghefghefgh
// T2S: Vd = abcdefghabcdefgh
// T4S: Vd = abcdefghabcdefghabcdefghabcdefgh
// T1D/T2D: invalid
void mov(FloatRegister Vd, SIMD_Arrangement T, u_int32_t imm32) {
assert(T != T1D && T != T2D, "invalid arrangement");
u_int32_t nimm32 = ~imm32;
if (T == T8B || T == T16B) { imm32 &= 0xff; nimm32 &= 0xff; }
if (T == T4H || T == T8H) { imm32 &= 0xffff; nimm32 &= 0xffff; }
u_int32_t x = imm32;
int movi_cnt = 0;
int movn_cnt = 0;
while (x) { if (x & 0xff) movi_cnt++; x >>= 8; }
x = nimm32;
while (x) { if (x & 0xff) movn_cnt++; x >>= 8; }
if (movn_cnt < movi_cnt) imm32 = nimm32;
unsigned lsl = 0;
while (imm32 && (imm32 & 0xff) == 0) { lsl += 8; imm32 >>= 8; }
if (movn_cnt < movi_cnt)
mvni(Vd, T, imm32 & 0xff, lsl);
else
movi(Vd, T, imm32 & 0xff, lsl);
imm32 >>= 8; lsl += 8;
while (imm32) {
while ((imm32 & 0xff) == 0) { lsl += 8; imm32 >>= 8; }
if (movn_cnt < movi_cnt)
bici(Vd, T, imm32 & 0xff, lsl);
else
orri(Vd, T, imm32 & 0xff, lsl);
lsl += 8; imm32 >>= 8;
}
}
// macro instructions for accessing and updating floating point
// status register
//

2
hotspot/src/cpu/aarch64/vm/register_aarch64.hpp

@ -186,7 +186,7 @@ class ConcreteRegisterImpl : public AbstractRegisterImpl {
// it's optoregs.
number_of_registers = (2 * RegisterImpl::number_of_registers +
2 * FloatRegisterImpl::number_of_registers +
4 * FloatRegisterImpl::number_of_registers +
1) // flags
};