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

8129551: aarch64: some regressions introduced by addition of vectorisation code

Browse Source 
Fix regressions

Reviewed-by: kvn
This commit is contained in:
Ed Nevill 2015-06-23 18:56:17 +00:00
parent 9ced0d90d5
commit d17ff6b63d
4 changed files with 54 additions and 40 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
hotspot/src/cpu/aarch64/vm/aarch64.ad
View File

@ -13276,7 +13276,7 @@ instruct replicate16B_imm(vecX dst, immI con)
ins_cost(INSN_COST); ins_cost(INSN_COST);
format %{ "movi $dst, $con\t# vector(16B)" %} format %{ "movi $dst, $con\t# vector(16B)" %}
ins_encode %{ ins_encode %{
__ mov(as_FloatRegister($dst$$reg), __ T16B, $con$$constant); __ mov(as_FloatRegister($dst$$reg), __ T16B, $con$$constant & 0xff);
%} %}
ins_pipe(pipe_class_default); ins_pipe(pipe_class_default);
%} %}
@ -13298,7 +13298,7 @@ instruct replicate8S_imm(vecX dst, immI con)
ins_cost(INSN_COST); ins_cost(INSN_COST);
format %{ "movi $dst, $con\t# vector(8H)" %} format %{ "movi $dst, $con\t# vector(8H)" %}
ins_encode %{ ins_encode %{
__ mov(as_FloatRegister($dst$$reg), __ T8H, $con$$constant); __ mov(as_FloatRegister($dst$$reg), __ T8H, $con$$constant & 0xffff);
%} %}
ins_pipe(pipe_class_default); ins_pipe(pipe_class_default);
%} %}

5
hotspot/src/cpu/aarch64/vm/assembler_aarch64.hpp
View File

@ -491,6 +491,11 @@ class Address VALUE_OBJ_CLASS_SPEC {
i->rf(_index, 16); i->rf(_index, 16);
i->f(_ext.option(), 15, 13); i->f(_ext.option(), 15, 13);
unsigned size = i->get(31, 30); 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;
}
if (size == 0) // It's a byte if (size == 0) // It's a byte
i->f(_ext.shift() >= 0, 12); i->f(_ext.shift() >= 0, 12);
else { else {

46
hotspot/src/cpu/aarch64/vm/macroAssembler_aarch64.cpp
View File

@ -1408,6 +1408,52 @@ void MacroAssembler::movptr(Register r, uintptr_t imm64) {
movk(r, imm64 & 0xffff, 32); movk(r, imm64 & 0xffff, 32);
} }
// Macro to mov replicated immediate to vector register.
// Vd will get the following values for different arrangements in T
// imm32 == hex 000000gh T8B: Vd = ghghghghghghghgh
// imm32 == hex 000000gh T16B: Vd = ghghghghghghghghghghghghghghghgh
// imm32 == hex 0000efgh T4H: Vd = efghefghefghefgh
// imm32 == hex 0000efgh T8H: Vd = efghefghefghefghefghefghefghefgh
// imm32 == hex abcdefgh T2S: Vd = abcdefghabcdefgh
// imm32 == hex abcdefgh T4S: Vd = abcdefghabcdefghabcdefghabcdefgh
// T1D/T2D: invalid
void MacroAssembler::mov(FloatRegister Vd, SIMD_Arrangement T, u_int32_t imm32) {
assert(T != T1D && T != T2D, "invalid arrangement");
if (T == T8B || T == T16B) {
assert((imm32 & ~0xff) == 0, "extraneous bits in unsigned imm32 (T8B/T16B)");
movi(Vd, T, imm32 & 0xff, 0);
return;
}
u_int32_t nimm32 = ~imm32;
if (T == T4H || T == T8H) {
assert((imm32 & ~0xffff) == 0, "extraneous bits in unsigned imm32 (T4H/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;
}
}
void MacroAssembler::mov_immediate64(Register dst, u_int64_t imm64) void MacroAssembler::mov_immediate64(Register dst, u_int64_t imm64)
{ {
#ifndef PRODUCT #ifndef PRODUCT

39
hotspot/src/cpu/aarch64/vm/macroAssembler_aarch64.hpp
View File

@ -465,44 +465,7 @@ public:
void movptr(Register r, uintptr_t imm64); void movptr(Register r, uintptr_t imm64);
// Macro to mov replicated immediate to vector register. void mov(FloatRegister Vd, SIMD_Arrangement T, u_int32_t imm32);
// 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 // macro instructions for accessing and updating floating point
// status register // status register