ARM64: Consolidate 32/64-bit constant handling in assembler.

Thanks to Peter Cawley. #1065
This commit is contained in:
Mike Pall 2023-09-09 16:30:14 +02:00
parent dfc122e45c
commit 5149b0a3a2
2 changed files with 42 additions and 37 deletions

View File

@ -606,7 +606,11 @@ static Reg ra_allock(ASMState *as, intptr_t k, RegSet allow)
IRIns *ir = IR(ref); IRIns *ir = IR(ref);
if ((ir->o == IR_KINT64 && k == (int64_t)ir_kint64(ir)->u64) || if ((ir->o == IR_KINT64 && k == (int64_t)ir_kint64(ir)->u64) ||
#if LJ_GC64 #if LJ_GC64
#if LJ_TARGET_ARM64
(ir->o == IR_KINT && (uint64_t)k == (uint32_t)ir->i) ||
#else
(ir->o == IR_KINT && k == ir->i) || (ir->o == IR_KINT && k == ir->i) ||
#endif
(ir->o == IR_KGC && k == (intptr_t)ir_kgc(ir)) || (ir->o == IR_KGC && k == (intptr_t)ir_kgc(ir)) ||
((ir->o == IR_KPTR || ir->o == IR_KKPTR) && ((ir->o == IR_KPTR || ir->o == IR_KKPTR) &&
k == (intptr_t)ir_kptr(ir)) k == (intptr_t)ir_kptr(ir))

View File

@ -20,7 +20,7 @@ static uint64_t get_k64val(ASMState *as, IRRef ref)
} else { } else {
lj_assertA(ir->o == IR_KINT || ir->o == IR_KNULL, lj_assertA(ir->o == IR_KINT || ir->o == IR_KNULL,
"bad 64 bit const IR op %d", ir->o); "bad 64 bit const IR op %d", ir->o);
return ir->i; /* Sign-extended. */ return (uint32_t)ir->i; /* Zero-extended. */
} }
} }
@ -152,11 +152,10 @@ nopair:
/* Prefer rematerialization of BASE/L from global_State over spills. */ /* Prefer rematerialization of BASE/L from global_State over spills. */
#define emit_canremat(ref) ((ref) <= ASMREF_L) #define emit_canremat(ref) ((ref) <= ASMREF_L)
/* Try to find an N-step delta relative to other consts with N < lim. */ /* Try to find a one-step delta relative to other consts. */
static int emit_kdelta(ASMState *as, Reg rd, uint64_t k, int lim) static int emit_kdelta(ASMState *as, Reg rd, uint64_t k, int is64)
{ {
RegSet work = (~as->freeset & RSET_GPR) | RID2RSET(RID_GL); RegSet work = (~as->freeset & RSET_GPR) | RID2RSET(RID_GL);
if (lim <= 1) return 0; /* Can't beat that. */
while (work) { while (work) {
Reg r = rset_picktop(work); Reg r = rset_picktop(work);
IRRef ref = regcost_ref(as->cost[r]); IRRef ref = regcost_ref(as->cost[r]);
@ -165,13 +164,14 @@ static int emit_kdelta(ASMState *as, Reg rd, uint64_t k, int lim)
uint64_t kx = ra_iskref(ref) ? (uint64_t)ra_krefk(as, ref) : uint64_t kx = ra_iskref(ref) ? (uint64_t)ra_krefk(as, ref) :
get_k64val(as, ref); get_k64val(as, ref);
int64_t delta = (int64_t)(k - kx); int64_t delta = (int64_t)(k - kx);
if (!is64) delta = (int64_t)(int32_t)delta; /* Sign-extend. */
if (delta == 0) { if (delta == 0) {
emit_dm(as, A64I_MOVx, rd, r); emit_dm(as, is64|A64I_MOVw, rd, r);
return 1; return 1;
} else { } else {
uint32_t k12 = emit_isk12(delta < 0 ? (int64_t)(~(uint64_t)delta+1u) : delta); uint32_t k12 = emit_isk12(delta < 0 ? (int64_t)(~(uint64_t)delta+1u) : delta);
if (k12) { if (k12) {
emit_dn(as, (delta < 0 ? A64I_SUBx : A64I_ADDx)^k12, rd, r); emit_dn(as, (delta < 0 ? A64I_SUBw : A64I_ADDw)^is64^k12, rd, r);
return 1; return 1;
} }
/* Do other ops or multi-step deltas pay off? Probably not. /* Do other ops or multi-step deltas pay off? Probably not.
@ -184,51 +184,52 @@ static int emit_kdelta(ASMState *as, Reg rd, uint64_t k, int lim)
return 0; /* Failed. */ return 0; /* Failed. */
} }
static void emit_loadk(ASMState *as, Reg rd, uint64_t u64, int is64) static void emit_loadk(ASMState *as, Reg rd, uint64_t u64)
{ {
int i, zeros = 0, ones = 0, neg; int zeros = 0, ones = 0, neg, lshift = 0;
if (!is64) u64 = (int64_t)(int32_t)u64; /* Sign-extend. */ int is64 = (u64 >> 32) ? A64I_X : 0, i = is64 ? 4 : 2;
/* Count homogeneous 16 bit fragments. */ /* Count non-homogeneous 16 bit fragments. */
for (i = 0; i < 4; i++) { while (--i >= 0) {
uint64_t frag = (u64 >> i*16) & 0xffff; uint32_t frag = (u64 >> i*16) & 0xffff;
zeros += (frag == 0); zeros += (frag != 0);
ones += (frag == 0xffff); ones += (frag != 0xffff);
} }
neg = ones > zeros; /* Use MOVN if it pays off. */ neg = ones < zeros; /* Use MOVN if it pays off. */
if ((neg ? ones : zeros) < 3) { /* Need 2+ ins. Try shorter K13 encoding. */ if ((neg ? ones : zeros) > 1) { /* Need 2+ ins. Try 1 ins encodings. */
uint32_t k13 = emit_isk13(u64, is64); uint32_t k13 = emit_isk13(u64, is64);
if (k13) { if (k13) {
emit_dn(as, (is64|A64I_ORRw)^k13, rd, RID_ZERO); emit_dn(as, (is64|A64I_ORRw)^k13, rd, RID_ZERO);
return; return;
} }
} if (emit_kdelta(as, rd, u64, is64)) {
if (!emit_kdelta(as, rd, u64, 4 - (neg ? ones : zeros))) { return;
int shift = 0, lshift = 0;
uint64_t n64 = neg ? ~u64 : u64;
if (n64 != 0) {
/* Find first/last fragment to be filled. */
shift = (63-emit_clz64(n64)) & ~15;
lshift = emit_ctz64(n64) & ~15;
} }
/* MOVK requires the original value (u64). */
while (shift > lshift) {
uint32_t u16 = (u64 >> shift) & 0xffff;
/* Skip fragments that are correctly filled by MOVN/MOVZ. */
if (u16 != (neg ? 0xffff : 0))
emit_d(as, is64 | A64I_MOVKw | A64F_U16(u16) | A64F_LSL16(shift), rd);
shift -= 16;
}
/* But MOVN needs an inverted value (n64). */
emit_d(as, (neg ? A64I_MOVNx : A64I_MOVZx) |
A64F_U16((n64 >> lshift) & 0xffff) | A64F_LSL16(lshift), rd);
} }
if (neg) {
u64 = ~u64;
if (!is64) u64 = (uint32_t)u64;
}
if (u64) {
/* Find first/last fragment to be filled. */
int shift = (63-emit_clz64(u64)) & ~15;
lshift = emit_ctz64(u64) & ~15;
for (; shift > lshift; shift -= 16) {
uint32_t frag = (u64 >> shift) & 0xffff;
if (frag == 0) continue; /* Will be correctly filled by MOVN/MOVZ. */
if (neg) frag ^= 0xffff; /* MOVK requires the original value. */
emit_d(as, is64 | A64I_MOVKw | A64F_U16(frag) | A64F_LSL16(shift), rd);
}
}
/* But MOVN needs an inverted value. */
emit_d(as, is64 | (neg ? A64I_MOVNw : A64I_MOVZw) |
A64F_U16((u64 >> lshift) & 0xffff) | A64F_LSL16(lshift), rd);
} }
/* Load a 32 bit constant into a GPR. */ /* Load a 32 bit constant into a GPR. */
#define emit_loadi(as, rd, i) emit_loadk(as, rd, i, 0) #define emit_loadi(as, rd, i) emit_loadk(as, rd, (uint32_t)i)
/* Load a 64 bit constant into a GPR. */ /* Load a 64 bit constant into a GPR. */
#define emit_loadu64(as, rd, i) emit_loadk(as, rd, i, A64I_X) #define emit_loadu64(as, rd, i) emit_loadk(as, rd, i)
#define glofs(as, k) \ #define glofs(as, k) \
((intptr_t)((uintptr_t)(k) - (uintptr_t)&J2GG(as->J)->g)) ((intptr_t)((uintptr_t)(k) - (uintptr_t)&J2GG(as->J)->g))