ShuriZma
/
xenia-canary
mirror of https://github.com/xenia-canary/xenia-canary.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Added lea based multiply optimization for pow2+1 values below 16. Added magicgu2 division magic number optimization for INT_64 and INT_32. Added in my HIR optimization subpass but disabled all but 3 of the optimization rules.

This commit is contained in:
chss95cs@gmail.com 2020-01-14 20:48:18 -08:00 committed by illusion
parent 91c2811559
commit 77f2ff8b34
4 changed files with 1546 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

143
src/xenia/cpu/backend/x64/x64_sequences.cc
View File

@ -1419,6 +1419,13 @@ struct MUL_I16 : Sequence<MUL_I16, I<OPCODE_MUL, I16Op, I16Op, I16Op>> {
}; };
struct MUL_I32 : Sequence<MUL_I32, I<OPCODE_MUL, I32Op, I32Op, I32Op>> { struct MUL_I32 : Sequence<MUL_I32, I<OPCODE_MUL, I32Op, I32Op, I32Op>> {
static void Emit(X64Emitter& e, const EmitArgType& i) { static void Emit(X64Emitter& e, const EmitArgType& i) {
if (i.src2.is_constant) {
uint32_t multiplier = i.src2.value->constant.u32;
if (multiplier == 3 || multiplier == 5 || multiplier == 9) {
e.lea(i.dest, e.ptr[i.src1.reg() * (multiplier - 1) + i.src1.reg()]);
return;
}
}
if (e.IsFeatureEnabled(kX64EmitBMI2)) { if (e.IsFeatureEnabled(kX64EmitBMI2)) {
// mulx: $1:$2 = EDX * $3 // mulx: $1:$2 = EDX * $3
@ -1461,6 +1468,14 @@ struct MUL_I32 : Sequence<MUL_I32, I<OPCODE_MUL, I32Op, I32Op, I32Op>> {
}; };
struct MUL_I64 : Sequence<MUL_I64, I<OPCODE_MUL, I64Op, I64Op, I64Op>> { struct MUL_I64 : Sequence<MUL_I64, I<OPCODE_MUL, I64Op, I64Op, I64Op>> {
static void Emit(X64Emitter& e, const EmitArgType& i) { static void Emit(X64Emitter& e, const EmitArgType& i) {
if (i.src2.is_constant) {
uint64_t multiplier = i.src2.value->constant.u64;
if (multiplier == 3 || multiplier == 5 || multiplier == 9) {
e.lea(i.dest,
e.ptr[i.src1.reg() * ((int)multiplier - 1) + i.src1.reg()]);
return;
}
}
if (e.IsFeatureEnabled(kX64EmitBMI2)) { if (e.IsFeatureEnabled(kX64EmitBMI2)) {
// mulx: $1:$2 = RDX * $3 // mulx: $1:$2 = RDX * $3
@ -1718,7 +1733,62 @@ struct MUL_HI_I64
}; };
EMITTER_OPCODE_TABLE(OPCODE_MUL_HI, MUL_HI_I8, MUL_HI_I16, MUL_HI_I32, EMITTER_OPCODE_TABLE(OPCODE_MUL_HI, MUL_HI_I8, MUL_HI_I16, MUL_HI_I32,
MUL_HI_I64); MUL_HI_I64);
/* from Hackers Delight - by Henry S. Warren Jr. Calculate magic number for
* unsigned division */
template <typename T>
auto magicu(T d) {
constexpr unsigned NBITS = sizeof(T) * CHAR_BIT;
constexpr unsigned NBITS_M1 = NBITS - 1;
constexpr T SIGNBIT = T(1) << NBITS_M1;
constexpr T POSMASK = ~SIGNBIT;
struct mu {
T M; // Magic number,
int a; // "add" indicator,
int s;
}; // and shift amount.
// Must have 1 <= d <= 2**32-1.
int p, gt = 0;
T nc, delta, q1, r1, q2, r2;
struct mu magu;
magu.a = 0; // Initialize "add" indicator.
nc = -1 - ((T) - (std::make_signed_t<T>)d) % d; // Unsigned arithmetic here.
p = NBITS_M1; // Init. p.
q1 = SIGNBIT / nc; // Init. q1 = 2**p/nc.
r1 = SIGNBIT - q1 * nc; // Init. r1 = rem(2**p, nc).
q2 = POSMASK / d; // Init. q2 = (2**p - 1)/d.
r2 = POSMASK - q2 * d; // Init. r2 = rem(2**p - 1, d).
do {
p = p + 1;
if (q1 >= SIGNBIT) gt = 1; // Means q1 > delta.
if (r1 >= nc - r1) {
q1 = 2 * q1 + 1; // Update q1.
r1 = 2 * r1 - nc;
} // Update r1.
else {
q1 = 2 * q1;
r1 = 2 * r1;
}
if (r2 + 1 >= d - r2) {
if (q2 >= POSMASK) magu.a = 1;
q2 = 2 * q2 + 1; // Update q2.
r2 = 2 * r2 + 1 - d;
} // Update r2.
else {
if (q2 >= SIGNBIT) magu.a = 1;
q2 = 2 * q2;
r2 = 2 * r2 + 1;
}
delta = d - 1 - r2;
} while (gt == 0 && (q1 < delta || (q1 == delta && r1 == 0)));
magu.M = q2 + 1; // Magic number
magu.s = p - NBITS; // and shift amount to return
return magu; // (magu.a was set above).
}
// ============================================================================ // ============================================================================
// OPCODE_DIV // OPCODE_DIV
// ============================================================================ // ============================================================================
@ -1821,13 +1891,40 @@ struct DIV_I32 : Sequence<DIV_I32, I<OPCODE_DIV, I32Op, I32Op, I32Op>> {
if (i.src2.is_constant) { if (i.src2.is_constant) {
assert_true(!i.src1.is_constant); assert_true(!i.src1.is_constant);
e.mov(e.ecx, i.src2.constant());
if (i.instr->flags & ARITHMETIC_UNSIGNED) { if (i.instr->flags & ARITHMETIC_UNSIGNED) {
e.mov(e.eax, i.src1); auto div_shift_info = magicu<unsigned>(i.src2.value->constant.u32);
// Zero upper bits. // disabling addflag for now. the code for handling it is incorrect
e.xor_(e.edx, e.edx); if (!div_shift_info.a) {
e.div(e.ecx); e.mov(i.dest, i.src1);
if (e.IsFeatureEnabled(kX64EmitAVX2)) {
e.mov(e.edx, div_shift_info.M);
e.mulx(i.dest, i.dest, i.dest);
} else {
e.mov(e.eax, div_shift_info.M);
e.mul(i.dest);
e.mov(i.dest, e.edx);
}
if (div_shift_info.a) {
// pg 228, hackers delight, 2nd edition. add initial input to
// product, result may carry we need to handle this carry, treat
// result as 33 bits with carry as bit 33
e.add(i.dest, i.src1);
e.rcr(i.dest, 1);
}
if (div_shift_info.s != 0) {
e.shr(i.dest, div_shift_info.s);
}
return;
} else {
e.mov(e.ecx, i.src2.constant());
e.mov(e.eax, i.src1);
// Zero upper bits.
e.xor_(e.edx, e.edx);
e.div(e.ecx);
}
} else { } else {
e.mov(e.ecx, i.src2.constant());
e.mov(e.eax, i.src1); e.mov(e.eax, i.src1);
e.cdq(); // edx:eax = sign-extend eax e.cdq(); // edx:eax = sign-extend eax
e.idiv(e.ecx); e.idiv(e.ecx);
@ -1869,13 +1966,39 @@ struct DIV_I64 : Sequence<DIV_I64, I<OPCODE_DIV, I64Op, I64Op, I64Op>> {
if (i.src2.is_constant) { if (i.src2.is_constant) {
assert_true(!i.src1.is_constant); assert_true(!i.src1.is_constant);
e.mov(e.rcx, i.src2.constant());
if (i.instr->flags & ARITHMETIC_UNSIGNED) { if (i.instr->flags & ARITHMETIC_UNSIGNED) {
e.mov(e.rax, i.src1); auto div_shift_info = magicu<uint64_t>(i.src2.value->constant.u64);
// Zero upper bits. // disabled addflag for now
e.xor_(e.rdx, e.rdx); if (!div_shift_info.a) {
e.div(e.rcx); if (e.IsFeatureEnabled(kX64EmitAVX2)) {
e.mov(e.rdx, div_shift_info.M);
e.mov(i.dest, i.src1);
e.mulx(i.dest, i.dest, i.dest);
} else {
e.mov(e.rax, div_shift_info.M);
e.mul(i.src1);
e.mov(i.dest, e.rdx);
}
if (div_shift_info.a) {
// pg 228, hackers delight, 2nd edition. add initial input to
// product, result may carry we need to handle this carry, treat
// result as 65 bits with carry as bit 65
e.add(i.dest, i.src1);
e.rcr(i.dest, 1);
}
if (div_shift_info.s) {
e.shr(i.dest, div_shift_info.s);
}
return;
} else {
e.mov(e.rcx, i.src2.constant());
e.mov(e.rax, i.src1);
// Zero upper bits.
e.xor_(e.rdx, e.rdx);
e.div(e.rcx);
}
} else { } else {
e.mov(e.rcx, i.src2.constant());
e.mov(e.rax, i.src1); e.mov(e.rax, i.src1);
e.cqo(); // rdx:rax = sign-extend rax e.cqo(); // rdx:rax = sign-extend rax
e.idiv(e.rcx); e.idiv(e.rcx);

520
src/xenia/cpu/compiler/passes/peephole_rules_impl.h Normal file
View File

@ -0,0 +1,520 @@
#pragma once
static size_t n_cvt_opts = 0, n_vec_shift_opts = 0,
n_useless_ctx_store_opts = 0;
static bool do_cvt_opt(HIRBuilder* builder, Block* block) {
bool did_change = false;
for (auto* insn = block->instr_head; insn; insn = insn->next) {
if (!is_op(insn, OPCODE_CONVERT)) {
continue;
}
/*if (insn->dest->type != FLOAT32_TYPE ||
insn->src1.value->type != FLOAT64_TYPE) {
continue;
}*/
if (!is_rvalue(insn->dest) || !is_rvalue(insn->src1)) continue;
Instr* target_next_use = find_next_use(insn->dest, block, insn);
if (!target_next_use) {
continue;
}
if (!is_op(target_next_use, OPCODE_CONVERT)) continue;
if (target_next_use->src1.value != insn->dest) continue;
if (target_next_use->dest->type != insn->src1.value->type ||
!is_rvalue(target_next_use->dest))
continue;
++n_cvt_opts;
make_assignment(target_next_use, insn->src1);
did_change = true;
}
return did_change;
}
static bool do_vector_shift_opt(HIRBuilder* builder, Block* block) {
bool did_change = false;
for (auto* insn = block->instr_head; insn; insn = insn->next) {
if (!is_op(insn, OPCODE_LOAD_VECTOR_SHR) &&
!is_op(insn, OPCODE_LOAD_VECTOR_SHL)) {
continue;
}
if (!is_rvalue(insn->dest) || !is_rvalue(insn->src1)) continue;
for (Instr* nextu = find_next_use(insn->src1, block, insn); nextu;
nextu = find_next_use(insn->src1, block, nextu)) {
if (!nextu || nextu->opcode != insn->opcode) continue;
++n_vec_shift_opts;
make_assignment(nextu, insn->dest);
did_change = true;
}
}
return did_change;
}
template <bool creg_only = false>
static bool elim_useless_ctx_stores(HIRBuilder* builder, Block* block) {
bool did_change = false;
for (auto* insn = block->instr_head; insn; insn = insn->next) {
if (!is_op(insn, OPCODE_STORE_CONTEXT)) continue;
ppc_ctx_vset_t trace{};
add_store_to_bitset(&trace, insn);
if (creg_only && !has_any_creg(&trace)) {
continue;
}
contextuse_res_t redefed = USED;
auto next =
find_next_context_use<creg_only>(&trace, block, insn->next, &redefed);
if (redefed == USED && !next) {
bool all_redef = true;
for (auto succ = block->outgoing_edge_head; succ;
succ = succ->outgoing_next) {
if (find_next_context_use<creg_only>(&trace, succ->dest, nullptr,
&redefed) &&
redefed == REDEFED) {
} else {
all_redef = false;
break;
}
}
if (!all_redef) {
continue;
}
// continue;
} else if (redefed == INDET || redefed == USED) {
continue;
}
++n_useless_ctx_store_opts;
make_nop(insn);
did_change = true;
}
return did_change;
}
static uint64_t n_replaced_assignments = 0;
static bool replace_assignment_uses(HIRBuilder* builder, Block* block) {
bool did_change = false;
for (auto* insn = block->instr_head; insn; insn = insn->next) {
if (!is_op(insn, OPCODE_ASSIGN)) continue;
if (!is_rvalue(insn->dest) || !is_rvalue(insn->src1)) continue;
replace_uses(builder, insn->dest, insn->src1.value);
n_replaced_assignments++;
make_nop(insn);
did_change = true;
}
return did_change;
}
static uint64_t n_replaced_loads = 0;
static bool optimize_repeated_loads(HIRBuilder* builder, Block* block) {
bool did_change = false;
for (auto* insn = block->instr_head; insn; insn = insn->next) {
if (!is_op(insn, OPCODE_LOAD)) {
continue;
}
if (!is_rvalue(insn->src1)) continue;
auto n = find_next_use(insn->src1, block, insn);
if (!n) continue;
if (!is_op(n, OPCODE_LOAD)) continue;
if (insn->dest->type != n->dest->type) continue;
if (has_any_intervening_store(insn, n)) continue;
make_assignment(n, insn->dest);
n_replaced_loads++;
did_change = true;
}
return did_change;
}
static uint64_t n_and1s_optimized = 0;
static bool and1_optimizer(HIRBuilder* builder, Block* block) {
bool did_change = false;
for (auto* insn = block->instr_head; insn; insn = insn->next) {
if (!is_op(insn, OPCODE_IS_TRUE) &&
!(is_op(insn, OPCODE_COMPARE_EQ) && is_const_x(insn->src2, 1ULL)) &&
!(is_op(insn, OPCODE_COMPARE_NE) && is_const_x(insn->src2, 0ULL))) {
continue;
}
auto def = insn->src1.value->def;
if (def_is_onebit(def)) {
if (insn->src1.value->type == insn->dest->type) {
make_assignment(insn, insn->src1);
} else {
insn->Replace(&OPCODE_TRUNCATE_info, 0);
insn->set_src1(def->dest);
}
n_and1s_optimized++;
did_change = true;
}
}
return did_change;
}
static uint64_t n_trunc_exts = 0;
static bool trunc_ext_optimizer(HIRBuilder* builder, Block* block) {
bool did_change = false;
for (auto* insn = block->instr_head; insn; insn = insn->next) {
if (!is_op(insn, OPCODE_TRUNCATE)) continue;
auto def = insn->src1.value->def;
if (!is_op(def, OPCODE_ZERO_EXTEND) && !is_op(def, OPCODE_SIGN_EXTEND))
continue;
if (def->src1.value->type != insn->dest->type) continue;
make_assignment(insn, def->src1);
++n_trunc_exts;
did_change = true;
// __debugbreak();
}
return did_change;
}
static uint64_t n_eliminated_unused_conversion_results = 0;
static bool unused_conversion_eliminator(HIRBuilder* builder, Block* block) {
bool did_change = false;
for (auto* insn = block->instr_head; insn; insn = insn->next) {
if (!is_elimable(insn)) continue;
if (!insn->dest->use_head || insn->dest->use_head->instr == nullptr) {
make_nop(insn);
did_change = true;
n_eliminated_unused_conversion_results++;
}
}
return did_change;
}
static uint64_t n_useless_operations = 0;
static bool useless_operation_eliminator(HIRBuilder* builder, Block* block) {
bool did_change = false;
auto signal_change = [&did_change](auto insn, auto& src) {
make_assignment(insn, insn->src1);
n_useless_operations++;
did_change = true;
};
for (auto* insn = block->instr_head; insn; insn = insn->next) {
if (is_op(insn, OPCODE_ADD) || is_op(insn, OPCODE_SUB) ||
is_op(insn, OPCODE_OR) || is_op(insn, OPCODE_XOR)) {
if (is_const_x(insn->src2, 0)) {
signal_change(insn, insn->src1);
continue;
}
}
if (is_op(insn, OPCODE_AND) &&
is_const_x(insn->src2, mask_for_typename(insn->src1.value->type))) {
signal_change(insn, insn->src1);
continue;
}
if ((is_op(insn, OPCODE_DIV) || is_op(insn, OPCODE_MUL)) &&
is_const_x(insn->src2, 1ULL)) {
signal_change(insn, insn->src1);
continue;
}
}
return did_change;
}
static uint64_t n_rotates_lowered_to_shifts = 0;
static bool rlwinm_sanitizer(HIRBuilder* builder, Block* block) {
bool did_change = false;
for (auto* insn = block->instr_head; insn; insn = insn->next) {
if (!is_op(insn, OPCODE_ROTATE_LEFT)) continue;
if (!insn->src2.value->IsConstant()) continue;
auto defed_value = insn->dest;
auto nxtuse = get_solo_use(defed_value);
if (!nxtuse) continue;
if (!is_op(nxtuse, OPCODE_AND)) continue;
if (!nxtuse->src2.value->IsConstant()) continue;
uint64_t v = extract_constant(nxtuse->src2.value);
uint64_t rotamt = extract_constant(insn->src2.value);
if ((v & ((1ULL << rotamt) - 1)) == 0ULL) {
// the bits rotated in to the left are discarded, can be shl
insn->opcode = &OPCODE_SHL_info;
n_rotates_lowered_to_shifts++;
did_change = true;
}
}
return did_change;
}
static uint64_t n_useless_rlwinm_ors_removed = 0;
static bool rlwinm_sanitizer_useless_or(HIRBuilder* builder, Block* block) {
bool did_change = false;
for (auto* insn = block->instr_head; insn; insn = insn->next) {
if (!is_op(insn, OPCODE_SHL) || !insn->src2.value->IsConstant()) continue;
auto solo_use_shl_result = get_solo_use(insn->dest);
if (!solo_use_shl_result) continue;
if (!is_op(solo_use_shl_result, OPCODE_OR) ||
solo_use_shl_result->src1.value != insn->dest) {
continue;
}
auto solo_use_or_result = get_solo_use(solo_use_shl_result->dest);
if (!solo_use_or_result) continue;
auto mask = mask_for_typename(insn->dest->type);
mask <<= extract_constant(insn->src2.value);
/* if (is_op(solo_use_or_result, OPCODE_TRUNCATE)) {
auto newmask = mask_for_typename(solo_use_or_result->dest->type);
if ((newmask & mask) == 0) {
// none of our shl bits survive
make_nop(insn);
make_assignment(solo_use_or_result, solo_use_shl_result->src2);
n_useless_rlwinm_ors_removed++;
did_change = true;
}
}
else*/
if (is_op(solo_use_or_result, OPCODE_AND) &&
solo_use_or_result->src2.value->IsConstant()) {
auto newmask = extract_constant(solo_use_or_result->src2.value);
if ((newmask & mask) == 0) {
// none of our shl bits survive
make_nop(insn);
auto newandoper = solo_use_shl_result->src2.value;
// make_nop(solo_use_shl_result);
make_assignment(solo_use_shl_result, solo_use_shl_result->src2);
n_useless_rlwinm_ors_removed++;
did_change = true;
}
}
}
return did_change;
}
static uint64_t n_ssa_merges = 0;
static bool ssa_merger(HIRBuilder* builder, Block* block) {
bool did_change = false;
for (auto* insn = block->instr_head; insn; insn = insn->next) {
if (!is_op(insn, OPCODE_ASSIGN)) continue;
auto solo = get_solo_use(insn->dest);
if (!solo || !is_like_assign(solo)) continue;
solo->set_src1(insn->src1.value);
make_nop(insn);
++n_ssa_merges;
did_change = true;
}
return did_change;
}
static uint64_t n_signbit_rols_optimized = 0;
static bool signbit_rol_optimizer(HIRBuilder* builder, Block* block) {
bool did_change = false;
for (auto* insn = block->instr_head; insn; insn = insn->next) {
if (!is_op(insn, OPCODE_ROTATE_LEFT) || !is_const_x(insn->src2, 1ULL))
continue;
auto solo_use = get_solo_use(insn->dest);
if (!solo_use) continue;
if (!is_op(solo_use, OPCODE_AND) || !is_const_x(solo_use->src2, 1ULL))
continue;
Value* shift_constant = constant_for_type(
builder, highbit_for_typename(insn->dest->type), INT8_TYPE);
if (!shift_constant) continue;
solo_use->Replace(&OPCODE_SHR_info, 0);
solo_use->set_src1(insn->src1.value);
solo_use->set_src2(shift_constant);
make_nop(insn);
did_change = true;
n_signbit_rols_optimized++;
}
return did_change;
}
static uint64_t n_rol_bitextracts_optimized = 0;
static bool rol_bitextract_optimizer(HIRBuilder* builder, Block* block) {
bool did_change = false;
for (auto* insn = block->instr_head; insn; insn = insn->next) {
if (!is_op(insn, OPCODE_ROTATE_LEFT) || !insn->src2.value->IsConstant())
continue;
auto defuse = get_solo_use(insn->dest);
if (!defuse) continue;
if (!is_op(defuse, OPCODE_AND) || !is_const_x(defuse->src2, 1)) continue;
unsigned bitsz = highbit_for_typename(insn->dest->type);
uint64_t realtarget =
(bitsz + 1) - extract_constant(insn->src2.value);
auto old_inp = insn->src1.value;
Value* shrconst = constant_for_type(builder, realtarget, INT8_TYPE);
if (!shrconst) continue;
insn->Replace(&OPCODE_SHR_info, 0);
insn->set_src1(old_inp);
insn->set_src2(shrconst);
n_rol_bitextracts_optimized++;
did_change = true;
}
return did_change;
}
static uint64_t n_shl_shr_masks_gen = 0;
static bool shl_shr_mask_optimizer(HIRBuilder* builder, Block* block) {
bool did_change = false;
for (auto* insn = block->instr_head; insn; insn = insn->next) {
if (!is_op(insn, OPCODE_SHL) || !insn->src2.value->IsConstant()) continue;
auto defuse = get_solo_use(insn->dest);
if (!defuse || !is_op(defuse, OPCODE_SHR) ||
!insn->src2.value->IsConstantEQ(defuse->src2.value))
continue;
uint64_t mask = mask_for_typename(insn->src1.value->type);
uint64_t shfactor = extract_constant(insn->src2.value);
uint64_t msk = mask;
mask <<= shfactor;
mask >>= shfactor;
mask &= msk;
auto oldsrc1 = insn->src1.value;
auto maskval = constant_for_type(builder, mask, oldsrc1->type);
if (!maskval) continue;
make_nop(insn);
defuse->Replace(&OPCODE_AND_info, 0);
defuse->set_src1(oldsrc1);
defuse->set_src2(maskval);
n_shl_shr_masks_gen++;
did_change = true;
}
return did_change;
}
static uint64_t n_redundant_loads_local = 0;
static bool redundant_local_slot_load_optimizer(HIRBuilder* builder,
Block* block) {
bool did_change = false;
for (auto* insn = block->instr_head; insn; insn = insn->next) {
if (!is_op(insn, OPCODE_LOAD_LOCAL)) {
continue;
}
auto next_load = find_next_local_load(insn->src1.value, block, insn);
if (!next_load) continue;
make_assignment(next_load, insn->dest);
++n_redundant_loads_local;
did_change = true;
}
return did_change;
}
/*
Re-enable these one by one or add flags to control them until each is proven to be stable
*/
#if 0
static const optblock_pass_t g_passes[] = {do_cvt_opt,
do_vector_shift_opt,
elim_useless_ctx_stores<false>,
replace_assignment_uses,
elim_useless_ctx_stores<true>,
optimize_repeated_loads,
and1_optimizer,
trunc_ext_optimizer,
unused_conversion_eliminator,
useless_operation_eliminator,
rlwinm_sanitizer,
rlwinm_sanitizer_useless_or,
ssa_merger,
signbit_rol_optimizer,
rol_bitextract_optimizer,
shl_shr_mask_optimizer,
redundant_local_slot_load_optimizer,
nop_deleter};
void dump_opts() {
FILE* lel = fopen("Optdump.txt", "w");
fprintf(
lel,
"Optimized %lld converts\nOptimized %lld vector shifts\nOptimized %lld "
"context stores\nOptimized away %lld useless assignments\nOptimized "
"away "
"%lld redundant loads."
"\nOptimized away %lld and 1 instructions\nOptimized away %lld useless "
"truncate-extends\nEliminated %lld unused operation "
"results.\nOptimized away %lld no-op operations.\nLowered %lld left "
"rotates to left shifts.\nRemoved %lld useless rlwinm rotate-or "
"sequences.\nShortened %lld assignment chains\nOptimized %lld signbit "
"rotate lefts.\nOptimized %lld rol bitextracts.\nConverted %lld shift "
"sequences to bitmasks.\n Optimized %lld redundant local loads.",
n_cvt_opts, n_vec_shift_opts, n_useless_ctx_store_opts,
n_replaced_assignments, n_replaced_loads, n_and1s_optimized, n_trunc_exts,
n_eliminated_unused_conversion_results, n_useless_operations,
n_rotates_lowered_to_shifts, n_useless_rlwinm_ors_removed, n_ssa_merges,
n_signbit_rols_optimized, n_rol_bitextracts_optimized,
n_shl_shr_masks_gen, n_redundant_loads_local);
fclose(lel);
}
#else
static const optblock_pass_t g_passes[] = {rlwinm_sanitizer_useless_or,
rlwinm_sanitizer,
nop_deleter,
rol_bitextract_optimizer,
signbit_rol_optimizer,
shl_shr_mask_optimizer
};
#endif

857
src/xenia/cpu/compiler/passes/repetitive_computation_merger_pass.cc Normal file
View File

@ -0,0 +1,857 @@
#include "repetitive_computation_merger_pass.h"
#include <bitset>
#include "xenia/cpu/ppc/ppc_context.h"
namespace xe {
namespace cpu {
namespace compiler {
namespace passes {
using namespace xe::cpu::hir;
static bool didit = false;
using xe::cpu::hir::HIRBuilder;
using xe::cpu::hir::Instr;
using xe::cpu::hir::Value;
using ppc_ctx_vset_underlying_t = std::bitset<sizeof(ppc::PPCContext)>;
struct alignas(__m256i) ppc_ctx_vset_t : public ppc_ctx_vset_underlying_t {
constexpr ppc_ctx_vset_t() : ppc_ctx_vset_underlying_t(), pad(){};
char pad[sizeof(ppc_ctx_vset_underlying_t) % sizeof(__m256i)];
};
#ifdef _MSC_VER
#define add_ppc_offset(field) \
__builtin_offsetof(ppc::PPCContext, field), \
__builtin_offsetof(ppc::PPCContext, field) + 1, \
__builtin_offsetof(ppc::PPCContext, field) + 2, \
__builtin_offsetof(ppc::PPCContext, field) + 3
#else
#define add_ppc_offset(field) offsetof(ppc::PPCContext, field)
#endif
constexpr unsigned bitset_cregs[] = {add_ppc_offset(cr0), add_ppc_offset(cr1),
add_ppc_offset(cr2), add_ppc_offset(cr3),
add_ppc_offset(cr4), add_ppc_offset(cr5),
add_ppc_offset(cr6), add_ppc_offset(cr7)};
RepetitiveComputationMergerPass::RepetitiveComputationMergerPass() {}
RepetitiveComputationMergerPass::~RepetitiveComputationMergerPass() {}
static Instr* create_instr(hir::HIRBuilder* builder, hir::Block* blk,
const hir::OpcodeInfo* opcode, unsigned flags,
Value* dest) {
Instr* instr = builder->arena()->Alloc<Instr>();
instr->next = NULL;
instr->prev = nullptr;
instr->ordinal = UINT32_MAX;
instr->block = blk;
instr->opcode = opcode;
instr->flags = flags;
instr->dest = dest;
instr->src1.value = instr->src2.value = instr->src3.value = NULL;
instr->src1_use = instr->src2_use = instr->src3_use = NULL;
if (dest) {
dest->def = instr;
}
return instr;
}
static uint64_t extract_constant(Value* v) {
switch (v->type) {
case INT8_TYPE:
return v->constant.u8;
case INT16_TYPE:
return v->constant.u16;
case INT32_TYPE:
return v->constant.u32;
case INT64_TYPE:
return v->constant.u64;
}
return 0;
}
static Instr* find_next_use(Value* val, hir::Block* blk, Instr* after,
Instr* end = nullptr) {
if (!after || after->next == end) {
return nullptr;
}
for (Instr* pos = after->next; pos != end; pos = pos->next) {
if (pos->opcode->flags & OPCODE_FLAG_VOLATILE) return nullptr;
if (pos->src1.value == val || pos->src2.value == val ||
pos->src3.value == val) {
return pos;
}
}
return nullptr;
}
static Instr* find_next_use(Instr::Op& val, hir::Block* blk, Instr* after,
Instr* end = nullptr) {
return find_next_use(val.value, blk, after, end);
}
static Instr* get_solo_use(Value* val) {
if (!val || !val->use_head || val->use_head->next) {
return nullptr;
}
return val->use_head->instr;
}
static void insert_after(Instr* i, Instr* after) {
after->prev = i;
after->next = i->next;
i->next = after;
if (after->next) {
after->next->prev = after;
}
}
static bool is_before_in_block(Instr* first, Instr* second, hir::Block* blk) {
for (auto s = blk->instr_head; s; s = s->next) {
if (s == first) {
return true;
} else if (s == second) {
return false;
}
}
return false;
}
static unsigned size_for_typename(TypeName type) {
switch (type) {
case INT8_TYPE:
return 1;
case INT16_TYPE:
return 2;
case INT32_TYPE:
return 4;
case INT64_TYPE:
return 8;
case FLOAT32_TYPE:
return 4;
case FLOAT64_TYPE:
return 8;
case VEC128_TYPE:
return 16;
}
xenia_assert(false);
return 0;
}
static Value* constant_for_type(HIRBuilder* builder, uint64_t v,
TypeName type) {
switch (type) {
case INT8_TYPE:
return builder->LoadConstantUint8((uint8_t)v);
case INT16_TYPE:
return builder->LoadConstantUint16((uint16_t)v);
case INT32_TYPE:
return builder->LoadConstantUint32((uint32_t)v);
case INT64_TYPE:
return builder->LoadConstantUint64((uint64_t)v);
}
return nullptr;
}
static unsigned bitsize_for_typename(TypeName type) {
return size_for_typename(type) * CHAR_BIT;
}
static unsigned highbit_for_typename(TypeName type) {
return bitsize_for_typename(type) - 1;
}
static uint64_t mask_for_typename(TypeName type) {
unsigned v = size_for_typename(type);
xenia_assert(v <= 8);
return (1ULL << ((CHAR_BIT * v))) - 1;
}
static bool is_const_x(Value* v, uint64_t val) {
return v->IsConstant() && extract_constant(v) == val;
}
static bool is_const_x(const Instr::Op& v, uint64_t val) {
return is_const_x(v.value, val);
}
static bool is_op(Instr* insn, Opcode opcode) {
return insn && insn->opcode && insn->opcode->num == opcode;
}
static bool has_any_intervening_memop(Instr* start, Instr* end) {
for (auto pos = start->next; pos != end; pos = pos->next) {
if (pos->opcode->flags & OPCODE_FLAG_MEMORY) return true;
}
return false;
}
static bool has_any_intervening_store(Instr* start, Instr* end) {
for (auto pos = start->next; pos != end; pos = pos->next) {
if (is_op(pos, OPCODE_STORE)) return true;
}
return false;
}
static bool is_rvalue(const Value* v) { return v && v->local_slot == nullptr; }
static bool is_rvalue(const Instr::Op& v) {
return v.value && v.value->local_slot == nullptr;
}
static void make_assignment(Instr* insn, Value* from) {
insn->Replace(&OPCODE_ASSIGN_info, 0);
insn->set_src1(from);
}
static void make_assignment(Instr* insn, Instr::Op& from) {
make_assignment(insn, from.value);
}
static void add_store_to_bitset(ppc_ctx_vset_t* out, Instr* insn) {
out->set(insn->src1.offset);
}
static bool load_target_in_bitset(const ppc_ctx_vset_t* to_test, Instr* insn) {
if (!is_op(insn, OPCODE_LOAD_CONTEXT)) {
return false;
}
return to_test->test(insn->src1.offset);
}
static bool store_target_in_bitset(const ppc_ctx_vset_t* to_test, Instr* insn) {
if (!is_op(insn, OPCODE_STORE_CONTEXT)) {
return false;
}
return to_test->test(insn->src1.offset);
}
static bool has_any_creg(const ppc_ctx_vset_t* b) {
for (auto&& bt : bitset_cregs) {
if (b->test(bt)) return true;
}
return false;
}
static Instr* find_next_local_store(Value* slot, hir::Block* blk, Instr* after,
Instr* end = nullptr) {
if (!after || after->next == end) {
return nullptr;
}
for (Instr* pos = after->next; pos != end; pos = pos->next) {
if (is_op(pos, OPCODE_STORE_LOCAL) && pos->src1.value == slot) {
return pos;
}
}
return nullptr;
}
static bool is_like_assign(Instr* insn) {
return is_op(insn, OPCODE_ASSIGN) || is_op(insn, OPCODE_SIGN_EXTEND) ||
is_op(insn, OPCODE_ZERO_EXTEND);
}
static bool produces_boolean(Instr* i) {
constexpr Opcode opcodes[] = {
OPCODE_IS_FALSE, OPCODE_IS_TRUE, OPCODE_COMPARE_EQ,
OPCODE_COMPARE_NE, OPCODE_COMPARE_SLT, OPCODE_COMPARE_SLE,
OPCODE_COMPARE_SGT, OPCODE_COMPARE_SGE, OPCODE_COMPARE_ULT,
OPCODE_COMPARE_ULE, OPCODE_COMPARE_UGT, OPCODE_COMPARE_UGE,
OPCODE_DID_SATURATE,
};
for (auto&& op : opcodes) {
if (is_op(i, op)) return true;
}
return false;
}
static void replace_uses(HIRBuilder* builder, Value* to_replace, Value* with) {
for (auto start = to_replace->use_head; start; start = start->next) {
auto i = start->instr;
if (i->src1.value == to_replace) {
i->set_src1(with);
}
if (i->src2.value == to_replace) {
i->set_src2(with);
}
if (i->src3.value == to_replace) {
i->set_src3(with);
}
}
}
static bool is_any_call(Instr* i) {
return is_op(i, OPCODE_CALL) || is_op(i, OPCODE_CALL_EXTERN) ||
is_op(i, OPCODE_CALL_INDIRECT) ||
is_op(i, OPCODE_CALL_INDIRECT_TRUE) || is_op(i, OPCODE_CALL_TRUE) ||
is_op(i, OPCODE_DEBUG_BREAK) || is_op(i, OPCODE_DEBUG_BREAK_TRUE) ||
is_op(i, OPCODE_RETURN);
}
static void make_nop(Instr* i) {
i->Replace(&OPCODE_NOP_info, 0);
// i->Remove();
}
static Instr* find_next_local_load(Value* slot, hir::Block* blk, Instr* after,
Instr* end = nullptr) {
if (!after || after->next == end) {
return nullptr;
}
for (Instr* pos = after->next; pos != end; pos = pos->next) {
if (is_op(pos, OPCODE_LOAD_LOCAL) && pos->src1.value == slot) {
return pos;
} else if (is_op(pos, OPCODE_STORE_LOCAL) && pos->src1.value == slot) {
return nullptr; // redefed b
}
}
return nullptr;
}
enum contextuse_res_t { USED, REDEFED, INDET };
template <bool break_on_call = false>
static Instr* find_next_context_use(const ppc_ctx_vset_t* val, hir::Block* blk,
Instr* start, contextuse_res_t* redefed,
Instr* end = nullptr) {
if (!start) {
start = blk->instr_head;
}
*redefed = USED;
for (Instr* pos = start; pos != end; pos = pos->next) {
if (pos->opcode->flags & OPCODE_FLAG_VOLATILE) {
*redefed = INDET;
return nullptr;
}
if (break_on_call && is_any_call(pos)) {
*redefed = REDEFED;
return nullptr;
}
if (load_target_in_bitset(val, pos)) {
return pos;
} else if (store_target_in_bitset(val, pos)) {
*redefed = REDEFED;
return nullptr;
}
}
return nullptr;
}
static bool def_is_onebit(Instr* def) {
if (((is_op(def, OPCODE_AND) || is_op(def, OPCODE_XOR)) &&
def->src2.value->IsConstant() && def->src2.value->constant.u64 == 1) ||
produces_boolean(def)) {
return true;
}
if (is_op(def, OPCODE_ASSIGN)) {
return def_is_onebit(def->src1.value->def);
}
if (is_op(def, OPCODE_AND) || is_op(def, OPCODE_OR) ||
is_op(def, OPCODE_XOR) || is_op(def, OPCODE_SHR) ||
is_op(def, OPCODE_MUL)) {
return def_is_onebit(def->src1.value->def) &&
def_is_onebit(def->src2.value->def);
}
if (is_op(def, OPCODE_ZERO_EXTEND)) {
return def_is_onebit(def->src1.value->def);
}
if (is_op(def, OPCODE_SIGN_EXTEND)) {
return def_is_onebit(def->src1.value->def);
}
if (is_op(def, OPCODE_LOAD_CONTEXT) && def->dest->type == INT8_TYPE) {
ppc_ctx_vset_t bset{};
bset.set(def->src1.offset);
if (has_any_creg(&bset)) {
return true;
}
}
return false;
}
static constexpr Opcode elimable_results[] = {
OPCODE_ADD, OPCODE_SUB, OPCODE_ASSIGN,
OPCODE_BYTE_SWAP, OPCODE_OR, OPCODE_XOR,
OPCODE_AND, OPCODE_NOT, OPCODE_NEG,
OPCODE_MUL, OPCODE_CONVERT, OPCODE_IS_FALSE,
OPCODE_IS_TRUE, OPCODE_COMPARE_EQ, OPCODE_COMPARE_NE,
OPCODE_COMPARE_SLT, OPCODE_COMPARE_SLE, OPCODE_COMPARE_SGT,
OPCODE_COMPARE_SGE, OPCODE_COMPARE_ULT, OPCODE_COMPARE_ULE,
OPCODE_COMPARE_UGT, OPCODE_COMPARE_UGE, OPCODE_DID_SATURATE,
OPCODE_DIV, OPCODE_SHL, OPCODE_SHR,
OPCODE_SHA, OPCODE_ABS, OPCODE_ADD_CARRY,
OPCODE_ZERO_EXTEND, OPCODE_SIGN_EXTEND, OPCODE_TRUNCATE,
OPCODE_CNTLZ, OPCODE_LOAD_LOCAL, OPCODE_LOAD,
OPCODE_LOAD_CONTEXT, OPCODE_LOG2, OPCODE_SQRT,
OPCODE_ROTATE_LEFT, OPCODE_ROUND
};
static bool is_elimable(Instr* i) {
for (auto&& opcode : elimable_results) {
if (is_op(i, opcode)) return true;
}
return false;
}
static bool nop_deleter(HIRBuilder* builder, Block* block) {
#if 0
Instr* pos = block->instr_head;
bool did_change = false;
while (pos) {
Instr* nextdude = pos->next;
if (is_op(pos, OPCODE_NOP)) {
pos->Remove();
did_change = true;
}
pos = nextdude;
}
return did_change;
#else
return false;
#endif
}
class optrule_t {
protected:
HIRBuilder* builder;
Block* block;
uint64_t n_runs;
public:
optrule_t() : builder(nullptr), block(nullptr), n_runs(0) {}
virtual const char* name() = 0;
virtual bool run(Instr* insn) = 0;
uint64_t get_execs() const { return n_runs; }
bool exec(HIRBuilder* blder, Block* blk);
};
bool optrule_t::exec(HIRBuilder* blder, Block* blk) {
bool did_change = false;
block = blk;
builder = blder;
for (auto* insn = block->instr_head; insn; insn = insn->next) {
if (run(insn)) {
did_change = true;
++n_runs;
}
}
return did_change;
}
#define DECL_RULE(typenam, printname) \
class typenam##_t : public optrule_t { \
virtual const char* name() override { return printname; } \
virtual bool run(Instr* insn); \
}; \
static typenam##_t typenam{};
#define IMPL_RULE(typenam) bool typenam##_t ::run(Instr* insn)
#define RULE(typenam, printname) \
DECL_RULE(typenam, printname); \
IMPL_RULE(typenam)
using optblock_pass_t = bool (*)(HIRBuilder*, Block*);
#if 1
#include "peephole_rules_impl.h"
#else
RULE(shr_shl_mask,
"Convert shift-left followed by shift-right to and constant") {
if (!is_op(insn, OPCODE_SHL) || !insn->src2.value->IsConstant()) return false;
auto defuse = get_solo_use(insn->dest);
if (!defuse || !is_op(defuse, OPCODE_SHR) ||
!insn->src2.value->IsConstantEQ(defuse->src2.value))
return false;
uint64_t mask = mask_for_typename(insn->src1.value->type);
uint64_t shfactor = extract_constant(insn->src2.value);
uint64_t msk = mask;
mask <<= shfactor;
mask >>= shfactor;
mask &= msk;
auto oldsrc1 = insn->src1.value;
auto maskval = constant_for_type(builder, mask, oldsrc1->type);
if (!maskval) return false;
make_nop(insn);
defuse->Replace(&OPCODE_AND_info, 0);
defuse->set_src1(oldsrc1);
defuse->set_src2(maskval);
return true;
}
RULE(cvt_op_optimizer, "Remove redundant conversion pair") {
if (!is_op(insn, OPCODE_CONVERT)) {
return false;
}
if (!is_rvalue(insn->dest) || !is_rvalue(insn->src1)) return false;
Instr* target_next_use = find_next_use(insn->dest, block, insn);
if (!target_next_use) {
return false;
}
if (!is_op(target_next_use, OPCODE_CONVERT)) return false;
if (target_next_use->src1.value != insn->dest) return false;
if (target_next_use->dest->type != insn->src1.value->type ||
!is_rvalue(target_next_use->dest))
return false;
make_assignment(target_next_use, insn->src1);
return true;
}
RULE(vector_shift_op, "Reuse load vector shift results") {
if (!is_op(insn, OPCODE_LOAD_VECTOR_SHR) &&
!is_op(insn, OPCODE_LOAD_VECTOR_SHL)) {
return false;
}
if (!is_rvalue(insn->dest) || !is_rvalue(insn->src1)) return false;
bool did_change = false;
for (Instr* nextu = find_next_use(insn->src1, block, insn); nextu;
nextu = find_next_use(insn->src1, block, nextu)) {
if (!nextu || nextu->opcode != insn->opcode) return false;
make_assignment(nextu, insn->dest);
did_change = true;
}
return did_change;
}
RULE(useless_ctx_store_elim, "Eliminate useless context stores") {
if (!is_op(insn, OPCODE_STORE_CONTEXT)) return false;
ppc_ctx_vset_t trace{};
add_store_to_bitset(&trace, insn);
if (true && !has_any_creg(&trace)) {
return false;
}
contextuse_res_t redefed = USED;
auto next =
find_next_context_use<true>(&trace, block, insn->next, &redefed);
if (redefed == USED && !next) {
bool all_redef = true;
for (auto succ = block->outgoing_edge_head; succ;
succ = succ->outgoing_next) {
if (find_next_context_use<true>(&trace, succ->dest, nullptr,
&redefed) &&
redefed == REDEFED) {
} else {
all_redef = false;
break;
}
}
if (!all_redef) {
return false;
}
// continue;
} else if (redefed == INDET || redefed == USED) {
return false;
}
make_nop(insn);
return true;
}
RULE(replace_assignment_uses, "Replace rvalue assignment uses") {
if (!is_op(insn, OPCODE_ASSIGN)) return false;
if (!is_rvalue(insn->dest) || !is_rvalue(insn->src1)) return false;
replace_uses(builder, insn->dest, insn->src1.value);
make_nop(insn);
return true;
}
RULE(optimize_rep_loads, "Remove repeated loads") {
if (!is_op(insn, OPCODE_LOAD)) {
return false;
}
if (!is_rvalue(insn->src1)) return false;
auto n = find_next_use(insn->src1, block, insn);
if (!n) return false;
if (!is_op(n, OPCODE_LOAD)) return false;
if (insn->dest->type != n->dest->type) return false;
if (has_any_intervening_store(insn, n)) return false;
make_assignment(n, insn->dest);
return true;
}
RULE(optimize_and1, "Replace onebit istrue evaluation with evaluated term") {
if (!is_op(insn, OPCODE_IS_TRUE) &&
!(is_op(insn, OPCODE_COMPARE_EQ) && is_const_x(insn->src2, 1ULL)) &&
!(is_op(insn, OPCODE_COMPARE_NE) && is_const_x(insn->src2, 0ULL))) {
return false;
}
auto def = insn->src1.value->def;
if (def_is_onebit(def)) {
if (insn->src1.value->type == insn->dest->type) {
make_assignment(insn, insn->src1);
} else {
insn->Replace(&OPCODE_TRUNCATE_info, 0);
insn->set_src1(def->dest);
}
return true;
}
return false;
}
RULE(truncate_extend_opt, "Remove useless truncate-extends") {
if (!is_op(insn, OPCODE_TRUNCATE)) return false;
auto def = insn->src1.value->def;
if (!is_op(def, OPCODE_ZERO_EXTEND) && !is_op(def, OPCODE_SIGN_EXTEND))
return false;
if (def->src1.value->type != insn->dest->type) return false;
make_assignment(insn, def->src1);
return true;
}
RULE(elim_unused_opres, "Eliminate operations with unused results") {
if (!is_elimable(insn)) return false;
if (!insn->dest->use_head || insn->dest->use_head->instr == nullptr) {
make_nop(insn);
return true;
}
return false;
}
RULE(elim_useless_op, "Eliminate useless operations") {
if (is_op(insn, OPCODE_ADD) || is_op(insn, OPCODE_SUB) ||
is_op(insn, OPCODE_OR) || is_op(insn, OPCODE_XOR)) {
if (is_const_x(insn->src2, 0)) {
make_assignment(insn, insn->src1);
return true;
}
}
if (is_op(insn, OPCODE_AND) &&
is_const_x(insn->src2, mask_for_typename(insn->src1.value->type))) {
make_assignment(insn, insn->src1);
return true;
}
if ((is_op(insn, OPCODE_DIV) || is_op(insn, OPCODE_MUL)) &&
is_const_x(insn->src2, 1ULL)) {
make_assignment(insn, insn->src1);
return true;
}
return false;
}
RULE(rol_to_shl, "Convert rotates to shifts") {
if (!is_op(insn, OPCODE_ROTATE_LEFT)) return false;
if (!insn->src2.value->IsConstant()) return false;
auto defed_value = insn->dest;
auto nxtuse = get_solo_use(defed_value);
if (!nxtuse) return false;
if (!is_op(nxtuse, OPCODE_AND)) return false;
if (!nxtuse->src2.value->IsConstant()) return false;
uint64_t v = extract_constant(nxtuse->src2.value);
uint64_t rotamt = extract_constant(insn->src2.value);
if ((v & ((1ULL << rotamt) - 1)) == 0ULL) {
// the bits rotated in to the left are discarded, can be shl
insn->opcode = &OPCODE_SHL_info;
return true;
}
return false;
}
RULE(remove_useless_or, "Remove useless or if bits are not used") {
if (!is_op(insn, OPCODE_SHL) || !insn->src2.value->IsConstant()) return false;
auto solo_use_shl_result = get_solo_use(insn->dest);
if (!solo_use_shl_result) return false;
if (!is_op(solo_use_shl_result, OPCODE_OR) ||
solo_use_shl_result->src1.value != insn->dest) {
return false;
}
auto solo_use_or_result = get_solo_use(solo_use_shl_result->dest);
if (!solo_use_or_result) return false;
auto mask = mask_for_typename(insn->dest->type);
mask <<= extract_constant(insn->src2.value);
if (is_op(solo_use_or_result, OPCODE_AND) &&
solo_use_or_result->src2.value->IsConstant()) {
auto newmask = extract_constant(solo_use_or_result->src2.value);
if ((newmask & mask) == 0) {
// none of our shl bits survive
make_nop(insn);
auto newandoper = solo_use_shl_result->src2.value;
// make_nop(solo_use_shl_result);
make_assignment(solo_use_shl_result, solo_use_shl_result->src2);
return true;
}
}
return false;
}
RULE(signbit_rol_opt, "Optimize rotates to extract signbit") {
if (!is_op(insn, OPCODE_ROTATE_LEFT) || !is_const_x(insn->src2, 1ULL))
return false;
auto solo_use = get_solo_use(insn->dest);
if (!solo_use) return false;
if (!is_op(solo_use, OPCODE_AND) || !is_const_x(solo_use->src2, 1ULL))
return false;
Value* shift_constant = constant_for_type(
builder, highbit_for_typename(insn->dest->type), INT8_TYPE);
if (!shift_constant) return false;
solo_use->Replace(&OPCODE_SHR_info, 0);
solo_use->set_src1(insn->src1.value);
solo_use->set_src2(shift_constant);
make_nop(insn);
return true;
}
RULE(bitextract_by_rol_opt, "Optimize rotates intended to extract one bit") {
if (!is_op(insn, OPCODE_ROTATE_LEFT) || !insn->src2.value->IsConstant())
return false;
auto defuse = get_solo_use(insn->dest);
if (!defuse) return false;
if (!is_op(defuse, OPCODE_AND) || !is_const_x(defuse->src2, 1)) return false;
unsigned bitsz = highbit_for_typename(insn->dest->type);
uint64_t realtarget =
(bitsz + 1) - extract_constant(insn->src2.value);
auto old_inp = insn->src1.value;
Value* shrconst = constant_for_type(builder, realtarget, INT8_TYPE);
if (!shrconst) return false;
insn->Replace(&OPCODE_SHR_info, 0);
insn->set_src1(old_inp);
insn->set_src2(shrconst);
return true;
}
RULE(optimize_redundant_local_load, "Optimize away redundant local loads") {
if (!is_op(insn, OPCODE_LOAD_LOCAL)) {
return false;
}
auto next_load = find_next_local_load(insn->src1.value, block, insn);
if (!next_load) return false;
make_assignment(next_load, insn->dest);
return true;
}
static optrule_t* g_optrules[] = {&shr_shl_mask, &cvt_op_optimizer, &vector_shift_op,
&useless_ctx_store_elim, &replace_assignment_uses, &optimize_rep_loads,
&optimize_and1, &truncate_extend_opt, &elim_unused_opres,
&elim_useless_op, &remove_useless_or, &signbit_rol_opt,
&bitextract_by_rol_opt,
&optimize_redundant_local_load};
static bool run_optrules(HIRBuilder* builder, Block* block) {
bool did_change = false;
for (auto&& rule : g_optrules) {
did_change |= rule->exec(builder, block);
}
return did_change;
}
void dump_opts() {
FILE* lel = fopen("Optdump.txt", "w");
for (auto&& rule : g_optrules) {
fprintf(lel, "(%s) matched and optimized %lld times.\n", rule->name(),
rule->get_execs());
}
fclose(lel);
}
static const optblock_pass_t g_passes[] = {run_optrules, nop_deleter};
#endif
static bool did_atexit = false;
bool RepetitiveComputationMergerPass::RunPerBlock(hir::HIRBuilder* builder,
hir::Block* block) {
bool did_change = false;
for (auto&& pass : g_passes) {
did_change |= pass(builder, block);
}
return did_change;
}
bool RepetitiveComputationMergerPass::Run(hir::HIRBuilder* builder) {
if (!did_atexit) {
// atexit(dump_opts);
did_atexit = true;
}
auto block = builder->first_block();
while (block) {
// rerun until no changes
while (RunPerBlock(builder, block))
;
block = block->next;
}
return true;
}
} // namespace passes
} // namespace compiler
} // namespace cpu
} // namespace xe

36
src/xenia/cpu/compiler/passes/repetitive_computation_merger_pass.h Normal file
View File

@ -0,0 +1,36 @@
/**
******************************************************************************
* Xenia : Xbox 360 Emulator Research Project *
******************************************************************************
* Copyright 2015 Ben Vanik. All rights reserved. *
* Released under the BSD license - see LICENSE in the root for more details. *
******************************************************************************
*/
#ifndef XENIA_CPU_COMPILER_PASSES_REPETITIVE_COMPUTATION_MERGER_PASS_H_
#define XENIA_CPU_COMPILER_PASSES_REPETITIVE_COMPUTATION_MERGER_PASS_H_
#include "xenia/cpu/compiler/compiler_pass.h"
namespace xe {
namespace cpu {
namespace compiler {
namespace passes {
class RepetitiveComputationMergerPass : public CompilerPass {
public:
RepetitiveComputationMergerPass();
~RepetitiveComputationMergerPass() override;
bool Run(hir::HIRBuilder* builder) override;
private:
bool RunPerBlock(hir::HIRBuilder* builder, hir::Block* block);
};
} // namespace passes
} // namespace compiler
} // namespace cpu
} // namespace xe
#endif // XENIA_CPU_COMPILER_PASSES_REPETITIVE_COMPUTATION_MERGER_PASS_H_