RPCS3/rpcs3
1
0
Fork 0
mirror of https://github.com/RPCS3/rpcs3.git synced 2024-11-26 04:32:35 +01:00
Code Issues Releases Wiki Activity

SPU LLVM: Optimize branches following byteswaps

Browse Source 
- The first element can be extracted via vmovd rather than vpextrd, which saves 1 uop.
This commit is contained in:
Malcolm Jestadt 2021-09-29 05:09:42 -04:00 committed by Ivan
parent f9ab077908
commit 86716dc37b
1 changed files with 59 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

60
rpcs3/Emu/Cell/SPURecompiler.cpp
View File

@ -9232,6 +9232,20 @@ public:
{
if (m_block) m_block->block_end = m_ir->GetInsertBlock();
const auto rt = get_vr<u8[16]>(op.rt);
// Checking for zero doeesn't care about the order of the bytes,
// so load the data before it's byteswapped
if (auto [ok, as] = match_expr(rt, byteswap(match<u8[16]>())); ok)
{
m_block->block_end = m_ir->GetInsertBlock();
const auto cond = eval(extract(bitcast<u32[4]>(as), 0) == 0);
const auto addr = eval(extract(get_vr(op.ra), 3) & 0x3fffc);
const auto target = add_block_indirect(op, addr);
m_ir->CreateCondBr(cond.value, target, add_block_next());
return;
}
// Check sign bit instead (optimization)
if (match_vr<s32[4], s64[2]>(op.rt, [&](auto c, auto MP)
{
@ -9263,6 +9277,21 @@ public:
{
if (m_block) m_block->block_end = m_ir->GetInsertBlock();
const auto rt = get_vr<u8[16]>(op.rt);
// Checking for zero doeesn't care about the order of the bytes,
// so load the data before it's byteswapped
if (auto [ok, as] = match_expr(rt, byteswap(match<u8[16]>())); ok)
{
m_block->block_end = m_ir->GetInsertBlock();
const auto cond = eval(extract(bitcast<u32[4]>(as), 0) != 0);
const auto addr = eval(extract(get_vr(op.ra), 3) & 0x3fffc);
const auto target = add_block_indirect(op, addr);
m_ir->CreateCondBr(cond.value, target, add_block_next());
return;
}
// Check sign bit instead (optimization)
if (match_vr<s32[4], s64[2]>(op.rt, [&](auto c, auto MP)
{
@ -9483,6 +9512,21 @@ public:
const u32 target = spu_branch_target(m_pos, op.i16);
const auto rt = get_vr<u8[16]>(op.rt);
// Checking for zero doeesn't care about the order of the bytes,
// so load the data before it's byteswapped
if (auto [ok, as] = match_expr(rt, byteswap(match<u8[16]>())); ok)
{
if (target != m_pos + 4)
{
m_block->block_end = m_ir->GetInsertBlock();
const auto cond = eval(extract(bitcast<u32[4]>(as), 0) == 0);
m_ir->CreateCondBr(cond.value, add_block(target), add_block(m_pos + 4));
return;
}
}
// Check sign bit instead (optimization)
if (match_vr<s32[4], s64[2]>(op.rt, [&](auto c, auto MP)
{
@ -9527,6 +9571,21 @@ public:
const u32 target = spu_branch_target(m_pos, op.i16);
const auto rt = get_vr<u8[16]>(op.rt);
// Checking for zero doeesn't care about the order of the bytes,
// so load the data before it's byteswapped
if (auto [ok, as] = match_expr(rt, byteswap(match<u8[16]>())); ok)
{
if (target != m_pos + 4)
{
m_block->block_end = m_ir->GetInsertBlock();
const auto cond = eval(extract(bitcast<u32[4]>(as), 0) != 0);
m_ir->CreateCondBr(cond.value, add_block(target), add_block(m_pos + 4));
return;
}
}
// Check sign bit instead (optimization)
if (match_vr<s32[4], s64[2]>(op.rt, [&](auto c, auto MP)
{
@ -9583,7 +9642,6 @@ public:
m_block->block_end = m_ir->GetInsertBlock();
const auto a = get_vr<s8[16]>(op.rt);
const auto cond = eval((bitcast<s16>(trunc<bool[16]>(a)) & 0x3000) == 0);
//const auto cond = eval((m & 0x3000) == 0);
m_ir->CreateCondBr(cond.value, add_block(target), add_block(m_pos + 4));
return true;
}