Impaczus/rpcs3
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

SPU LLVM: Use TBL for AArch64 ROTQBY shuffles

Browse Source
This commit is contained in:
Whatcookie
2026-05-28 08:30:08 -04:00
committed by Ani
parent d25972e196
commit 2d1be09180
1 changed files with 69 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files
rpcs3/Emu/Cell/SPULLVMRecompiler.cpp
+69 -30
View File
@@ -5783,6 +5783,45 @@ public:
return zshuffle(std::forward<TA>(a), 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0); return zshuffle(std::forward<TA>(a), 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0);
} }
static auto rotqby_reverse_base()
{
return build<u8[16]>(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0);
}
static auto rotqby_forward_base()
{
return build<u8[16]>(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
}
static auto rotqby_zero_base()
{
#ifdef ARCH_ARM64
return rotqby_forward_base();
#else
return build<u8[16]>(112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127);
#endif
}
static auto rotqby_reverse_zero_base()
{
#ifdef ARCH_ARM64
return rotqby_reverse_base();
#else
return build<u8[16]>(127, 126, 125, 124, 123, 122, 121, 120, 119, 118, 117, 116, 115, 114, 113, 112);
#endif
}
// For use in rotqby family of instructions only
template <typename T, typename U>
auto pshufb_for_x86_and_tbl_for_aarch64(T&& a, U&& b)
{
#ifdef ARCH_ARM64
return tbl(std::forward<T>(a), std::forward<U>(b));
#else
return pshufb(std::forward<T>(a), std::forward<U>(b));
#endif
}
template <typename T, typename U> template <typename T, typename U>
static llvm_calli<u8[16], T, U> rotqbybi(T&& a, U&& b) static llvm_calli<u8[16], T, U> rotqbybi(T&& a, U&& b)
{ {
@@ -5799,7 +5838,7 @@ public:
// Data with swapped endian from a load instruction // Data with swapped endian from a load instruction
if (auto [ok, as] = match_expr(a, byteswap(match<u8[16]>())); ok) if (auto [ok, as] = match_expr(a, byteswap(match<u8[16]>())); ok)
{ {
const auto sc = build<u8[16]>(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0); const auto sc = rotqby_reverse_base();
const auto sh = sc + (splat_scalar(b) >> 3); const auto sh = sc + (splat_scalar(b) >> 3);
if (m_use_avx512_icl) if (m_use_avx512_icl)
@@ -5807,9 +5846,9 @@ public:
return eval(vpermb(as, sh)); return eval(vpermb(as, sh));
} }
return eval(pshufb(as, (sh & 0xf))); return eval(pshufb_for_x86_and_tbl_for_aarch64(as, (sh & 0xf)));
} }
const auto sc = build<u8[16]>(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15); const auto sc = rotqby_forward_base();
const auto sh = sc - (splat_scalar(b) >> 3); const auto sh = sc - (splat_scalar(b) >> 3);
if (m_use_avx512_icl) if (m_use_avx512_icl)
@@ -5817,7 +5856,7 @@ public:
return eval(vpermb(a, sh)); return eval(vpermb(a, sh));
} }
return eval(pshufb(a, (sh & 0xf))); return eval(pshufb_for_x86_and_tbl_for_aarch64(a, (sh & 0xf)));
}); });
set_vr(op.rt, rotqbybi(get_vr<u8[16]>(op.ra), get_vr<u8[16]>(op.rb))); set_vr(op.rt, rotqbybi(get_vr<u8[16]>(op.ra), get_vr<u8[16]>(op.rb)));
@@ -5845,15 +5884,15 @@ public:
// Data with swapped endian from a load instruction // Data with swapped endian from a load instruction
if (auto [ok, as] = match_expr(a, byteswap(match<u8[16]>())); ok) if (auto [ok, as] = match_expr(a, byteswap(match<u8[16]>())); ok)
{ {
const auto sc = build<u8[16]>(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0); const auto sc = rotqby_reverse_base();
const auto sh = sc - splat_scalar(minusbx); const auto sh = sc - splat_scalar(minusbx);
set_vr(op.rt, pshufb(as, sh)); set_vr(op.rt, pshufb_for_x86_and_tbl_for_aarch64(as, sh));
return; return;
} }
const auto sc = build<u8[16]>(112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127); const auto sc = rotqby_zero_base();
const auto sh = sc + splat_scalar(minusbx); const auto sh = sc + splat_scalar(minusbx);
set_vr(op.rt, pshufb(a, sh)); set_vr(op.rt, pshufb_for_x86_and_tbl_for_aarch64(a, sh));
} }
void SHLQBYBI(spu_opcode_t op) void SHLQBYBI(spu_opcode_t op)
@@ -5864,15 +5903,15 @@ public:
// Data with swapped endian from a load instruction // Data with swapped endian from a load instruction
if (auto [ok, as] = match_expr(a, byteswap(match<u8[16]>())); ok) if (auto [ok, as] = match_expr(a, byteswap(match<u8[16]>())); ok)
{ {
const auto sc = build<u8[16]>(127, 126, 125, 124, 123, 122, 121, 120, 119, 118, 117, 116, 115, 114, 113, 112); const auto sc = rotqby_reverse_zero_base();
const auto sh = sc + (splat_scalar(b) >> 3); const auto sh = sc + (splat_scalar(b) >> 3);
set_vr(op.rt, pshufb(as, sh)); set_vr(op.rt, pshufb_for_x86_and_tbl_for_aarch64(as, sh));
return; return;
} }
const auto sc = build<u8[16]>(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15); const auto sc = rotqby_forward_base();
const auto sh = sc - (splat_scalar(b) >> 3); const auto sh = sc - (splat_scalar(b) >> 3);
set_vr(op.rt, pshufb(a, sh)); set_vr(op.rt, pshufb_for_x86_and_tbl_for_aarch64(a, sh));
} }
template <typename RT, typename T> template <typename RT, typename T>
@@ -6026,7 +6065,7 @@ public:
// Data with swapped endian from a load instruction // Data with swapped endian from a load instruction
if (auto [ok, as] = match_expr(a, byteswap(match<u8[16]>())); ok) if (auto [ok, as] = match_expr(a, byteswap(match<u8[16]>())); ok)
{ {
const auto sc = build<u8[16]>(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0); const auto sc = rotqby_reverse_base();
const auto sh = eval(sc + splat_scalar(b)); const auto sh = eval(sc + splat_scalar(b));
if (m_use_avx512_icl) if (m_use_avx512_icl)
@@ -6035,11 +6074,11 @@ public:
return; return;
} }
set_vr(op.rt, pshufb(as, (sh & 0xf))); set_vr(op.rt, pshufb_for_x86_and_tbl_for_aarch64(as, (sh & 0xf)));
return; return;
} }
const auto sc = build<u8[16]>(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15); const auto sc = rotqby_forward_base();
const auto sh = eval(sc - splat_scalar(b)); const auto sh = eval(sc - splat_scalar(b));
if (m_use_avx512_icl) if (m_use_avx512_icl)
@@ -6048,7 +6087,7 @@ public:
return; return;
} }
set_vr(op.rt, pshufb(a, (sh & 0xf))); set_vr(op.rt, pshufb_for_x86_and_tbl_for_aarch64(a, (sh & 0xf)));
} }
void ROTQMBY(spu_opcode_t op) void ROTQMBY(spu_opcode_t op)
@@ -6067,15 +6106,15 @@ public:
// Data with swapped endian from a load instruction // Data with swapped endian from a load instruction
if (auto [ok, as] = match_expr(a, byteswap(match<u8[16]>())); ok) if (auto [ok, as] = match_expr(a, byteswap(match<u8[16]>())); ok)
{ {
const auto sc = build<u8[16]>(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0); const auto sc = rotqby_reverse_base();
const auto sh = sc - (splat_scalar(minusbx) & 0x1f); const auto sh = sc - (splat_scalar(minusbx) & 0x1f);
set_vr(op.rt, pshufb(as, sh)); set_vr(op.rt, pshufb_for_x86_and_tbl_for_aarch64(as, sh));
return; return;
} }
const auto sc = build<u8[16]>(112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127); const auto sc = rotqby_zero_base();
const auto sh = sc + (splat_scalar(minusbx) & 0x1f); const auto sh = sc + (splat_scalar(minusbx) & 0x1f);
set_vr(op.rt, pshufb(a, sh)); set_vr(op.rt, pshufb_for_x86_and_tbl_for_aarch64(a, sh));
} }
void SHLQBY(spu_opcode_t op) void SHLQBY(spu_opcode_t op)
@@ -6086,15 +6125,15 @@ public:
// Data with swapped endian from a load instruction // Data with swapped endian from a load instruction
if (auto [ok, as] = match_expr(a, byteswap(match<u8[16]>())); ok) if (auto [ok, as] = match_expr(a, byteswap(match<u8[16]>())); ok)
{ {
const auto sc = build<u8[16]>(127, 126, 125, 124, 123, 122, 121, 120, 119, 118, 117, 116, 115, 114, 113, 112); const auto sc = rotqby_reverse_zero_base();
const auto sh = sc + (splat_scalar(b) & 0x1f); const auto sh = sc + (splat_scalar(b) & 0x1f);
set_vr(op.rt, pshufb(as, sh)); set_vr(op.rt, pshufb_for_x86_and_tbl_for_aarch64(as, sh));
return; return;
} }
const auto sc = build<u8[16]>(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15); const auto sc = rotqby_forward_base();
const auto sh = sc - (splat_scalar(b) & 0x1f); const auto sh = sc - (splat_scalar(b) & 0x1f);
set_vr(op.rt, pshufb(a, sh)); set_vr(op.rt, pshufb_for_x86_and_tbl_for_aarch64(a, sh));
} }
template <typename T> template <typename T>
@@ -6192,26 +6231,26 @@ public:
void ROTQBYI(spu_opcode_t op) void ROTQBYI(spu_opcode_t op)
{ {
const auto a = get_vr<u8[16]>(op.ra); const auto a = get_vr<u8[16]>(op.ra);
const auto sc = build<u8[16]>(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15); const auto sc = rotqby_forward_base();
const auto sh = (sc - get_imm<u8[16]>(op.i7, false)) & 0xf; const auto sh = (sc - get_imm<u8[16]>(op.i7, false)) & 0xf;
set_vr(op.rt, pshufb(a, sh)); set_vr(op.rt, pshufb_for_x86_and_tbl_for_aarch64(a, sh));
} }
void ROTQMBYI(spu_opcode_t op) void ROTQMBYI(spu_opcode_t op)
{ {
const auto a = get_vr<u8[16]>(op.ra); const auto a = get_vr<u8[16]>(op.ra);
const auto sc = build<u8[16]>(112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127); const auto sc = rotqby_zero_base();
const auto sh = sc + (-get_imm<u8[16]>(op.i7, false) & 0x1f); const auto sh = sc + (-get_imm<u8[16]>(op.i7, false) & 0x1f);
set_vr(op.rt, pshufb(a, sh)); set_vr(op.rt, pshufb_for_x86_and_tbl_for_aarch64(a, sh));
} }
void SHLQBYI(spu_opcode_t op) void SHLQBYI(spu_opcode_t op)
{ {
if (get_reg_raw(op.ra) && !op.i7) return set_reg_fixed(op.rt, get_reg_raw(op.ra), false); // For expressions matching if (get_reg_raw(op.ra) && !op.i7) return set_reg_fixed(op.rt, get_reg_raw(op.ra), false); // For expressions matching
const auto a = get_vr<u8[16]>(op.ra); const auto a = get_vr<u8[16]>(op.ra);
const auto sc = build<u8[16]>(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15); const auto sc = rotqby_forward_base();
const auto sh = sc - (get_imm<u8[16]>(op.i7, false) & 0x1f); const auto sh = sc - (get_imm<u8[16]>(op.i7, false) & 0x1f);
set_vr(op.rt, pshufb(a, sh)); set_vr(op.rt, pshufb_for_x86_and_tbl_for_aarch64(a, sh));
} }
void CGT(spu_opcode_t op) void CGT(spu_opcode_t op)