emit_x64: Remove SSSE3 implementation of PackedHalvingAddU8

It is much slower than the SSE2 implementation, so there's no point keeping it around.
This commit is contained in:
MerryMage 2017-11-25 15:38:36 +00:00
parent c4b40909f7
commit 18f11972c6

View file

@ -1742,56 +1742,25 @@ void EmitX64::EmitPackedSubS16(RegAlloc& reg_alloc, IR::Block& block, IR::Inst*
void EmitX64::EmitPackedHalvingAddU8(RegAlloc& reg_alloc, IR::Block&, IR::Inst* inst) {
auto args = reg_alloc.GetArgumentInfo(inst);
// This code path requires SSSE3 because of the PSHUFB instruction.
// A fallback implementation is provided below.
if (code->DoesCpuSupport(Xbyak::util::Cpu::tSSSE3)) {
Xbyak::Xmm xmm_a = reg_alloc.UseScratchXmm(args[0]);
Xbyak::Xmm xmm_b = reg_alloc.UseScratchXmm(args[1]);
Xbyak::Reg32 reg_a = reg_alloc.UseScratchGpr(args[0]).cvt32();
Xbyak::Reg32 reg_b = reg_alloc.UseGpr(args[1]).cvt32();
Xbyak::Reg32 xor_a_b = reg_alloc.ScratchGpr().cvt32();
Xbyak::Reg32 and_a_b = reg_a;
Xbyak::Reg32 result = reg_a;
Xbyak::Xmm xmm_mask = reg_alloc.ScratchXmm();
Xbyak::Reg64 mask = reg_alloc.ScratchGpr();
// This relies on the equality x+y == ((x&y) << 1) + (x^y).
// Note that x^y always contains the LSB of the result.
// Since we want to calculate (x+y)/2, we can instead calculate (x&y) + ((x^y)>>1).
// We mask by 0x7F to remove the LSB so that it doesn't leak into the field below.
// Set the mask to expand the values
// 0xAABBCCDD becomes 0x00AA00BB00CC00DD
code->mov(mask, 0x8003800280018000);
code->movq(xmm_mask, mask);
code->mov(xor_a_b, reg_a);
code->and_(and_a_b, reg_b);
code->xor_(xor_a_b, reg_b);
code->shr(xor_a_b, 1);
code->and_(xor_a_b, 0x7F7F7F7F);
code->add(result, xor_a_b);
// Expand each 8-bit value to 16-bit
code->pshufb(xmm_a, xmm_mask);
code->pshufb(xmm_b, xmm_mask);
// Add the individual 16-bit values
code->paddw(xmm_a, xmm_b);
// Shift the 16-bit values to the right to halve them
code->psrlw(xmm_a, 1);
// Set the mask to pack the values again
// 0x00AA00BB00CC00DD becomes 0xAABBCCDD
code->mov(mask, 0x06040200);
code->movq(xmm_mask, mask);
// Shuffle them back to 8-bit values
code->pshufb(xmm_a, xmm_mask);
reg_alloc.DefineValue(inst, xmm_a);
} else {
// Fallback implementation in case the CPU doesn't support SSSE3
Xbyak::Reg32 reg_a = reg_alloc.UseScratchGpr(args[0]).cvt32();
Xbyak::Reg32 reg_b = reg_alloc.UseGpr(args[1]).cvt32();
Xbyak::Reg32 xor_a_b = reg_alloc.ScratchGpr().cvt32();
Xbyak::Reg32 and_a_b = reg_a;
Xbyak::Reg32 result = reg_a;
code->mov(xor_a_b, reg_a);
code->and_(and_a_b, reg_b);
code->xor_(xor_a_b, reg_b);
code->shr(xor_a_b, 1);
code->and_(xor_a_b, 0x7F7F7F7F);
code->add(result, xor_a_b);
reg_alloc.DefineValue(inst, result);
}
reg_alloc.DefineValue(inst, result);
}
void EmitX64::EmitPackedHalvingAddU16(RegAlloc& reg_alloc, IR::Block&, IR::Inst* inst) {