diff --git a/Common/ArmEmitter.cpp b/Common/ArmEmitter.cpp index b81a0049aa..f9d8a2a2de 100644 --- a/Common/ArmEmitter.cpp +++ b/Common/ArmEmitter.cpp @@ -801,6 +801,17 @@ void ARMXEmitter::CLZ(ARMReg rd, ARMReg rm) Write32(condition | (0x16F << 16) | (rd << 12) | (0xF1 << 4) | rm); } +void ARMXEmitter::PLD(ARMReg rn, int offset, bool forWrite) { + _dbg_assert_msg_(JIT, offset < 0x3ff && offset > -0x3ff, "PLD: Max 12 bits of offset allowed"); + + bool U = offset >= 0; + if (offset < 0) offset = -offset; + bool R = !forWrite; + // Conditions not allowed + Write32((0xF5 << 24) | (U << 23) | (R << 22) | (1 << 20) | ((int)rn << 16) | (0xF << 12) | offset); +} + + void ARMXEmitter::BFI(ARMReg rd, ARMReg rn, u8 lsb, u8 width) { u32 msb = (lsb + width - 1); diff --git a/Common/ArmEmitter.h b/Common/ArmEmitter.h index 6157127f25..9de2e83e37 100644 --- a/Common/ArmEmitter.h +++ b/Common/ArmEmitter.h @@ -541,6 +541,7 @@ public: void BFI(ARMReg rd, ARMReg rn, u8 lsb, u8 width); void UBFX(ARMReg dest, ARMReg op2, u8 lsb, u8 width); void CLZ(ARMReg rd, ARMReg rm); + void PLD(ARMReg rd, int offset, bool forWrite = false); // Using just MSR here messes with our defines on the PPC side of stuff (when this code was in dolphin...) // Just need to put an underscore here, bit annoying. diff --git a/GPU/GLES/VertexDecoder.cpp b/GPU/GLES/VertexDecoder.cpp index 405a899616..d4aa9eb0b8 100644 --- a/GPU/GLES/VertexDecoder.cpp +++ b/GPU/GLES/VertexDecoder.cpp @@ -1058,6 +1058,8 @@ JittedVertexDecoder VertexDecoderJitCache::Compile(const VertexDecoder &dec) { } JumpTarget loopStart = GetCodePtr(); + // Preload data cache ahead of reading. TODO: Experiment with the offset. + PLD(srcReg, 64); for (int i = 0; i < dec.numSteps_; i++) { if (!CompileStep(dec, i)) { // Reset the code ptr and return zero to indicate that we failed. @@ -1265,13 +1267,14 @@ void VertexDecoderJitCache::Jit_TcU16ThroughDouble() { } void VertexDecoderJitCache::Jit_TcU8Prescale() { - if (false && cpu_info.bNEON) { + if (cpu_info.bNEON) { // TODO: Needs testing ADD(scratchReg, srcReg, dec_->tcoff); VLD1_lane(I_16, neonScratchReg, scratchReg, 0, false); - VMOVL(I_8 | I_UNSIGNED, neonScratchRegQ, neonScratchReg); // Widen to 32-bit + VMOVL(I_8 | I_UNSIGNED, neonScratchRegQ, neonScratchReg); // Widen to 16-bit VMOVL(I_16 | I_UNSIGNED, neonScratchRegQ, neonScratchReg); // Widen to 32-bit VCVT(F_32 | I_UNSIGNED, neonScratchRegQ, neonScratchRegQ); + ADD(scratchReg2, dstReg, dec_->decFmt.uvoff); VMUL(F_32, neonScratchReg, neonScratchReg, neonUVScaleReg); VADD(F_32, neonScratchReg, neonScratchReg, neonUVOffsetReg); VST1(F_32, neonScratchReg, scratchReg2, 1, ALIGN_NONE); @@ -1294,12 +1297,13 @@ void VertexDecoderJitCache::Jit_TcU8Prescale() { } void VertexDecoderJitCache::Jit_TcU16Prescale() { - if (false && cpu_info.bNEON) { + if (cpu_info.bNEON) { // TODO: Needs testing ADD(scratchReg, srcReg, dec_->tcoff); VLD1_lane(I_32, neonScratchReg, scratchReg, 0, false); VMOVL(I_16 | I_UNSIGNED, neonScratchRegQ, neonScratchReg); // Widen to 32-bit VCVT(F_32 | I_UNSIGNED, neonScratchRegQ, neonScratchRegQ); + ADD(scratchReg2, dstReg, dec_->decFmt.uvoff); VMUL(F_32, neonScratchReg, neonScratchReg, neonUVScaleReg); VADD(F_32, neonScratchReg, neonScratchReg, neonUVOffsetReg); VST1(F_32, neonScratchReg, scratchReg2, 1, ALIGN_NONE); diff --git a/android/jni/ArmEmitterTest.cpp b/android/jni/ArmEmitterTest.cpp index d36b8491fc..3cf6e4ced9 100644 --- a/android/jni/ArmEmitterTest.cpp +++ b/android/jni/ArmEmitterTest.cpp @@ -52,8 +52,6 @@ void TestCode::Generate() VLD1_all_lanes(F_32, Q2, R1, true); ADD(R0, R0, 12); VLD1_lane(F_32, D4, R0, 1, true); - u32 word = *(u32 *)(GetCodePtr() - 4); - ILOG("Instruction Word: %08x", word); // VMUL(F_32, Q2, Q0, Q1); VST1(F_32, D4, R2, 2); */ @@ -70,6 +68,10 @@ void TestCode::Generate() VST1(I_32, D2, R1, 2); VST1(I_32, D4, R2, 2); VST1(I_32, D6, R3, 2); + PLD(R1, 32); + u32 word = *(u32 *)(GetCodePtr() - 4); + ILOG("Instruction Word: %08x", word); + // This works!