static intx allocate_prefetch_distance() {
// This method should be called before allocate_prefetch_style().
//
// Hardware prefetching (distance/size in bytes):
// Pentium 3 - 64 / 32
// Pentium 4 - 256 / 128
// Athlon - 64 / 32 ????
// Opteron - 128 / 64 only when 2 sequential cache lines accessed
// Core - 128 / 64
//
// Software prefetching (distance in bytes / instruction with best score):
// Pentium 3 - 128 / prefetchnta
// Pentium 4 - 512 / prefetchnta
// Athlon - 128 / prefetchnta
// Opteron - 256 / prefetchnta
// Core - 256 / prefetchnta
// It will be used only when AllocatePrefetchStyle > 0
intx count = AllocatePrefetchDistance;
if (count < 0) { // default ?
if (is_amd()) { // AMD
if (supports_sse2())
count = 256; // Opteron
else
count = 128; // Athlon
} else { // Intel
if (supports_sse2())
if (cpu_family() == 6) {
count = 256; // Pentium M, Core, Core2
} else {
count = 512; // Pentium 4
}
else
count = 128; // Pentium 3 (and all other old CPUs)
}
}
return count;
}
void
Downscaler::DownscaleInputLine()
{
typedef skia::ConvolutionFilter1D::Fixed FilterValue;
MOZ_ASSERT(mOutputBuffer);
MOZ_ASSERT(mCurrentOutLine < mTargetSize.height,
"Writing past end of output");
int32_t filterOffset = 0;
int32_t filterLength = 0;
MOZ_ASSERT(mCurrentOutLine < mYFilter->num_values());
auto filterValues =
mYFilter->FilterForValue(mCurrentOutLine, &filterOffset, &filterLength);
int32_t currentOutLine = mFlipVertically
? mTargetSize.height - (mCurrentOutLine + 1)
: mCurrentOutLine;
MOZ_ASSERT(currentOutLine >= 0);
uint8_t* outputLine =
&mOutputBuffer[currentOutLine * mTargetSize.width * sizeof(uint32_t)];
skia::ConvolveVertically(static_cast<const FilterValue*>(filterValues),
filterLength, mWindow.get(), mXFilter->num_values(),
outputLine, mHasAlpha, supports_sse2());
mCurrentOutLine += 1;
if (mCurrentOutLine == mTargetSize.height) {
// We're done.
return;
}
int32_t newFilterOffset = 0;
int32_t newFilterLength = 0;
GetFilterOffsetAndLength(mYFilter, mCurrentOutLine,
&newFilterOffset, &newFilterLength);
int diff = newFilterOffset - filterOffset;
MOZ_ASSERT(diff >= 0, "Moving backwards in the filter?");
// Shift the buffer. We're just moving pointers here, so this is cheap.
mLinesInBuffer -= diff;
mLinesInBuffer = max(mLinesInBuffer, 0);
for (int32_t i = 0; i < mLinesInBuffer; ++i) {
swap(mWindow[i], mWindow[filterLength - mLinesInBuffer + i]);
}
}
void
Downscaler::CommitRow()
{
MOZ_ASSERT(mOutputBuffer, "Should have a current frame");
MOZ_ASSERT(mCurrentInLine < mOriginalSize.height, "Past end of input");
MOZ_ASSERT(mCurrentOutLine < mTargetSize.height, "Past end of output");
int32_t filterOffset = 0;
int32_t filterLength = 0;
GetFilterOffsetAndLength(mYFilter, mCurrentOutLine,
&filterOffset, &filterLength);
int32_t inLineToRead = filterOffset + mLinesInBuffer;
MOZ_ASSERT(mCurrentInLine <= inLineToRead, "Reading past end of input");
if (mCurrentInLine == inLineToRead) {
skia::ConvolveHorizontally(mRowBuffer.get(), *mXFilter,
mWindow[mLinesInBuffer++], mHasAlpha,
supports_sse2());
}
MOZ_ASSERT(mCurrentOutLine < mTargetSize.height,
"Writing past end of output");
while (mLinesInBuffer == filterLength) {
DownscaleInputLine();
if (mCurrentOutLine == mTargetSize.height) {
break; // We're done.
}
GetFilterOffsetAndLength(mYFilter, mCurrentOutLine,
&filterOffset, &filterLength);
}
mCurrentInLine += 1;
// If we're at the end of the part of the original image that has data, commit
// rows to shift us to the end.
if (mCurrentInLine == (mFrameRect.y + mFrameRect.height)) {
SkipToRow(mOriginalSize.height - 1);
}
}
void VM_Version::get_processor_features() {
_cpu = 4; // 486 by default
_model = 0;
_stepping = 0;
_cpuFeatures = 0;
_logical_processors_per_package = 1;
if (!Use486InstrsOnly) {
// Get raw processor info
getPsrInfo_stub(&_cpuid_info);
assert_is_initialized();
_cpu = extended_cpu_family();
_model = extended_cpu_model();
_stepping = cpu_stepping();
if (cpu_family() > 4) { // it supports CPUID
_cpuFeatures = feature_flags();
// Logical processors are only available on P4s and above,
// and only if hyperthreading is available.
_logical_processors_per_package = logical_processor_count();
}
}
_supports_cx8 = supports_cmpxchg8();
#ifdef _LP64
// OS should support SSE for x64 and hardware should support at least SSE2.
if (!VM_Version::supports_sse2()) {
vm_exit_during_initialization("Unknown x64 processor: SSE2 not supported");
}
// in 64 bit the use of SSE2 is the minimum
if (UseSSE < 2) UseSSE = 2;
#endif
#ifdef AMD64
// flush_icache_stub have to be generated first.
// That is why Icache line size is hard coded in ICache class,
// see icache_x86.hpp. It is also the reason why we can't use
// clflush instruction in 32-bit VM since it could be running
// on CPU which does not support it.
//
// The only thing we can do is to verify that flushed
// ICache::line_size has correct value.
guarantee(_cpuid_info.std_cpuid1_edx.bits.clflush != 0, "clflush is not supported");
// clflush_size is size in quadwords (8 bytes).
guarantee(_cpuid_info.std_cpuid1_ebx.bits.clflush_size == 8, "such clflush size is not supported");
#endif
// If the OS doesn't support SSE, we can't use this feature even if the HW does
if (!os::supports_sse())
_cpuFeatures &= ~(CPU_SSE|CPU_SSE2|CPU_SSE3|CPU_SSSE3|CPU_SSE4A|CPU_SSE4_1|CPU_SSE4_2);
if (UseSSE < 4) {
_cpuFeatures &= ~CPU_SSE4_1;
_cpuFeatures &= ~CPU_SSE4_2;
}
if (UseSSE < 3) {
_cpuFeatures &= ~CPU_SSE3;
_cpuFeatures &= ~CPU_SSSE3;
_cpuFeatures &= ~CPU_SSE4A;
}
if (UseSSE < 2)
_cpuFeatures &= ~CPU_SSE2;
if (UseSSE < 1)
_cpuFeatures &= ~CPU_SSE;
if (UseAVX < 2)
_cpuFeatures &= ~CPU_AVX2;
if (UseAVX < 1)
_cpuFeatures &= ~CPU_AVX;
if (logical_processors_per_package() == 1) {
// HT processor could be installed on a system which doesn't support HT.
_cpuFeatures &= ~CPU_HT;
}
char buf[256];
jio_snprintf(buf, sizeof(buf), "(%u cores per cpu, %u threads per core) family %d model %d stepping %d%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s",
cores_per_cpu(), threads_per_core(),
cpu_family(), _model, _stepping,
(supports_cmov() ? ", cmov" : ""),
(supports_cmpxchg8() ? ", cx8" : ""),
(supports_fxsr() ? ", fxsr" : ""),
(supports_mmx() ? ", mmx" : ""),
(supports_sse() ? ", sse" : ""),
(supports_sse2() ? ", sse2" : ""),
(supports_sse3() ? ", sse3" : ""),
(supports_ssse3()? ", ssse3": ""),
(supports_sse4_1() ? ", sse4.1" : ""),
(supports_sse4_2() ? ", sse4.2" : ""),
(supports_popcnt() ? ", popcnt" : ""),
(supports_avx() ? ", avx" : ""),
(supports_avx2() ? ", avx2" : ""),
(supports_mmx_ext() ? ", mmxext" : ""),
(supports_3dnow_prefetch() ? ", 3dnowpref" : ""),
(supports_lzcnt() ? ", lzcnt": ""),
(supports_sse4a() ? ", sse4a": ""),
(supports_ht() ? ", ht": ""),
(supports_tsc() ? ", tsc": ""),
(supports_tscinv_bit() ? ", tscinvbit": ""),
(supports_tscinv() ? ", tscinv": ""));
_features_str = strdup(buf);
// UseSSE is set to the smaller of what hardware supports and what
// the command line requires. I.e., you cannot set UseSSE to 2 on
// older Pentiums which do not support it.
if (UseSSE > 4) UseSSE=4;
if (UseSSE < 0) UseSSE=0;
if (!supports_sse4_1()) // Drop to 3 if no SSE4 support
UseSSE = MIN2((intx)3,UseSSE);
if (!supports_sse3()) // Drop to 2 if no SSE3 support
UseSSE = MIN2((intx)2,UseSSE);
if (!supports_sse2()) // Drop to 1 if no SSE2 support
UseSSE = MIN2((intx)1,UseSSE);
if (!supports_sse ()) // Drop to 0 if no SSE support
UseSSE = 0;
if (UseAVX > 2) UseAVX=2;
if (UseAVX < 0) UseAVX=0;
if (!supports_avx2()) // Drop to 1 if no AVX2 support
UseAVX = MIN2((intx)1,UseAVX);
if (!supports_avx ()) // Drop to 0 if no AVX support
UseAVX = 0;
// On new cpus instructions which update whole XMM register should be used
// to prevent partial register stall due to dependencies on high half.
//
// UseXmmLoadAndClearUpper == true --> movsd(xmm, mem)
// UseXmmLoadAndClearUpper == false --> movlpd(xmm, mem)
// UseXmmRegToRegMoveAll == true --> movaps(xmm, xmm), movapd(xmm, xmm).
// UseXmmRegToRegMoveAll == false --> movss(xmm, xmm), movsd(xmm, xmm).
if( is_amd() ) { // AMD cpus specific settings
if( supports_sse2() && FLAG_IS_DEFAULT(UseAddressNop) ) {
// Use it on new AMD cpus starting from Opteron.
UseAddressNop = true;
}
if( supports_sse2() && FLAG_IS_DEFAULT(UseNewLongLShift) ) {
// Use it on new AMD cpus starting from Opteron.
UseNewLongLShift = true;
}
if( FLAG_IS_DEFAULT(UseXmmLoadAndClearUpper) ) {
if( supports_sse4a() ) {
UseXmmLoadAndClearUpper = true; // use movsd only on '10h' Opteron
} else {
UseXmmLoadAndClearUpper = false;
}
}
if( FLAG_IS_DEFAULT(UseXmmRegToRegMoveAll) ) {
if( supports_sse4a() ) {
UseXmmRegToRegMoveAll = true; // use movaps, movapd only on '10h'
} else {
UseXmmRegToRegMoveAll = false;
}
}
if( FLAG_IS_DEFAULT(UseXmmI2F) ) {
if( supports_sse4a() ) {
UseXmmI2F = true;
} else {
UseXmmI2F = false;
}
}
if( FLAG_IS_DEFAULT(UseXmmI2D) ) {
if( supports_sse4a() ) {
UseXmmI2D = true;
} else {
UseXmmI2D = false;
}
}
if( FLAG_IS_DEFAULT(UseSSE42Intrinsics) ) {
if( supports_sse4_2() && UseSSE >= 4 ) {
UseSSE42Intrinsics = true;
}
}
// Use count leading zeros count instruction if available.
if (supports_lzcnt()) {
if (FLAG_IS_DEFAULT(UseCountLeadingZerosInstruction)) {
UseCountLeadingZerosInstruction = true;
}
}
// some defaults for AMD family 15h
if ( cpu_family() == 0x15 ) {
// On family 15h processors default is no sw prefetch
if (FLAG_IS_DEFAULT(AllocatePrefetchStyle)) {
AllocatePrefetchStyle = 0;
}
// Also, if some other prefetch style is specified, default instruction type is PREFETCHW
if (FLAG_IS_DEFAULT(AllocatePrefetchInstr)) {
AllocatePrefetchInstr = 3;
}
// On family 15h processors use XMM and UnalignedLoadStores for Array Copy
if( FLAG_IS_DEFAULT(UseXMMForArrayCopy) ) {
UseXMMForArrayCopy = true;
}
if( FLAG_IS_DEFAULT(UseUnalignedLoadStores) && UseXMMForArrayCopy ) {
UseUnalignedLoadStores = true;
}
}
}
if( is_intel() ) { // Intel cpus specific settings
if( FLAG_IS_DEFAULT(UseStoreImmI16) ) {
UseStoreImmI16 = false; // don't use it on Intel cpus
}
if( cpu_family() == 6 || cpu_family() == 15 ) {
if( FLAG_IS_DEFAULT(UseAddressNop) ) {
// Use it on all Intel cpus starting from PentiumPro
UseAddressNop = true;
}
}
if( FLAG_IS_DEFAULT(UseXmmLoadAndClearUpper) ) {
UseXmmLoadAndClearUpper = true; // use movsd on all Intel cpus
}
if( FLAG_IS_DEFAULT(UseXmmRegToRegMoveAll) ) {
if( supports_sse3() ) {
UseXmmRegToRegMoveAll = true; // use movaps, movapd on new Intel cpus
} else {
UseXmmRegToRegMoveAll = false;
}
}
if( cpu_family() == 6 && supports_sse3() ) { // New Intel cpus
#ifdef COMPILER2
if( FLAG_IS_DEFAULT(MaxLoopPad) ) {
// For new Intel cpus do the next optimization:
// don't align the beginning of a loop if there are enough instructions
// left (NumberOfLoopInstrToAlign defined in c2_globals.hpp)
// in current fetch line (OptoLoopAlignment) or the padding
// is big (> MaxLoopPad).
// Set MaxLoopPad to 11 for new Intel cpus to reduce number of
// generated NOP instructions. 11 is the largest size of one
// address NOP instruction '0F 1F' (see Assembler::nop(i)).
MaxLoopPad = 11;
}
#endif // COMPILER2
if( FLAG_IS_DEFAULT(UseXMMForArrayCopy) ) {
UseXMMForArrayCopy = true; // use SSE2 movq on new Intel cpus
}
if( supports_sse4_2() && supports_ht() ) { // Newest Intel cpus
if( FLAG_IS_DEFAULT(UseUnalignedLoadStores) && UseXMMForArrayCopy ) {
UseUnalignedLoadStores = true; // use movdqu on newest Intel cpus
}
}
if( supports_sse4_2() && UseSSE >= 4 ) {
if( FLAG_IS_DEFAULT(UseSSE42Intrinsics)) {
UseSSE42Intrinsics = true;
}
}
}
}
// Use population count instruction if available.
if (supports_popcnt()) {
if (FLAG_IS_DEFAULT(UsePopCountInstruction)) {
UsePopCountInstruction = true;
}
} else if (UsePopCountInstruction) {
warning("POPCNT instruction is not available on this CPU");
FLAG_SET_DEFAULT(UsePopCountInstruction, false);
}
#ifdef COMPILER2
if (UseFPUForSpilling) {
if (UseSSE < 2) {
// Only supported with SSE2+
FLAG_SET_DEFAULT(UseFPUForSpilling, false);
}
}
#endif
assert(0 <= ReadPrefetchInstr && ReadPrefetchInstr <= 3, "invalid value");
assert(0 <= AllocatePrefetchInstr && AllocatePrefetchInstr <= 3, "invalid value");
// set valid Prefetch instruction
if( ReadPrefetchInstr < 0 ) ReadPrefetchInstr = 0;
if( ReadPrefetchInstr > 3 ) ReadPrefetchInstr = 3;
if( ReadPrefetchInstr == 3 && !supports_3dnow_prefetch() ) ReadPrefetchInstr = 0;
if( !supports_sse() && supports_3dnow_prefetch() ) ReadPrefetchInstr = 3;
if( AllocatePrefetchInstr < 0 ) AllocatePrefetchInstr = 0;
if( AllocatePrefetchInstr > 3 ) AllocatePrefetchInstr = 3;
if( AllocatePrefetchInstr == 3 && !supports_3dnow_prefetch() ) AllocatePrefetchInstr=0;
if( !supports_sse() && supports_3dnow_prefetch() ) AllocatePrefetchInstr = 3;
// Allocation prefetch settings
intx cache_line_size = prefetch_data_size();
if( cache_line_size > AllocatePrefetchStepSize )
AllocatePrefetchStepSize = cache_line_size;
assert(AllocatePrefetchLines > 0, "invalid value");
if( AllocatePrefetchLines < 1 ) // set valid value in product VM
AllocatePrefetchLines = 3;
assert(AllocateInstancePrefetchLines > 0, "invalid value");
if( AllocateInstancePrefetchLines < 1 ) // set valid value in product VM
AllocateInstancePrefetchLines = 1;
AllocatePrefetchDistance = allocate_prefetch_distance();
AllocatePrefetchStyle = allocate_prefetch_style();
if( is_intel() && cpu_family() == 6 && supports_sse3() ) {
if( AllocatePrefetchStyle == 2 ) { // watermark prefetching on Core
#ifdef _LP64
AllocatePrefetchDistance = 384;
#else
AllocatePrefetchDistance = 320;
#endif
}
if( supports_sse4_2() && supports_ht() ) { // Nehalem based cpus
AllocatePrefetchDistance = 192;
AllocatePrefetchLines = 4;
#ifdef COMPILER2
if (AggressiveOpts && FLAG_IS_DEFAULT(UseFPUForSpilling)) {
FLAG_SET_DEFAULT(UseFPUForSpilling, true);
}
#endif
}
}
assert(AllocatePrefetchDistance % AllocatePrefetchStepSize == 0, "invalid value");
#ifdef _LP64
// Prefetch settings
PrefetchCopyIntervalInBytes = prefetch_copy_interval_in_bytes();
PrefetchScanIntervalInBytes = prefetch_scan_interval_in_bytes();
PrefetchFieldsAhead = prefetch_fields_ahead();
#endif
#ifndef PRODUCT
if (PrintMiscellaneous && Verbose) {
tty->print_cr("Logical CPUs per core: %u",
logical_processors_per_package());
tty->print("UseSSE=%d",UseSSE);
if (UseAVX > 0) {
tty->print(" UseAVX=%d",UseAVX);
}
tty->cr();
tty->print("Allocation");
if (AllocatePrefetchStyle <= 0 || UseSSE == 0 && !supports_3dnow_prefetch()) {
tty->print_cr(": no prefetching");
} else {
tty->print(" prefetching: ");
if (UseSSE == 0 && supports_3dnow_prefetch()) {
tty->print("PREFETCHW");
} else if (UseSSE >= 1) {
if (AllocatePrefetchInstr == 0) {
tty->print("PREFETCHNTA");
} else if (AllocatePrefetchInstr == 1) {
tty->print("PREFETCHT0");
} else if (AllocatePrefetchInstr == 2) {
tty->print("PREFETCHT2");
} else if (AllocatePrefetchInstr == 3) {
tty->print("PREFETCHW");
}
}
if (AllocatePrefetchLines > 1) {
tty->print_cr(" at distance %d, %d lines of %d bytes", AllocatePrefetchDistance, AllocatePrefetchLines, AllocatePrefetchStepSize);
} else {
tty->print_cr(" at distance %d, one line of %d bytes", AllocatePrefetchDistance, AllocatePrefetchStepSize);
}
}
if (PrefetchCopyIntervalInBytes > 0) {
tty->print_cr("PrefetchCopyIntervalInBytes %d", PrefetchCopyIntervalInBytes);
}
if (PrefetchScanIntervalInBytes > 0) {
tty->print_cr("PrefetchScanIntervalInBytes %d", PrefetchScanIntervalInBytes);
}
if (PrefetchFieldsAhead > 0) {
tty->print_cr("PrefetchFieldsAhead %d", PrefetchFieldsAhead);
}
}
#endif // !PRODUCT
}