thirdparty/scummvm
2
0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

PSP: optimized memcpy some more and fixed memcpy testing

Browse source 
Found that the particular implementation was producing messy assembly for misaligned copies. Improved it and also fixed up wrapping the memcpy, which would cause endless prints in case memcpy testing is asked for.

svn-id: r51503
This commit is contained in:
Yotam Barnoy 2010-07-30 09:32:54 +00:00
parent 2fc1461f2b
commit 1532685d82
2 changed files with 38 additions and 36 deletions
Download patch file Download diff file Expand all files Collapse all files

73
backends/platform/psp/memory.cpp
View file

@ -39,9 +39,17 @@
extern "C" {
#ifdef TEST_MEMORY_COPY /* we won't be able to run in this case b/c of printouts */
extern void *__real_memcpy(void *dst, void *src, size_t bytes);
#endif
void *__wrap_memcpy(void *dst, void *src, size_t bytes) {
#ifdef TEST_MEMORY_COPY /* we won't be able to run in this case */
return __real_memcpy(dst, src, bytes);
#else
PspMemory::fastCopy((byte *)dst, (byte *)src, bytes);
return dst;
#endif
}
}
@ -291,43 +299,31 @@ void PspMemory::copy32Misaligned(uint32 *dst32, const byte *src, uint32 bytes, u
PSP_DEBUG_PRINT("copy32Misaligned: dst32[%p], src[%p], bytes[%d], alignSrc[%d]\n", dst32, src, bytes, alignSrc);
uint32 *src32 = (uint32 *)(((uint32)src) & 0xFFFFFFFC); // remove misalignment
uint32 offset;
uint32 shiftValue, lastShiftValue;
switch (alignSrc) {
case 1:
offset = misaligned32Detail(dst32, src32, bytes, alignSrc, 8, 24);
shiftValue = 8;
lastShiftValue = 24;
break;
case 2:
offset = misaligned32Detail(dst32, src32, bytes, alignSrc, 16, 16);
shiftValue = 16;
lastShiftValue = 16;
break;
default: /* 3 */
offset = misaligned32Detail(dst32, src32, bytes, alignSrc, 24, 8);
shiftValue = 24;
lastShiftValue = 8;
break;
}
uint32 remainingBytes = bytes & 3;
if (remainingBytes) {
byte *dst = (byte *)dst32;
src += offset;
dst += offset;
copy8(dst, src, remainingBytes);
}
}
// returns offset in dst
uint32 PspMemory::misaligned32Detail(uint32 *dst32, uint32 *src32, uint32 bytes, uint32 alignSrc, const uint32 shiftValue, const uint32 lastShiftValue) {
uint32 *origDst32 = dst32;
register uint32 dstWord, srcWord;
PSP_DEBUG_PRINT("misaligned32Detail(): alignSrc[%d], dst32[%p], src32[%p], bytes[%d]\n", alignSrc, dst32, src32, bytes);
uint32 dstWord, srcWord;
// Try to do groups of 4 words
uint32 words4 = bytes >> 4;
srcWord = src32[0];
while (words4--) {
srcWord = *src32; // preload 1st word so we read ahead
for (; words4; words4--) {
dstWord = srcWord >> shiftValue;
srcWord = src32[1];
dstWord |= srcWord << lastShiftValue;
@ -348,22 +344,29 @@ uint32 PspMemory::misaligned32Detail(uint32 *dst32, uint32 *src32, uint32 bytes,
dst32 += 4;
}
uint32 words = (bytes & 0xF) >> 2;
uint32 words = (bytes & 0xF) >> 2; // now get remaining words
// we read one word ahead of what we write
// setup the first read
if (words) {
src32++; // we already loaded the value, so just increment
while (words--) {
dstWord = srcWord >> shiftValue;
srcWord = *src32++;
dstWord |= srcWord << lastShiftValue;
*dst32++ = dstWord;
}
for (; words ;words--) {
dstWord = srcWord >> shiftValue;
srcWord = src32[1]; // we still go one ahead
src32++;
dstWord |= srcWord << lastShiftValue;
*dst32++ = dstWord;
}
return (byte *)dst32 - (byte *)origDst32;
uint32 bytesLeft = bytes & 3; // and remaining bytes
if (bytesLeft) {
byte *dst8 = (byte *)dst32;
byte *src8 = ((byte *)src32) + ((uint32)src & 0x3); // get exact location we should be at
for(; bytesLeft; bytesLeft--) {
*dst8++ = *src8++;
}
}
}
// More challenging -- need to shift