Macintosh-SW/tenfourfox
1
0
Fork 0
mirror of https://github.com/classilla/tenfourfox.git synced 2024-05-31 09:41:37 +00:00
Code Issues Projects Releases Wiki Activity

#571: fix per M817058

Browse Source
This commit is contained in:
Cameron Kaiser 2019-08-29 15:05:56 -07:00
parent 78ce567c46
commit 8a9eccaae8
1 changed files with 20 additions and 19 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

39
dom/base/nsTextFragmentVMX.cpp
View File

@ -17,6 +17,9 @@ FirstNon8Bit(const char16_t *str, const char16_t *end)
{ {
const uint32_t numUnicharsPerVector = 8; const uint32_t numUnicharsPerVector = 8;
const uint32_t numCharsPerVector = 16; const uint32_t numCharsPerVector = 16;
// Paranoia. If this assertion is wrong, change the vector loop below.
MOZ_ASSERT((numCharsPerVector / numUnicharsPerVector) == sizeof(char16_t));
register vector unsigned short vect; register vector unsigned short vect;
typedef Non8BitParameters<sizeof(size_t)> p; typedef Non8BitParameters<sizeof(size_t)> p;
@ -39,27 +42,25 @@ if ((len - alignLen) > (numUnicharsPerVector - 1)) {
register const vector unsigned short gtcompare = register const vector unsigned short gtcompare =
vec_mergel( vec_splat_s8( 0 ), vec_splat_s8( -1 ) ); vec_mergel( vec_splat_s8( 0 ), vec_splat_s8( -1 ) );
// Check one VMX register (16 bytes) at a time.
// This is simpler on AltiVec and involves no mucking about with masks,
// since the vec_any_gt intrinsic does exactly what we want.
const int32_t vectWalkEnd = ((len - i) / numUnicharsPerVector) * numUnicharsPerVector; const int32_t vectWalkEnd = ((len - i) / numUnicharsPerVector) * numUnicharsPerVector;
// We use this a lot, so let's calculate it now. i2 = i * sizeof(char16_t);
const int32_t vectFactor = (numCharsPerVector/numUnicharsPerVector);
int32_t i2 = i * vectFactor;
while (1) { while (1) {
#define CheckForASCII \ // Check one VMX register (8 unichars) at a time. The vec_any_gt
vect = vec_ld(i2, (unsigned short *)str); \ // intrinsic does exactly what we want. This loop is manually unrolled;
if (vec_any_gt(vect, gtcompare)) \ // it yields notable performance improvements this way.
return (i2 / vectFactor); \ #define CheckForASCII \
i2 += numCharsPerVector; \ vect = vec_ld(i2, reinterpret_cast<const unsigned short*>(str)); \
if (!(i2 < vectWalkEnd)) \ if (vec_any_gt(vect, gtcompare)) return i; \
break; i += numUnicharsPerVector; \
CheckForASCII if (!(i < vectWalkEnd)) break; \
CheckForASCII i2 += numCharsPerVector;
}
i = i2 / vectFactor; CheckForASCII CheckForASCII
}
else { #undef CheckForASCII
}
} else {
// Align ourselves to a word boundary. // Align ourselves to a word boundary.
alignLen = alignLen =
std::min(len, int32_t(((-NS_PTR_TO_UINT32(str)) & alignMask) / sizeof(char16_t))); std::min(len, int32_t(((-NS_PTR_TO_UINT32(str)) & alignMask) / sizeof(char16_t)));