ENABLE_FAST_BLT is typically not assigned a value even when it is defined,so "#if" form is tecnically a syntax error.

This is not the case when being called from "fuzz" or "bench" testapplications. It may also not be accurate if a fast path has beensynthesised from a combination of copyBitsFallback and one or more otherfast paths.

In some places, sourceForm and destForm were being compared to determinewhich code path to follow. However, when being called from fuzz or othertest tools, these structures aren't used to pass parameters, so the pointershaven't been initialised and default to 0, so the wrong code path is followed.Detect such cases and initialise them from sourceBits and destBits instead,since these will perform the same under equality tests.

This shortcut is triggered more frequently than it used to be, due toimprovements in copyLoop() that avoid buffer overruns.

When classed as "wide" because each line is long enough to warrant pipelinedprefetching as we go along, the inner loop is unrolled enough that there isat least one prefetch instruction per iteration. Loading the source imagecan only be done in atoms of 32 bit words due to big-endian packing, sowhen destination pixels are 8 times wider (or more) than source pixels, theloads happen less frequently than the store atoms (quadwords) and aconditional branch per subblock is required to decide whether to do a loador not, depending on the skew and the number of pixels remaining to process.The 'x' register is only updated once per loop, so an assembly-time constantderived from the unrolling subblock number needs to be factored in, butsince the number of pixels remaining decreases as the subblock numberincreases, this should have been a subtraction.In practice, since only the least-significant bits of the result matter,addition and subtraction behave the same when the source:destination pixelratio is 8, so the only operations affected were 1->16bpp, 2->32bpp and1->32bpp. The exact threshold that counts as "wide" depends on the prefetchdistance that was selected empirically, but typically would require anoperation that is several hundreds of pixels wide.

In fastPathDepthConv (which combines sourceWord colour-depth conversion withanother fast path for another combinationRule at a constant colour depth)and fastPathRightToLeft, it could overflow the temporary buffer and therebycorrupt other local variables if the last chunk of a pixel row was 2048bytes (or just under). This was most likely to happen with 32bpp destinationimages and widths of about 512 pixels.

For images that were wide enough to invoke intra-line preloads, there was aregister clash between the preload address calculation and one of theregisters holding the deskewed source pixels (this only occurred once perdestination cacheline).

The halftone array is accessed using a hard-coded multiplier of 4 bytes,therefore the type of each element needs to be 32 bit on every platform.`sqInt` is not appropriate for this use, since it is a 64-bit type on64-bit platforms. Rather than unilaterally introduce C99 stdint types,use `unsigned int` since this wil be 32-bit on both current fast pathbinary targets.

Sometimes, colour maps are used such that all entries except the firstcontain the same value. Combined with the fact that only source colour 0uses colour map entry 0 (any other colours for which all non-0 bits wouldotherwise be discarded during index generation are forced to use entry 1instead), this effectively acts as a 2-entry (or 1bpp) map, depending onwhether the source colour is 0 or not. This is far more efficiently codedin any fast path by a test against zero, than by a table lookup - it freesup 2 KB, 16 KB or 128 KB of data cache space, depending on whether a 9-,12- or 15-bit colour map was used. There is an up-front cost to scanningthe colour map to see if its entries are of this nature, however in most"normal" colour maps, this scan will rapidly be aborted.

This runs approx 2.6x faster when benchmarked on Cortex-A72 in AArch64.

…nationThis makes better use of existing fast paths, and applies to all platforms.