summaryrefslogtreecommitdiff
path: root/src/gallium/drivers/cell/spu
diff options
context:
space:
mode:
authorIan Romanick <idr@us.ibm.com>2008-02-22 17:51:55 -0800
committerIan Romanick <idr@us.ibm.com>2008-02-25 16:18:08 -0800
commit2efa7e9489541b6a86c3d46e3d58cbf5bf399189 (patch)
treeee99e351f474506a9b3fee96737a7e94b5325aee /src/gallium/drivers/cell/spu
parenta63fd641a01a50e1be51664bf863e01ddaf61d3e (diff)
cell: Fix off-by-one error in spu_dcache_fetch_unaligned
This time the off-by-one error caused an extra qword to be fetched under certain circumstances when the source ea was not qword aligned.
Diffstat (limited to 'src/gallium/drivers/cell/spu')
-rw-r--r--src/gallium/drivers/cell/spu/spu_dcache.c50
1 files changed, 32 insertions, 18 deletions
diff --git a/src/gallium/drivers/cell/spu/spu_dcache.c b/src/gallium/drivers/cell/spu/spu_dcache.c
index 3baeaea998..a1701d80d1 100644
--- a/src/gallium/drivers/cell/spu/spu_dcache.c
+++ b/src/gallium/drivers/cell/spu/spu_dcache.c
@@ -49,43 +49,57 @@
/**
* Fetch between arbitrary number of bytes from an unaligned address
+ *
+ * \param dst Destination data buffer
+ * \param ea Main memory effective address of source data
+ * \param size Number of bytes to read
+ *
+ * \warning
+ * As is hinted by the type of the \c dst pointer, this function writes
+ * multiples of 16-bytes.
*/
void
spu_dcache_fetch_unaligned(qword *dst, unsigned ea, unsigned size)
{
const int shift = ea & 0x0f;
- const unsigned aligned_start_ea = ea & ~0x0f;
- const unsigned aligned_end_ea = ROUNDUP16(ea + size);
- const unsigned num_entries = (aligned_end_ea - aligned_start_ea) / 16;
+ const unsigned read_size = ROUNDUP16(size + shift);
+ const unsigned last_read = ROUNDUP16(ea + size);
+ const qword *const last_write = dst + (ROUNDUP16(size) / 16);
unsigned i;
if (shift == 0) {
/* Data is already aligned. Fetch directly into the destination buffer.
*/
- for (i = 0; i < num_entries; i++) {
- dst[i] = cache_rd(data, ea + (i * 16));
+ for (i = 0; i < size; i += 16) {
+ *(dst++) = cache_rd(data, ea + i);
}
} else {
- qword tmp[2] ALIGN16_ATTRIB;
-
+ qword hi;
- tmp[0] = cache_rd(data, (ea & ~0x0f));
- for (i = 0; i < (num_entries & ~1); i++) {
- const unsigned curr = i & 1;
- const unsigned next = curr ^ 1;
- tmp[next] = cache_rd(data, (ea & ~0x0f) + (next * 16));
-
- dst[i] = si_or((qword) spu_slqwbyte(tmp[curr], shift),
- (qword) spu_rlmaskqwbyte(tmp[next], shift - 16));
+ /* Please exercise extreme caution when modifying this code. This code
+ * must not read past the end of the page containing the source data,
+ * and it must not write more than ((size + 15) / 16) qwords to the
+ * destination buffer.
+ */
+ ea &= ~0x0f;
+ hi = cache_rd(data, ea);
+ for (i = 16; i < read_size; i += 16) {
+ qword lo = cache_rd(data, ea + i);
+
+ *(dst++) = si_or((qword) spu_slqwbyte(hi, shift),
+ (qword) spu_rlmaskqwbyte(lo, shift - 16));
+ hi = lo;
}
- if (i < num_entries) {
- dst[i] = si_or((qword) spu_slqwbyte(tmp[(i & 1)], shift),
- si_il(0));
+ if (dst != last_write) {
+ *(dst++) = si_or((qword) spu_slqwbyte(hi, shift), si_il(0));
}
}
+
+ ASSERT((ea + i) == last_read);
+ ASSERT(dst == last_write);
}