< prev index next >
src/share/vm/opto/superword.cpp
Print this page
@@ -233,10 +233,17 @@
if (vw > vw_best) {
// Do not vectorize a memory access with more elements per vector
// if unaligned memory access is not allowed because number of
// iterations in pre-loop will be not enough to align it.
create_pack = false;
+ } else {
+ SWPointer p2(best_align_to_mem_ref, this);
+ if (align_to_ref_p.invar() != p2.invar()) {
+ // Do not vectorize memory accesses with different invariants
+ // if unaligned memory accesses are not allowed.
+ create_pack = false;
+ }
}
}
} else {
if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
// Can't allow vectorization of unaligned memory accesses with the
@@ -454,12 +461,27 @@
// Stride one accesses are alignable if offset is aligned to memory operation size.
// Offset can be unaligned when UseUnalignedAccesses is used.
if (ABS(span) == mem_size && (ABS(offset) % mem_size) == 0) {
return true;
}
- // If initial offset from start of object is computable,
- // compute alignment within the vector.
+ // If the initial offset from start of the object is computable,
+ // check if the pre-loop can align the final offset accordingly.
+ //
+ // In other words: Can we find an i such that the offset
+ // after i pre-loop iterations is aligned to vw?
+ // (init_offset + pre_loop) % vw == 0 (1)
+ // where
+ // pre_loop = i * span
+ // is the number of bytes added to the offset by i pre-loop iterations.
+ //
+ // For this to hold we need pre_loop to increase init_offset by
+ // pre_loop = vw - (init_offset % vw)
+ //
+ // This is only possible if pre_loop is divisible by span because each
+ // pre-loop iteration increases the initial offset by 'span' bytes:
+ // (vw - (init_offset % vw)) % span == 0
+ //
int vw = vector_width_in_bytes(p.mem());
assert(vw > 1, "sanity");
if (vw % span == 0) {
Node* init_nd = pre_end->init_trip();
if (init_nd->is_Con() && p.invar() == NULL) {
@@ -473,31 +495,55 @@
} else {
assert(span < 0, "nonzero stride * scale");
return (init_offset % vw) % -span == 0;
}
}
+ } else if (span % vw == 0) {
+ // If span is a multiple of vw, we can simplify formula (1) to:
+ // (init_offset + i * span) % vw == 0
+ // =>
+ // (init_offset % vw) + ((i * span) % vw) == 0
+ // =>
+ // init_offset % vw == 0
+ //
+ // Because we add a multiple of vw to the initial offset, the
+ // final offset is a multiple of vw iff init_offset is a multiple.
+ //
+ Node* init_nd = pre_end->init_trip();
+ if (init_nd->is_Con() && p.invar() == NULL) {
+ int init = init_nd->bottom_type()->is_int()->get_con();
+ int init_offset = init * p.scale_in_bytes() + offset;
+ assert(init_offset >= 0, "positive offset from object start");
+ return (init_offset % vw) == 0;
+ }
}
return false;
}
//---------------------------get_iv_adjustment---------------------------
// Calculate loop's iv adjustment for this memory ops.
int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
SWPointer align_to_ref_p(mem_ref, this);
int offset = align_to_ref_p.offset_in_bytes();
int scale = align_to_ref_p.scale_in_bytes();
+ int elt_size = align_to_ref_p.memory_size();
int vw = vector_width_in_bytes(mem_ref);
assert(vw > 1, "sanity");
+ int iv_adjustment;
+ if (scale != 0) {
int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
// At least one iteration is executed in pre-loop by default. As result
// several iterations are needed to align memory operations in main-loop even
// if offset is 0.
int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw));
- int elt_size = align_to_ref_p.memory_size();
assert(((ABS(iv_adjustment_in_bytes) % elt_size) == 0),
err_msg_res("(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size));
- int iv_adjustment = iv_adjustment_in_bytes/elt_size;
+ iv_adjustment = iv_adjustment_in_bytes/elt_size;
+ } else {
+ // This memory op is not dependent on iv (scale == 0)
+ iv_adjustment = 0;
+ }
#ifndef PRODUCT
if (TraceSuperWord)
tty->print_cr("\noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d",
offset, iv_adjustment, elt_size, scale, iv_stride(), vw);
< prev index next >