| Message ID | 20140920040451.GC21760@kam.mff.cuni.cz |
|---|---|
| State | New |
| Headers | show |
On 09/19/14 22:04, Jan Hubicka wrote: > Hi, > int_bit_position is used by ipa-devirt's type walking code. It is currently a bottleneck > since I introduced speculation into contextes (I plan to solve this by changing the > way i cache results). But this patch seems to make sense anyway: we do not need to go > through folding: > tree > bit_from_pos (tree offset, tree bitpos) > { > if (TREE_CODE (offset) == PLUS_EXPR) > offset = size_binop (PLUS_EXPR, > fold_convert (bitsizetype, TREE_OPERAND (offset, 0)), > fold_convert (bitsizetype, TREE_OPERAND (offset, 1))); > else > offset = fold_convert (bitsizetype, offset); > return size_binop (PLUS_EXPR, bitpos, > size_binop (MULT_EXPR, offset, bitsize_unit_node)); > } > > Because all the code cares only about constant offsets, we do not need to go through fold_convert, > because all the codes go via int_bit_position that already expects result to be host wide int, > it seems to make sense to implement quick path for that. > > Bootstrap/regtest x86_64 in progress, OK? > > Honza > > * stor-layout.c (int_bit_from_pos): New function. > * stor-layout.h (int_bit_from_pos): Declare. > * tree.c (int_bit_from_pos): Use it. Just as a note to anyone else that peeks at this code, tree_to_shwi will verify the nodes are constant during a checking build. I'd consider a different name for the function that somehow indicates the inputs are constants. jeff Please consider an assert or other checking code to ensure that OFFSET and BITPOS are constants. Oh, I see that tree_to_shwi will get that checking when it Jeff
Index: stor-layout.c =================================================================== --- stor-layout.c (revision 215409) +++ stor-layout.c (working copy) @@ -858,6 +858,20 @@ size_binop (MULT_EXPR, offset, bitsize_unit_node)); } +/* Like int_bit_from_pos, but return the result as HOST_WIDE_INT. + OFFSET and BITPOS must be constant. */ + +HOST_WIDE_INT +int_bit_from_pos (tree offset, tree bitpos) +{ + HOST_WIDE_INT off; + if (TREE_CODE (offset) == PLUS_EXPR) + off = tree_to_shwi (TREE_OPERAND (offset, 0)) + tree_to_shwi (TREE_OPERAND (offset, 1)); + else + off = tree_to_shwi (offset); + return off * BITS_PER_UNIT + tree_to_shwi (bitpos); +} + /* Return the combined truncated byte position for the byte offset OFFSET and the bit position BITPOS. */ Index: stor-layout.h =================================================================== --- stor-layout.h (revision 215409) +++ stor-layout.h (working copy) @@ -27,6 +27,7 @@ unsigned int); extern record_layout_info start_record_layout (tree); extern tree bit_from_pos (tree, tree); +extern HOST_WIDE_INT int_bit_from_pos (tree, tree); extern tree byte_from_pos (tree, tree); extern void pos_from_bit (tree *, tree *, unsigned int, tree); extern void normalize_offset (tree *, tree *, unsigned int); Index: tree.c =================================================================== --- tree.c (revision 215409) +++ tree.c (working copy) @@ -2839,7 +2839,8 @@ HOST_WIDE_INT int_bit_position (const_tree field) { - return tree_to_shwi (bit_position (field)); + return int_bit_from_pos (DECL_FIELD_OFFSET (field), + DECL_FIELD_BIT_OFFSET (field)); } /* Return the byte position of FIELD, in bytes from the start of the record.
Hi, int_bit_position is used by ipa-devirt's type walking code. It is currently a bottleneck since I introduced speculation into contextes (I plan to solve this by changing the way i cache results). But this patch seems to make sense anyway: we do not need to go through folding: tree bit_from_pos (tree offset, tree bitpos) { if (TREE_CODE (offset) == PLUS_EXPR) offset = size_binop (PLUS_EXPR, fold_convert (bitsizetype, TREE_OPERAND (offset, 0)), fold_convert (bitsizetype, TREE_OPERAND (offset, 1))); else offset = fold_convert (bitsizetype, offset); return size_binop (PLUS_EXPR, bitpos, size_binop (MULT_EXPR, offset, bitsize_unit_node)); } Because all the code cares only about constant offsets, we do not need to go through fold_convert, because all the codes go via int_bit_position that already expects result to be host wide int, it seems to make sense to implement quick path for that. Bootstrap/regtest x86_64 in progress, OK? Honza * stor-layout.c (int_bit_from_pos): New function. * stor-layout.h (int_bit_from_pos): Declare. * tree.c (int_bit_from_pos): Use it.