| Message ID | CABu31nNHZKdm98XCwT8ULBgyaE0s7WsPCuqt3ueZpm5AGJaE4w@mail.gmail.com |
|---|---|
| State | New |
| Headers | show |
On Fri, Jun 29, 2012 at 4:35 PM, Steven Bosscher <stevenb.gcc@gmail.com> wrote: > Hello, > > This patch removes the CASE_USE_BIT_TESTS target macro. The default > value of the macro is defined in stmt.c, where the only user of the > macro is also. No target re-defines the macro. > > (I wonder if the test is necessary at all. AFAICT all targets support > shifts in word_mode. The macro was originally written to test for > ashlsi3, which is _not_ supported on all targets -- but word_mode > shifts are. Oh well, another day perhaps...) > > Bootstrapped and tested on powerpc64-unknown-linux-gnu. OK for trunk? Ok. (poison CASE_USE_BIT_TESTS?) Thanks, Richard. > Ciao! > Steven > > > * stmt.c (CASE_USE_BIT_TESTS): Fold away into its only user ... > (expand_switch_using_bit_tests_p): ...here. > * doc/tm.texi.in (CASE_USE_BIT_TESTS): Remove documentation. > * doc/tm.texi (CASE_USE_BIT_TESTS): Regenerate. > > Index: stmt.c > =================================================================== > --- stmt.c (revision 189073) > +++ stmt.c (working copy) > @@ -1722,13 +1722,6 @@ add_case_node (struct case_node *head, tree type, > /* Maximum number of case bit tests. */ > #define MAX_CASE_BIT_TESTS 3 > > -/* By default, enable case bit tests on targets with ashlsi3. */ > -#ifndef CASE_USE_BIT_TESTS > -#define CASE_USE_BIT_TESTS (optab_handler (ashl_optab, word_mode) \ > - != CODE_FOR_nothing) > -#endif > - > - > /* A case_bit_test represents a set of case nodes that may be > selected from using a bit-wise comparison. HI and LO hold > the integer to be tested against, LABEL contains the label > @@ -1888,8 +1881,10 @@ bool > expand_switch_using_bit_tests_p (tree index_expr, tree range, > unsigned int uniq, unsigned int count) > { > - return (CASE_USE_BIT_TESTS > - && ! TREE_CONSTANT (index_expr) > + if (optab_handler (ashl_optab, word_mode) == CODE_FOR_nothing) > + return false; > + > + return (! TREE_CONSTANT (index_expr) > && compare_tree_int (range, GET_MODE_BITSIZE (word_mode)) < 0 > && compare_tree_int (range, 0) > 0 > && lshift_cheap_p () > Index: doc/tm.texi > =================================================================== > --- doc/tm.texi (revision 189074) > +++ doc/tm.texi (working copy) > @@ -10306,16 +10306,6 @@ The default is four for machines with a @code{case > five otherwise. This is best for most machines. > @end deftypefn > > -@defmac CASE_USE_BIT_TESTS > -Define this macro to be a C expression to indicate whether C switch > -statements may be implemented by a sequence of bit tests. This is > -advantageous on processors that can efficiently implement left shift > -of 1 by the number of bits held in a register, but inappropriate on > -targets that would require a loop. By default, this macro returns > -@code{true} if the target defines an @code{ashlsi3} pattern, and > -@code{false} otherwise. > -@end defmac > - > @defmac WORD_REGISTER_OPERATIONS > Define this macro if operations between registers with integral mode > smaller than a word are always performed on the entire register. > Index: doc/tm.texi.in > =================================================================== > --- doc/tm.texi.in (revision 189074) > +++ doc/tm.texi.in (working copy) > @@ -10180,16 +10180,6 @@ The default is four for machines with a @code{case > five otherwise. This is best for most machines. > @end deftypefn > > -@defmac CASE_USE_BIT_TESTS > -Define this macro to be a C expression to indicate whether C switch > -statements may be implemented by a sequence of bit tests. This is > -advantageous on processors that can efficiently implement left shift > -of 1 by the number of bits held in a register, but inappropriate on > -targets that would require a loop. By default, this macro returns > -@code{true} if the target defines an @code{ashlsi3} pattern, and > -@code{false} otherwise. > -@end defmac > - > @defmac WORD_REGISTER_OPERATIONS > Define this macro if operations between registers with integral mode > smaller than a word are always performed on the entire register.
On Fri, Jun 29, 2012 at 4:42 PM, Richard Guenther <richard.guenther@gmail.com> wrote: > On Fri, Jun 29, 2012 at 4:35 PM, Steven Bosscher <stevenb.gcc@gmail.com> wrote: >> Hello, >> >> This patch removes the CASE_USE_BIT_TESTS target macro. ... > Ok. (poison CASE_USE_BIT_TESTS?) Right, I've done that in the patch I commited. Ciao! Steven
Steven Bosscher schrieb: > Hello, > > This patch removes the CASE_USE_BIT_TESTS target macro. The default > value of the macro is defined in stmt.c, where the only user of the > macro is also. No target re-defines the macro. > > (I wonder if the test is necessary at all. AFAICT all targets support > shifts in word_mode. The macro was originally written to test for > ashlsi3, which is _not_ supported on all targets -- but word_mode > shifts are. Oh well, another day perhaps...) Even if a target supports ward_mode shifts with variable offsets it might be the case that it's not appropriate to use it in switch/case expansion because it is too expensive. For example, avr supports word_mode shifts but because there is no barrel shifter, a loop has to be used. Likewise for "odd" shift offsets with const_int offsets. Johann > Bootstrapped and tested on powerpc64-unknown-linux-gnu. OK for trunk? > > Ciao! > Steven > > > * stmt.c (CASE_USE_BIT_TESTS): Fold away into its only user ... > (expand_switch_using_bit_tests_p): ...here. > * doc/tm.texi.in (CASE_USE_BIT_TESTS): Remove documentation. > * doc/tm.texi (CASE_USE_BIT_TESTS): Regenerate.
On Sat, Jun 30, 2012 at 10:58 AM, Georg-Johann Lay <gjl@gcc.gnu.org> wrote: > Steven Bosscher schrieb: >> >> Hello, >> >> This patch removes the CASE_USE_BIT_TESTS target macro. The default >> value of the macro is defined in stmt.c, where the only user of the >> macro is also. No target re-defines the macro. >> >> (I wonder if the test is necessary at all. AFAICT all targets support >> shifts in word_mode. The macro was originally written to test for >> ashlsi3, which is _not_ supported on all targets -- but word_mode >> shifts are. Oh well, another day perhaps...) > > > Even if a target supports ward_mode shifts with variable offsets > it might be the case that it's not appropriate to use it in switch/case > expansion because it is too expensive. > > For example, avr supports word_mode shifts but because there is no > barrel shifter, a loop has to be used. Likewise for "odd" shift > offsets with const_int offsets. Cost is an orthogonal issue. The CASE_USE_BIT_TESTS macro tests whether the operation is supported at all for a target. There is a cost check after that to see if the shift is cheap, see lshift_cheap_p(). Ciao! Steven
Index: stmt.c =================================================================== --- stmt.c (revision 189073) +++ stmt.c (working copy) @@ -1722,13 +1722,6 @@ add_case_node (struct case_node *head, tree type, /* Maximum number of case bit tests. */ #define MAX_CASE_BIT_TESTS 3 -/* By default, enable case bit tests on targets with ashlsi3. */ -#ifndef CASE_USE_BIT_TESTS -#define CASE_USE_BIT_TESTS (optab_handler (ashl_optab, word_mode) \ - != CODE_FOR_nothing) -#endif - - /* A case_bit_test represents a set of case nodes that may be selected from using a bit-wise comparison. HI and LO hold the integer to be tested against, LABEL contains the label @@ -1888,8 +1881,10 @@ bool expand_switch_using_bit_tests_p (tree index_expr, tree range, unsigned int uniq, unsigned int count) { - return (CASE_USE_BIT_TESTS - && ! TREE_CONSTANT (index_expr) + if (optab_handler (ashl_optab, word_mode) == CODE_FOR_nothing) + return false; + + return (! TREE_CONSTANT (index_expr) && compare_tree_int (range, GET_MODE_BITSIZE (word_mode)) < 0 && compare_tree_int (range, 0) > 0 && lshift_cheap_p () Index: doc/tm.texi =================================================================== --- doc/tm.texi (revision 189074) +++ doc/tm.texi (working copy) @@ -10306,16 +10306,6 @@ The default is four for machines with a @code{case five otherwise. This is best for most machines. @end deftypefn -@defmac CASE_USE_BIT_TESTS -Define this macro to be a C expression to indicate whether C switch -statements may be implemented by a sequence of bit tests. This is -advantageous on processors that can efficiently implement left shift -of 1 by the number of bits held in a register, but inappropriate on -targets that would require a loop. By default, this macro returns -@code{true} if the target defines an @code{ashlsi3} pattern, and -@code{false} otherwise. -@end defmac - @defmac WORD_REGISTER_OPERATIONS Define this macro if operations between registers with integral mode smaller than a word are always performed on the entire register. Index: doc/tm.texi.in =================================================================== --- doc/tm.texi.in (revision 189074) +++ doc/tm.texi.in (working copy) @@ -10180,16 +10180,6 @@ The default is four for machines with a @code{case five otherwise. This is best for most machines. @end deftypefn -@defmac CASE_USE_BIT_TESTS -Define this macro to be a C expression to indicate whether C switch -statements may be implemented by a sequence of bit tests. This is -advantageous on processors that can efficiently implement left shift -of 1 by the number of bits held in a register, but inappropriate on -targets that would require a loop. By default, this macro returns -@code{true} if the target defines an @code{ashlsi3} pattern, and -@code{false} otherwise. -@end defmac - @defmac WORD_REGISTER_OPERATIONS Define this macro if operations between registers with integral mode smaller than a word are always performed on the entire register.