> The patch was successfully tested on {i686, arm, mips}-linux, both GCC > testsuites and SPEC2000 runs. For all targets there was no observable code > difference in SPEC2000 benchmarks, so the example does not trigger very > often. Still, it speeds up CoreMark by about 1%. > > OK for trunk?
Yes, modulo the following nits: @@ -4938,11 +4938,13 @@ find_split_point (rtx *loc, rtx insn, bool set_src) IN_DEST is nonzero if we are processing the SET_DEST of a SET. + IN_COND is nonzero if we are on top level of the condition. "...we are at the top level of a condition." @@ -5221,10 +5225,12 @@ subst (rtx x, rtx from, rtx to, int in_dest, int unique_copy) expression. OP0_MODE is the original mode of XEXP (x, 0). IN_DEST is nonzero - if we are inside a SET_DEST. */ + if we are inside a SET_DEST. IN_COND is nonzero if we are on the top level + of a condition. */ Likewise. @@ -5717,7 +5723,16 @@ combine_simplify_rtx (rtx x, enum machine_mode op0_mode, int in_dest) ZERO_EXTRACT is indeed appropriate, it will be placed back by the call to make_compound_operation in the SET case. */ - if (STORE_FLAG_VALUE == 1 + if (in_cond) + /* Don't apply below optimizations if the caller would + prefer a comparison rather than a value. + E.g., for the condition in an IF_THEN_ELSE most targets need + an explicit comparison. */ + { + ; + } Remove the superfluous parentheses and move the comment to a new paragraph of the big comment just above. No need to retest, just make sure this still compiles, thanks in advance. -- Eric Botcazou