------- Comment #5 from uros at kss-loka dot si 2006-09-05 09:35 ------- The problem here is following:
We already have the patterns, that would satisfy combined instruction (*lshrsi3_cmp) in above testcase. However, combiner rejects combined instruction because the register that holds shifted result is unused! The problematic part is in combine.c, around line 2236 (please read the comment, which describes exactly the situation we have here). This part of code is activated only when the register that holds the result of arith operation is keept alive. This is quite strange - even if the result is unused, resulting code will be still smaller as we avoid extra CC setting instruction. The patch bellow (currently under testing, but so far OK) forces generation of combined instruction even if the arithmetic result is unused. Index: combine.c =================================================================== --- combine.c (revision 116691) +++ combine.c (working copy) @@ -2244,7 +2244,7 @@ needed, and make the PARALLEL by just replacing I2DEST in I3SRC with I2SRC. Later we will make the PARALLEL that contains I2. */ - if (i1 == 0 && added_sets_2 && GET_CODE (PATTERN (i3)) == SET + if (i1 == 0 && GET_CODE (PATTERN (i3)) == SET && GET_CODE (SET_SRC (PATTERN (i3))) == COMPARE && XEXP (SET_SRC (PATTERN (i3)), 1) == const0_rtx && rtx_equal_p (XEXP (SET_SRC (PATTERN (i3)), 0), i2dest)) @@ -2254,6 +2254,13 @@ enum machine_mode compare_mode; #endif + /* To force generation of the combined comparison and arithmetic + operation PARALLEL, pretend that the set in I2 is to be used, + even if it is dead after I2. This results in better generated + code, as only CC setting arithmetic instruction will be + emitted in conditionals. */ + added_sets_2 = 1; + newpat = PATTERN (i3); SUBST (XEXP (SET_SRC (newpat), 0), i2src); Compiling testcase with this patch results in following code: fct: movl 4(%esp), %eax shrl $5, %eax je .L2 jmp fct1 .p2align 4,,7 .L2: jmp fct2 -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=28946