| Message ID | 20201007225800.9536-8-amodra@gmail.com |
|---|---|
| State | New |
| Headers | show |
| Series | rs6000_rtx_costs V2 | expand |
On Thu, 2020-10-08 at 09:27 +1030, Alan Modra via Gcc-patches wrote: > The aim of this patch is to make rtx_costs for SETs closer to > insn_cost for SETs. One visible effect on powerpc code is increased > if-conversion. > > * config/rs6000/rs6000.c (rs6000_rtx_costs): Reduce cost of SET > operands. > > diff --git a/gcc/config/rs6000/rs6000.c b/gcc/config/rs6000/rs6000.c > index 76aedbfae6f..d455aa52427 100644 > --- a/gcc/config/rs6000/rs6000.c > +++ b/gcc/config/rs6000/rs6000.c > @@ -21684,6 +21684,35 @@ rs6000_rtx_costs (rtx x, machine_mode mode, > int outer_code, > } > return false; > > + case SET: > + /* On entry the value in *TOTAL is the number of general > purpose > + regs being set, multiplied by COSTS_N_INSNS (1). Handle > + costing of set operands specially since in most cases we have > + an instruction rather than just a piece of RTL and should > + return a cost comparable to insn_cost. That's a little > + complicated because in some cases the cost of SET operands is > + non-zero, see point 5 above and cost of PLUS for example, and > + in others it is zero, for example for (set (reg) (reg)). > + But (set (reg) (reg)) has the same insn_cost as > + (set (reg) (plus (reg) (reg))). Hack around this by > + subtracting COSTS_N_INSNS (1) from the operand cost in cases > + were we add at least COSTS_N_INSNS (1) for some operation. s/were/where/ :-) > + However, don't do so for constants. Constants might cost > + more than zero when they require more than one instruction, > + and we do want the cost of extra instructions. */ > + { > + rtx_code src_code = GET_CODE (SET_SRC (x)); > + if (src_code == CONST_INT > + || src_code == CONST_DOUBLE > + || src_code == CONST_WIDE_INT) > + return false; > + int set_cost = (rtx_cost (SET_SRC (x), mode, SET, 1, speed) > + + rtx_cost (SET_DEST (x), mode, SET, 0, > speed)); > + if (set_cost >= COSTS_N_INSNS (1)) > + *total += set_cost - COSTS_N_INSNS (1); > + return true; > + } > + > default: > return false; > } lgtm, thanks -Will
Ping. On Tue, Jan 12, 2021 at 02:02:27PM +1030, Alan Modra wrote: > Ping > https://gcc.gnu.org/pipermail/gcc-patches/2020-October/555758.html > > On Thu, Oct 08, 2020 at 09:27:59AM +1030, Alan Modra wrote: > > The aim of this patch is to make rtx_costs for SETs closer to > > insn_cost for SETs. One visible effect on powerpc code is increased > > if-conversion. > > > > * config/rs6000/rs6000.c (rs6000_rtx_costs): Reduce cost of SET > > operands. > > > > diff --git a/gcc/config/rs6000/rs6000.c b/gcc/config/rs6000/rs6000.c > > index 76aedbfae6f..d455aa52427 100644 > > --- a/gcc/config/rs6000/rs6000.c > > +++ b/gcc/config/rs6000/rs6000.c > > @@ -21684,6 +21684,35 @@ rs6000_rtx_costs (rtx x, machine_mode mode, int outer_code, > > } > > return false; > > > > + case SET: > > + /* On entry the value in *TOTAL is the number of general purpose > > + regs being set, multiplied by COSTS_N_INSNS (1). Handle > > + costing of set operands specially since in most cases we have > > + an instruction rather than just a piece of RTL and should > > + return a cost comparable to insn_cost. That's a little > > + complicated because in some cases the cost of SET operands is > > + non-zero, see point 5 above and cost of PLUS for example, and > > + in others it is zero, for example for (set (reg) (reg)). > > + But (set (reg) (reg)) has the same insn_cost as > > + (set (reg) (plus (reg) (reg))). Hack around this by > > + subtracting COSTS_N_INSNS (1) from the operand cost in cases > > + were we add at least COSTS_N_INSNS (1) for some operation. > > + However, don't do so for constants. Constants might cost > > + more than zero when they require more than one instruction, > > + and we do want the cost of extra instructions. */ > > + { > > + rtx_code src_code = GET_CODE (SET_SRC (x)); > > + if (src_code == CONST_INT > > + || src_code == CONST_DOUBLE > > + || src_code == CONST_WIDE_INT) > > + return false; > > + int set_cost = (rtx_cost (SET_SRC (x), mode, SET, 1, speed) > > + + rtx_cost (SET_DEST (x), mode, SET, 0, speed)); > > + if (set_cost >= COSTS_N_INSNS (1)) > > + *total += set_cost - COSTS_N_INSNS (1); > > + return true; > > + } > > + > > default: > > return false; > > }
diff --git a/gcc/config/rs6000/rs6000.c b/gcc/config/rs6000/rs6000.c index 76aedbfae6f..d455aa52427 100644 --- a/gcc/config/rs6000/rs6000.c +++ b/gcc/config/rs6000/rs6000.c @@ -21684,6 +21684,35 @@ rs6000_rtx_costs (rtx x, machine_mode mode, int outer_code, } return false; + case SET: + /* On entry the value in *TOTAL is the number of general purpose + regs being set, multiplied by COSTS_N_INSNS (1). Handle + costing of set operands specially since in most cases we have + an instruction rather than just a piece of RTL and should + return a cost comparable to insn_cost. That's a little + complicated because in some cases the cost of SET operands is + non-zero, see point 5 above and cost of PLUS for example, and + in others it is zero, for example for (set (reg) (reg)). + But (set (reg) (reg)) has the same insn_cost as + (set (reg) (plus (reg) (reg))). Hack around this by + subtracting COSTS_N_INSNS (1) from the operand cost in cases + were we add at least COSTS_N_INSNS (1) for some operation. + However, don't do so for constants. Constants might cost + more than zero when they require more than one instruction, + and we do want the cost of extra instructions. */ + { + rtx_code src_code = GET_CODE (SET_SRC (x)); + if (src_code == CONST_INT + || src_code == CONST_DOUBLE + || src_code == CONST_WIDE_INT) + return false; + int set_cost = (rtx_cost (SET_SRC (x), mode, SET, 1, speed) + + rtx_cost (SET_DEST (x), mode, SET, 0, speed)); + if (set_cost >= COSTS_N_INSNS (1)) + *total += set_cost - COSTS_N_INSNS (1); + return true; + } + default: return false; }