Skip to content

Commit 274dae9

Browse files
dotdashdotdash
committed
Improve codegen for the various "with overflow" intrinsics
We're currently possibly introducing an unneeded temporary, make use of InsertValue which is said to kick us off of FastISel and we generate loads/stores of first class aggregates, which is bad as well. Let's not do all these things.
1 parent 2375743 commit 274dae9

File tree

1 file changed

+29
-35
lines changed

1 file changed

+29
-35
lines changed

‎src/librustc_trans/trans/intrinsic.rs

+29-35
Original file line numberDiff line numberDiff line change
@@ -616,171 +616,171 @@ pub fn trans_intrinsic_call<'a, 'blk, 'tcx>(mut bcx: Block<'blk, 'tcx>,
616616
(_, "i8_add_with_overflow") =>
617617
with_overflow_intrinsic(bcx,
618618
"llvm.sadd.with.overflow.i8",
619-
ret_ty,
620619
llargs[0],
621620
llargs[1],
621+
llresult,
622622
call_debug_location),
623623
(_, "i16_add_with_overflow") =>
624624
with_overflow_intrinsic(bcx,
625625
"llvm.sadd.with.overflow.i16",
626-
ret_ty,
627626
llargs[0],
628627
llargs[1],
628+
llresult,
629629
call_debug_location),
630630
(_, "i32_add_with_overflow") =>
631631
with_overflow_intrinsic(bcx,
632632
"llvm.sadd.with.overflow.i32",
633-
ret_ty,
634633
llargs[0],
635634
llargs[1],
635+
llresult,
636636
call_debug_location),
637637
(_, "i64_add_with_overflow") =>
638638
with_overflow_intrinsic(bcx,
639639
"llvm.sadd.with.overflow.i64",
640-
ret_ty,
641640
llargs[0],
642641
llargs[1],
642+
llresult,
643643
call_debug_location),
644644

645645
(_, "u8_add_with_overflow") =>
646646
with_overflow_intrinsic(bcx,
647647
"llvm.uadd.with.overflow.i8",
648-
ret_ty,
649648
llargs[0],
650649
llargs[1],
650+
llresult,
651651
call_debug_location),
652652
(_, "u16_add_with_overflow") =>
653653
with_overflow_intrinsic(bcx,
654654
"llvm.uadd.with.overflow.i16",
655-
ret_ty,
656655
llargs[0],
657656
llargs[1],
657+
llresult,
658658
call_debug_location),
659659
(_, "u32_add_with_overflow") =>
660660
with_overflow_intrinsic(bcx,
661661
"llvm.uadd.with.overflow.i32",
662-
ret_ty,
663662
llargs[0],
664663
llargs[1],
664+
llresult,
665665
call_debug_location),
666666
(_, "u64_add_with_overflow") =>
667667
with_overflow_intrinsic(bcx,
668668
"llvm.uadd.with.overflow.i64",
669-
ret_ty,
670669
llargs[0],
671670
llargs[1],
671+
llresult,
672672
call_debug_location),
673673
(_, "i8_sub_with_overflow") =>
674674
with_overflow_intrinsic(bcx,
675675
"llvm.ssub.with.overflow.i8",
676-
ret_ty,
677676
llargs[0],
678677
llargs[1],
678+
llresult,
679679
call_debug_location),
680680
(_, "i16_sub_with_overflow") =>
681681
with_overflow_intrinsic(bcx,
682682
"llvm.ssub.with.overflow.i16",
683-
ret_ty,
684683
llargs[0],
685684
llargs[1],
685+
llresult,
686686
call_debug_location),
687687
(_, "i32_sub_with_overflow") =>
688688
with_overflow_intrinsic(bcx,
689689
"llvm.ssub.with.overflow.i32",
690-
ret_ty,
691690
llargs[0],
692691
llargs[1],
692+
llresult,
693693
call_debug_location),
694694
(_, "i64_sub_with_overflow") =>
695695
with_overflow_intrinsic(bcx,
696696
"llvm.ssub.with.overflow.i64",
697-
ret_ty,
698697
llargs[0],
699698
llargs[1],
699+
llresult,
700700
call_debug_location),
701701
(_, "u8_sub_with_overflow") =>
702702
with_overflow_intrinsic(bcx,
703703
"llvm.usub.with.overflow.i8",
704-
ret_ty,
705704
llargs[0],
706705
llargs[1],
706+
llresult,
707707
call_debug_location),
708708
(_, "u16_sub_with_overflow") =>
709709
with_overflow_intrinsic(bcx,
710710
"llvm.usub.with.overflow.i16",
711-
ret_ty,
712711
llargs[0],
713712
llargs[1],
713+
llresult,
714714
call_debug_location),
715715
(_, "u32_sub_with_overflow") =>
716716
with_overflow_intrinsic(bcx,
717717
"llvm.usub.with.overflow.i32",
718-
ret_ty,
719718
llargs[0],
720719
llargs[1],
720+
llresult,
721721
call_debug_location),
722722
(_, "u64_sub_with_overflow") =>
723723
with_overflow_intrinsic(bcx,
724724
"llvm.usub.with.overflow.i64",
725-
ret_ty,
726725
llargs[0],
727726
llargs[1],
727+
llresult,
728728
call_debug_location),
729729
(_, "i8_mul_with_overflow") =>
730730
with_overflow_intrinsic(bcx,
731731
"llvm.smul.with.overflow.i8",
732-
ret_ty,
733732
llargs[0],
734733
llargs[1],
734+
llresult,
735735
call_debug_location),
736736
(_, "i16_mul_with_overflow") =>
737737
with_overflow_intrinsic(bcx,
738738
"llvm.smul.with.overflow.i16",
739-
ret_ty,
740739
llargs[0],
741740
llargs[1],
741+
llresult,
742742
call_debug_location),
743743
(_, "i32_mul_with_overflow") =>
744744
with_overflow_intrinsic(bcx,
745745
"llvm.smul.with.overflow.i32",
746-
ret_ty,
747746
llargs[0],
748747
llargs[1],
748+
llresult,
749749
call_debug_location),
750750
(_, "i64_mul_with_overflow") =>
751751
with_overflow_intrinsic(bcx,
752752
"llvm.smul.with.overflow.i64",
753-
ret_ty,
754753
llargs[0],
755754
llargs[1],
755+
llresult,
756756
call_debug_location),
757757
(_, "u8_mul_with_overflow") =>
758758
with_overflow_intrinsic(bcx,
759759
"llvm.umul.with.overflow.i8",
760-
ret_ty,
761760
llargs[0],
762761
llargs[1],
762+
llresult,
763763
call_debug_location),
764764
(_, "u16_mul_with_overflow") =>
765765
with_overflow_intrinsic(bcx,
766766
"llvm.umul.with.overflow.i16",
767-
ret_ty,
768767
llargs[0],
769768
llargs[1],
769+
llresult,
770770
call_debug_location),
771771
(_, "u32_mul_with_overflow") =>
772772
with_overflow_intrinsic(bcx,
773773
"llvm.umul.with.overflow.i32",
774-
ret_ty,
775774
llargs[0],
776775
llargs[1],
776+
llresult,
777777
call_debug_location),
778778
(_, "u64_mul_with_overflow") =>
779779
with_overflow_intrinsic(bcx,
780780
"llvm.umul.with.overflow.i64",
781-
ret_ty,
782781
llargs[0],
783782
llargs[1],
783+
llresult,
784784
call_debug_location),
785785

786786
(_, "unchecked_udiv") => UDiv(bcx, llargs[0], llargs[1], call_debug_location),
@@ -1053,9 +1053,9 @@ fn count_zeros_intrinsic(bcx: Block,
10531053

10541054
fn with_overflow_intrinsic<'blk, 'tcx>(bcx: Block<'blk, 'tcx>,
10551055
name: &'static str,
1056-
t: Ty<'tcx>,
10571056
a: ValueRef,
10581057
b: ValueRef,
1058+
out: ValueRef,
10591059
call_debug_location: DebugLoc)
10601060
-> ValueRef {
10611061
let llfn = bcx.ccx().get_intrinsic(&name);
@@ -1064,16 +1064,10 @@ fn with_overflow_intrinsic<'blk, 'tcx>(bcx: Block<'blk, 'tcx>,
10641064
let val = Call(bcx, llfn, &[a, b], None, call_debug_location);
10651065
let result = ExtractValue(bcx, val, 0);
10661066
let overflow = ZExt(bcx, ExtractValue(bcx, val, 1), Type::bool(bcx.ccx()));
1067-
let ret = C_undef(type_of::type_of(bcx.ccx(), t));
1068-
let ret = InsertValue(bcx, ret, result, 0);
1069-
let ret = InsertValue(bcx, ret, overflow, 1);
1070-
if !arg_is_indirect(bcx.ccx(), t) {
1071-
let tmp = alloc_ty(bcx, t, "tmp");
1072-
Store(bcx, ret, tmp);
1073-
load_ty(bcx, tmp, t)
1074-
} else {
1075-
ret
1076-
}
1067+
Store(bcx, result, StructGEP(bcx, out, 0));
1068+
Store(bcx, overflow, StructGEP(bcx, out, 1));
1069+
1070+
C_nil(bcx.ccx())
10771071
}
10781072

10791073
fn try_intrinsic<'blk, 'tcx>(bcx: Block<'blk, 'tcx>,

0 commit comments

Comments
 (0)