b9808e5660
Loop unrolling tends to produce chains of `%x1 = add %x0, 1; %x2 = add %x1, 1; ...` with one add per unrolled iteration. This patch simplifies these adds to `%xN = add %x0, N` directly during unrolling, rather than waiting for InstCombine to do so. The motivation for this is that having a single add (rather than an add chain) on the induction variable makes it a simple recurrence, which we specially recognize in a number of places. This allows InstCombine to directly perform folds with that knowledge, instead of first folding the add chains, and then doing other folds in another InstCombine iteration. Due to the reduced number of InstCombine iterations, this also results in a small compile-time improvement. Differential Revision: https://reviews.llvm.org/D153540
236 lines
11 KiB
LLVM
236 lines
11 KiB
LLVM
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
|
|
; RUN: opt < %s -passes=loop-unroll -S -unroll-count=4 | FileCheck %s
|
|
; Test phi update after partial unroll.
|
|
|
|
declare i1 @check() nounwind
|
|
|
|
define void @test1(i32 %i, i32 %j) nounwind uwtable ssp {
|
|
;
|
|
; CHECK-LABEL: @test1(
|
|
; CHECK-NEXT: entry:
|
|
; CHECK-NEXT: [[COND1:%.*]] = call zeroext i1 @check()
|
|
; CHECK-NEXT: br i1 [[COND1]], label [[IF_THEN:%.*]], label [[IF_ELSE_LR_PH:%.*]]
|
|
; CHECK: if.else.lr.ph:
|
|
; CHECK-NEXT: br label [[IF_ELSE:%.*]]
|
|
; CHECK: if.else:
|
|
; CHECK-NEXT: [[SUB:%.*]] = phi i32 [ [[I:%.*]], [[IF_ELSE_LR_PH]] ], [ [[SUB5_3:%.*]], [[IF_ELSE_3:%.*]] ]
|
|
; CHECK-NEXT: [[SUB5:%.*]] = sub i32 [[SUB]], [[J:%.*]]
|
|
; CHECK-NEXT: [[COND2:%.*]] = call zeroext i1 @check()
|
|
; CHECK-NEXT: br i1 [[COND2]], label [[IF_THEN_LOOPEXIT:%.*]], label [[IF_ELSE_1:%.*]]
|
|
; CHECK: if.else.1:
|
|
; CHECK-NEXT: [[SUB5_1:%.*]] = sub i32 [[SUB5]], [[J]]
|
|
; CHECK-NEXT: [[COND2_1:%.*]] = call zeroext i1 @check()
|
|
; CHECK-NEXT: br i1 [[COND2_1]], label [[IF_THEN_LOOPEXIT]], label [[IF_ELSE_2:%.*]]
|
|
; CHECK: if.else.2:
|
|
; CHECK-NEXT: [[SUB5_2:%.*]] = sub i32 [[SUB5_1]], [[J]]
|
|
; CHECK-NEXT: [[COND2_2:%.*]] = call zeroext i1 @check()
|
|
; CHECK-NEXT: br i1 [[COND2_2]], label [[IF_THEN_LOOPEXIT]], label [[IF_ELSE_3]]
|
|
; CHECK: if.else.3:
|
|
; CHECK-NEXT: [[SUB5_3]] = sub i32 [[SUB5_2]], [[J]]
|
|
; CHECK-NEXT: [[COND2_3:%.*]] = call zeroext i1 @check()
|
|
; CHECK-NEXT: br i1 [[COND2_3]], label [[IF_THEN_LOOPEXIT]], label [[IF_ELSE]], !llvm.loop [[LOOP0:![0-9]+]]
|
|
; CHECK: if.then.loopexit:
|
|
; CHECK-NEXT: [[SUB5_LCSSA:%.*]] = phi i32 [ [[SUB5]], [[IF_ELSE]] ], [ [[SUB5_1]], [[IF_ELSE_1]] ], [ [[SUB5_2]], [[IF_ELSE_2]] ], [ [[SUB5_3]], [[IF_ELSE_3]] ]
|
|
; CHECK-NEXT: br label [[IF_THEN]]
|
|
; CHECK: if.then:
|
|
; CHECK-NEXT: [[I_TR:%.*]] = phi i32 [ [[I]], [[ENTRY:%.*]] ], [ [[SUB5_LCSSA]], [[IF_THEN_LOOPEXIT]] ]
|
|
; CHECK-NEXT: ret void
|
|
;
|
|
entry:
|
|
%cond1 = call zeroext i1 @check()
|
|
br i1 %cond1, label %if.then, label %if.else.lr.ph
|
|
|
|
if.else.lr.ph: ; preds = %entry
|
|
br label %if.else
|
|
|
|
if.else: ; preds = %if.else, %if.else.lr.ph
|
|
%sub = phi i32 [ %i, %if.else.lr.ph ], [ %sub5, %if.else ]
|
|
%sub5 = sub i32 %sub, %j
|
|
%cond2 = call zeroext i1 @check()
|
|
br i1 %cond2, label %if.then, label %if.else
|
|
|
|
if.then: ; preds = %if.else, %entry
|
|
%i.tr = phi i32 [ %i, %entry ], [ %sub5, %if.else ]
|
|
ret void
|
|
|
|
}
|
|
|
|
; PR7318: assertion failure after doing a simple loop unroll
|
|
;
|
|
define i32 @test2(ptr nocapture %p, i32 %n) nounwind readonly {
|
|
;
|
|
; CHECK-LABEL: @test2(
|
|
; CHECK-NEXT: entry:
|
|
; CHECK-NEXT: [[TMP0:%.*]] = icmp sgt i32 [[N:%.*]], 0
|
|
; CHECK-NEXT: br i1 [[TMP0]], label [[BB_NPH:%.*]], label [[BB2:%.*]]
|
|
; CHECK: bb.nph:
|
|
; CHECK-NEXT: [[TMP:%.*]] = zext i32 [[N]] to i64
|
|
; CHECK-NEXT: br label [[BB:%.*]]
|
|
; CHECK: bb:
|
|
; CHECK-NEXT: [[INDVAR:%.*]] = phi i64 [ 0, [[BB_NPH]] ], [ [[INDVAR_NEXT_3:%.*]], [[BB1_3:%.*]] ]
|
|
; CHECK-NEXT: [[S_01:%.*]] = phi i32 [ 0, [[BB_NPH]] ], [ [[TMP8:%.*]], [[BB1_3]] ]
|
|
; CHECK-NEXT: [[SCEVGEP:%.*]] = getelementptr i32, ptr [[P:%.*]], i64 [[INDVAR]]
|
|
; CHECK-NEXT: [[TMP1:%.*]] = load i32, ptr [[SCEVGEP]], align 1
|
|
; CHECK-NEXT: [[TMP2:%.*]] = add nsw i32 [[TMP1]], [[S_01]]
|
|
; CHECK-NEXT: br label [[BB1:%.*]]
|
|
; CHECK: bb1:
|
|
; CHECK-NEXT: [[INDVAR_NEXT:%.*]] = add nuw nsw i64 [[INDVAR]], 1
|
|
; CHECK-NEXT: [[EXITCOND:%.*]] = icmp ne i64 [[INDVAR_NEXT]], [[TMP]]
|
|
; CHECK-NEXT: br i1 [[EXITCOND]], label [[BB_1:%.*]], label [[BB1_BB2_CRIT_EDGE:%.*]]
|
|
; CHECK: bb.1:
|
|
; CHECK-NEXT: [[SCEVGEP_1:%.*]] = getelementptr i32, ptr [[P]], i64 [[INDVAR_NEXT]]
|
|
; CHECK-NEXT: [[TMP3:%.*]] = load i32, ptr [[SCEVGEP_1]], align 1
|
|
; CHECK-NEXT: [[TMP4:%.*]] = add nsw i32 [[TMP3]], [[TMP2]]
|
|
; CHECK-NEXT: br label [[BB1_1:%.*]]
|
|
; CHECK: bb1.1:
|
|
; CHECK-NEXT: [[INDVAR_NEXT_1:%.*]] = add nuw nsw i64 [[INDVAR]], 2
|
|
; CHECK-NEXT: [[EXITCOND_1:%.*]] = icmp ne i64 [[INDVAR_NEXT_1]], [[TMP]]
|
|
; CHECK-NEXT: br i1 [[EXITCOND_1]], label [[BB_2:%.*]], label [[BB1_BB2_CRIT_EDGE]]
|
|
; CHECK: bb.2:
|
|
; CHECK-NEXT: [[SCEVGEP_2:%.*]] = getelementptr i32, ptr [[P]], i64 [[INDVAR_NEXT_1]]
|
|
; CHECK-NEXT: [[TMP5:%.*]] = load i32, ptr [[SCEVGEP_2]], align 1
|
|
; CHECK-NEXT: [[TMP6:%.*]] = add nsw i32 [[TMP5]], [[TMP4]]
|
|
; CHECK-NEXT: br label [[BB1_2:%.*]]
|
|
; CHECK: bb1.2:
|
|
; CHECK-NEXT: [[INDVAR_NEXT_2:%.*]] = add nuw nsw i64 [[INDVAR]], 3
|
|
; CHECK-NEXT: [[EXITCOND_2:%.*]] = icmp ne i64 [[INDVAR_NEXT_2]], [[TMP]]
|
|
; CHECK-NEXT: br i1 [[EXITCOND_2]], label [[BB_3:%.*]], label [[BB1_BB2_CRIT_EDGE]]
|
|
; CHECK: bb.3:
|
|
; CHECK-NEXT: [[SCEVGEP_3:%.*]] = getelementptr i32, ptr [[P]], i64 [[INDVAR_NEXT_2]]
|
|
; CHECK-NEXT: [[TMP7:%.*]] = load i32, ptr [[SCEVGEP_3]], align 1
|
|
; CHECK-NEXT: [[TMP8]] = add nsw i32 [[TMP7]], [[TMP6]]
|
|
; CHECK-NEXT: br label [[BB1_3]]
|
|
; CHECK: bb1.3:
|
|
; CHECK-NEXT: [[INDVAR_NEXT_3]] = add i64 [[INDVAR]], 4
|
|
; CHECK-NEXT: [[EXITCOND_3:%.*]] = icmp ne i64 [[INDVAR_NEXT_3]], [[TMP]]
|
|
; CHECK-NEXT: br i1 [[EXITCOND_3]], label [[BB]], label [[BB1_BB2_CRIT_EDGE]], !llvm.loop [[LOOP2:![0-9]+]]
|
|
; CHECK: bb1.bb2_crit_edge:
|
|
; CHECK-NEXT: [[DOTLCSSA:%.*]] = phi i32 [ [[TMP2]], [[BB1]] ], [ [[TMP4]], [[BB1_1]] ], [ [[TMP6]], [[BB1_2]] ], [ [[TMP8]], [[BB1_3]] ]
|
|
; CHECK-NEXT: br label [[BB2]]
|
|
; CHECK: bb2:
|
|
; CHECK-NEXT: [[S_0_LCSSA:%.*]] = phi i32 [ [[DOTLCSSA]], [[BB1_BB2_CRIT_EDGE]] ], [ 0, [[ENTRY:%.*]] ]
|
|
; CHECK-NEXT: ret i32 [[S_0_LCSSA]]
|
|
;
|
|
entry:
|
|
%0 = icmp sgt i32 %n, 0 ; <i1> [#uses=1]
|
|
br i1 %0, label %bb.nph, label %bb2
|
|
|
|
bb.nph: ; preds = %entry
|
|
%tmp = zext i32 %n to i64 ; <i64> [#uses=1]
|
|
br label %bb
|
|
|
|
bb: ; preds = %bb.nph, %bb1
|
|
%indvar = phi i64 [ 0, %bb.nph ], [ %indvar.next, %bb1 ] ; <i64> [#uses=2]
|
|
%s.01 = phi i32 [ 0, %bb.nph ], [ %2, %bb1 ] ; <i32> [#uses=1]
|
|
%scevgep = getelementptr i32, ptr %p, i64 %indvar ; <ptr> [#uses=1]
|
|
%1 = load i32, ptr %scevgep, align 1 ; <i32> [#uses=1]
|
|
%2 = add nsw i32 %1, %s.01 ; <i32> [#uses=2]
|
|
br label %bb1
|
|
|
|
bb1: ; preds = %bb
|
|
%indvar.next = add i64 %indvar, 1 ; <i64> [#uses=2]
|
|
%exitcond = icmp ne i64 %indvar.next, %tmp ; <i1> [#uses=1]
|
|
br i1 %exitcond, label %bb, label %bb1.bb2_crit_edge
|
|
|
|
bb1.bb2_crit_edge: ; preds = %bb1
|
|
%.lcssa = phi i32 [ %2, %bb1 ] ; <i32> [#uses=1]
|
|
br label %bb2
|
|
|
|
bb2: ; preds = %bb1.bb2_crit_edge, %entry
|
|
%s.0.lcssa = phi i32 [ %.lcssa, %bb1.bb2_crit_edge ], [ 0, %entry ] ; <i32> [#uses=1]
|
|
ret i32 %s.0.lcssa
|
|
}
|
|
|
|
; Check phi update for loop with an early-exit.
|
|
;
|
|
define i32 @test3() nounwind uwtable ssp align 2 {
|
|
;
|
|
; CHECK-LABEL: @test3(
|
|
; CHECK-NEXT: entry:
|
|
; CHECK-NEXT: [[COND1:%.*]] = call zeroext i1 @check()
|
|
; CHECK-NEXT: br i1 [[COND1]], label [[RETURN:%.*]], label [[IF_END:%.*]]
|
|
; CHECK: if.end:
|
|
; CHECK-NEXT: br label [[DO_BODY:%.*]]
|
|
; CHECK: do.body:
|
|
; CHECK-NEXT: [[COND2:%.*]] = call zeroext i1 @check()
|
|
; CHECK-NEXT: br i1 [[COND2]], label [[EXIT:%.*]], label [[DO_COND:%.*]]
|
|
; CHECK: exit:
|
|
; CHECK-NEXT: [[TMP7_I:%.*]] = load i32, ptr undef, align 8
|
|
; CHECK-NEXT: br i1 undef, label [[DO_COND]], label [[LAND_LHS_TRUE:%.*]]
|
|
; CHECK: land.lhs.true:
|
|
; CHECK-NEXT: br i1 true, label [[RETURN_LOOPEXIT:%.*]], label [[DO_COND]]
|
|
; CHECK: do.cond:
|
|
; CHECK-NEXT: [[COND3:%.*]] = call zeroext i1 @check()
|
|
; CHECK-NEXT: br i1 [[COND3]], label [[DO_END:%.*]], label [[DO_BODY_1:%.*]]
|
|
; CHECK: do.body.1:
|
|
; CHECK-NEXT: [[COND2_1:%.*]] = call zeroext i1 @check()
|
|
; CHECK-NEXT: br i1 [[COND2_1]], label [[EXIT_1:%.*]], label [[DO_COND_1:%.*]]
|
|
; CHECK: exit.1:
|
|
; CHECK-NEXT: [[TMP7_I_1:%.*]] = load i32, ptr undef, align 8
|
|
; CHECK-NEXT: br i1 undef, label [[DO_COND_1]], label [[LAND_LHS_TRUE_1:%.*]]
|
|
; CHECK: land.lhs.true.1:
|
|
; CHECK-NEXT: br i1 true, label [[RETURN_LOOPEXIT]], label [[DO_COND_1]]
|
|
; CHECK: do.cond.1:
|
|
; CHECK-NEXT: [[COND3_1:%.*]] = call zeroext i1 @check()
|
|
; CHECK-NEXT: br i1 [[COND3_1]], label [[DO_END]], label [[DO_BODY_2:%.*]]
|
|
; CHECK: do.body.2:
|
|
; CHECK-NEXT: [[COND2_2:%.*]] = call zeroext i1 @check()
|
|
; CHECK-NEXT: br i1 [[COND2_2]], label [[EXIT_2:%.*]], label [[DO_COND_2:%.*]]
|
|
; CHECK: exit.2:
|
|
; CHECK-NEXT: [[TMP7_I_2:%.*]] = load i32, ptr undef, align 8
|
|
; CHECK-NEXT: br i1 undef, label [[DO_COND_2]], label [[LAND_LHS_TRUE_2:%.*]]
|
|
; CHECK: land.lhs.true.2:
|
|
; CHECK-NEXT: br i1 true, label [[RETURN_LOOPEXIT]], label [[DO_COND_2]]
|
|
; CHECK: do.cond.2:
|
|
; CHECK-NEXT: [[COND3_2:%.*]] = call zeroext i1 @check()
|
|
; CHECK-NEXT: br i1 [[COND3_2]], label [[DO_END]], label [[DO_BODY_3:%.*]]
|
|
; CHECK: do.body.3:
|
|
; CHECK-NEXT: [[COND2_3:%.*]] = call zeroext i1 @check()
|
|
; CHECK-NEXT: br i1 [[COND2_3]], label [[EXIT_3:%.*]], label [[DO_COND_3:%.*]]
|
|
; CHECK: exit.3:
|
|
; CHECK-NEXT: [[TMP7_I_3:%.*]] = load i32, ptr undef, align 8
|
|
; CHECK-NEXT: br i1 undef, label [[DO_COND_3]], label [[LAND_LHS_TRUE_3:%.*]]
|
|
; CHECK: land.lhs.true.3:
|
|
; CHECK-NEXT: br i1 true, label [[RETURN_LOOPEXIT]], label [[DO_COND_3]]
|
|
; CHECK: do.cond.3:
|
|
; CHECK-NEXT: [[COND3_3:%.*]] = call zeroext i1 @check()
|
|
; CHECK-NEXT: br i1 [[COND3_3]], label [[DO_END]], label [[DO_BODY]], !llvm.loop [[LOOP3:![0-9]+]]
|
|
; CHECK: do.end:
|
|
; CHECK-NEXT: br label [[RETURN]]
|
|
; CHECK: return.loopexit:
|
|
; CHECK-NEXT: [[TMP7_I_LCSSA:%.*]] = phi i32 [ [[TMP7_I]], [[LAND_LHS_TRUE]] ], [ [[TMP7_I_1]], [[LAND_LHS_TRUE_1]] ], [ [[TMP7_I_2]], [[LAND_LHS_TRUE_2]] ], [ [[TMP7_I_3]], [[LAND_LHS_TRUE_3]] ]
|
|
; CHECK-NEXT: br label [[RETURN]]
|
|
; CHECK: return:
|
|
; CHECK-NEXT: [[RETVAL_0:%.*]] = phi i32 [ 0, [[DO_END]] ], [ 0, [[ENTRY:%.*]] ], [ [[TMP7_I_LCSSA]], [[RETURN_LOOPEXIT]] ]
|
|
; CHECK-NEXT: ret i32 [[RETVAL_0]]
|
|
;
|
|
entry:
|
|
%cond1 = call zeroext i1 @check()
|
|
br i1 %cond1, label %return, label %if.end
|
|
|
|
if.end: ; preds = %entry
|
|
br label %do.body
|
|
|
|
do.body: ; preds = %do.cond, %if.end
|
|
%cond2 = call zeroext i1 @check()
|
|
br i1 %cond2, label %exit, label %do.cond
|
|
|
|
exit: ; preds = %do.body
|
|
%tmp7.i = load i32, ptr undef, align 8
|
|
br i1 undef, label %do.cond, label %land.lhs.true
|
|
|
|
land.lhs.true: ; preds = %exit
|
|
br i1 undef, label %return, label %do.cond
|
|
|
|
do.cond: ; preds = %land.lhs.true, %exit, %do.body
|
|
%cond3 = call zeroext i1 @check()
|
|
br i1 %cond3, label %do.end, label %do.body
|
|
|
|
do.end: ; preds = %do.cond
|
|
br label %return
|
|
|
|
return: ; preds = %do.end, %land.lhs.true, %entry
|
|
%retval.0 = phi i32 [ 0, %do.end ], [ 0, %entry ], [ %tmp7.i, %land.lhs.true ]
|
|
ret i32 %retval.0
|
|
}
|