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
91 lines
4.9 KiB
LLVM
91 lines
4.9 KiB
LLVM
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
|
|
; RUN: opt < %s -S -unroll-runtime -unroll-count=2 -passes=loop-unroll -unroll-runtime-epilog=true | FileCheck %s -check-prefix=EPILOG
|
|
; RUN: opt < %s -S -unroll-runtime -unroll-count=2 -passes=loop-unroll -unroll-runtime-epilog=false | FileCheck %s -check-prefix=PROLOG
|
|
target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
|
|
|
|
; This test case documents how runtime loop unrolling handles the case
|
|
; when the backedge-count is -1.
|
|
|
|
; If %N, the backedge-taken count, is -1 then %0 unsigned-overflows
|
|
; and is 0. %xtraiter too is 0, signifying that the total trip-count
|
|
; is divisible by 2. The prologue then branches to the unrolled loop
|
|
; and executes the 2^32 iterations there, in groups of 2.
|
|
|
|
define i32 @foo(i32 %N) {
|
|
; EPILOG-LABEL: @foo(
|
|
; EPILOG-NEXT: entry:
|
|
; EPILOG-NEXT: [[TMP0:%.*]] = add i32 [[N:%.*]], 1
|
|
; EPILOG-NEXT: [[XTRAITER:%.*]] = and i32 [[TMP0]], 1
|
|
; EPILOG-NEXT: [[TMP1:%.*]] = icmp ult i32 [[N]], 1
|
|
; EPILOG-NEXT: br i1 [[TMP1]], label [[WHILE_END_UNR_LCSSA:%.*]], label [[ENTRY_NEW:%.*]]
|
|
; EPILOG: entry.new:
|
|
; EPILOG-NEXT: [[UNROLL_ITER:%.*]] = sub i32 [[TMP0]], [[XTRAITER]]
|
|
; EPILOG-NEXT: br label [[WHILE_BODY:%.*]]
|
|
; EPILOG: while.body:
|
|
; EPILOG-NEXT: [[I:%.*]] = phi i32 [ 0, [[ENTRY_NEW]] ], [ [[INC_1:%.*]], [[WHILE_BODY]] ]
|
|
; EPILOG-NEXT: [[NITER:%.*]] = phi i32 [ 0, [[ENTRY_NEW]] ], [ [[NITER_NEXT_1:%.*]], [[WHILE_BODY]] ]
|
|
; EPILOG-NEXT: [[INC:%.*]] = add nuw nsw i32 [[I]], 1
|
|
; EPILOG-NEXT: [[INC_1]] = add i32 [[I]], 2
|
|
; EPILOG-NEXT: [[NITER_NEXT_1]] = add i32 [[NITER]], 2
|
|
; EPILOG-NEXT: [[NITER_NCMP_1:%.*]] = icmp eq i32 [[NITER_NEXT_1]], [[UNROLL_ITER]]
|
|
; EPILOG-NEXT: br i1 [[NITER_NCMP_1]], label [[WHILE_END_UNR_LCSSA_LOOPEXIT:%.*]], label [[WHILE_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
|
|
; EPILOG: while.end.unr-lcssa.loopexit:
|
|
; EPILOG-NEXT: [[I_LCSSA_PH_PH:%.*]] = phi i32 [ [[INC]], [[WHILE_BODY]] ]
|
|
; EPILOG-NEXT: [[I_UNR_PH:%.*]] = phi i32 [ [[INC_1]], [[WHILE_BODY]] ]
|
|
; EPILOG-NEXT: br label [[WHILE_END_UNR_LCSSA]]
|
|
; EPILOG: while.end.unr-lcssa:
|
|
; EPILOG-NEXT: [[I_LCSSA_PH:%.*]] = phi i32 [ undef, [[ENTRY:%.*]] ], [ [[I_LCSSA_PH_PH]], [[WHILE_END_UNR_LCSSA_LOOPEXIT]] ]
|
|
; EPILOG-NEXT: [[I_UNR:%.*]] = phi i32 [ 0, [[ENTRY]] ], [ [[I_UNR_PH]], [[WHILE_END_UNR_LCSSA_LOOPEXIT]] ]
|
|
; EPILOG-NEXT: [[LCMP_MOD:%.*]] = icmp ne i32 [[XTRAITER]], 0
|
|
; EPILOG-NEXT: br i1 [[LCMP_MOD]], label [[WHILE_BODY_EPIL_PREHEADER:%.*]], label [[WHILE_END:%.*]]
|
|
; EPILOG: while.body.epil.preheader:
|
|
; EPILOG-NEXT: br label [[WHILE_BODY_EPIL:%.*]]
|
|
; EPILOG: while.body.epil:
|
|
; EPILOG-NEXT: br label [[WHILE_END]]
|
|
; EPILOG: while.end:
|
|
; EPILOG-NEXT: [[I_LCSSA:%.*]] = phi i32 [ [[I_LCSSA_PH]], [[WHILE_END_UNR_LCSSA]] ], [ [[I_UNR]], [[WHILE_BODY_EPIL]] ]
|
|
; EPILOG-NEXT: ret i32 [[I_LCSSA]]
|
|
;
|
|
; PROLOG-LABEL: @foo(
|
|
; PROLOG-NEXT: entry:
|
|
; PROLOG-NEXT: [[TMP0:%.*]] = add i32 [[N:%.*]], 1
|
|
; PROLOG-NEXT: [[XTRAITER:%.*]] = and i32 [[TMP0]], 1
|
|
; PROLOG-NEXT: [[LCMP_MOD:%.*]] = icmp ne i32 [[XTRAITER]], 0
|
|
; PROLOG-NEXT: br i1 [[LCMP_MOD]], label [[WHILE_BODY_PROL_PREHEADER:%.*]], label [[WHILE_BODY_PROL_LOOPEXIT:%.*]]
|
|
; PROLOG: while.body.prol.preheader:
|
|
; PROLOG-NEXT: br label [[WHILE_BODY_PROL:%.*]]
|
|
; PROLOG: while.body.prol:
|
|
; PROLOG-NEXT: br label [[WHILE_BODY_PROL_LOOPEXIT]]
|
|
; PROLOG: while.body.prol.loopexit:
|
|
; PROLOG-NEXT: [[I_LCSSA_UNR:%.*]] = phi i32 [ undef, [[ENTRY:%.*]] ], [ 0, [[WHILE_BODY_PROL]] ]
|
|
; PROLOG-NEXT: [[I_UNR:%.*]] = phi i32 [ 0, [[ENTRY]] ], [ 1, [[WHILE_BODY_PROL]] ]
|
|
; PROLOG-NEXT: [[TMP1:%.*]] = icmp ult i32 [[N]], 1
|
|
; PROLOG-NEXT: br i1 [[TMP1]], label [[WHILE_END:%.*]], label [[ENTRY_NEW:%.*]]
|
|
; PROLOG: entry.new:
|
|
; PROLOG-NEXT: br label [[WHILE_BODY:%.*]]
|
|
; PROLOG: while.body:
|
|
; PROLOG-NEXT: [[I:%.*]] = phi i32 [ [[I_UNR]], [[ENTRY_NEW]] ], [ [[INC_1:%.*]], [[WHILE_BODY]] ]
|
|
; PROLOG-NEXT: [[INC:%.*]] = add i32 [[I]], 1
|
|
; PROLOG-NEXT: [[CMP_1:%.*]] = icmp eq i32 [[INC]], [[N]]
|
|
; PROLOG-NEXT: [[INC_1]] = add i32 [[I]], 2
|
|
; PROLOG-NEXT: br i1 [[CMP_1]], label [[WHILE_END_UNR_LCSSA:%.*]], label [[WHILE_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
|
|
; PROLOG: while.end.unr-lcssa:
|
|
; PROLOG-NEXT: [[I_LCSSA_PH:%.*]] = phi i32 [ [[INC]], [[WHILE_BODY]] ]
|
|
; PROLOG-NEXT: br label [[WHILE_END]]
|
|
; PROLOG: while.end:
|
|
; PROLOG-NEXT: [[I_LCSSA:%.*]] = phi i32 [ [[I_LCSSA_UNR]], [[WHILE_BODY_PROL_LOOPEXIT]] ], [ [[I_LCSSA_PH]], [[WHILE_END_UNR_LCSSA]] ]
|
|
; PROLOG-NEXT: ret i32 [[I_LCSSA]]
|
|
;
|
|
entry:
|
|
br label %while.body
|
|
|
|
while.body: ; preds = %while.body, %entry
|
|
%i = phi i32 [ 0, %entry ], [ %inc, %while.body ]
|
|
%cmp = icmp eq i32 %i, %N
|
|
%inc = add i32 %i, 1
|
|
br i1 %cmp, label %while.end, label %while.body
|
|
|
|
while.end: ; preds = %while.body
|
|
ret i32 %i
|
|
}
|