[llvm] 0c9c6dd - [IndVars] ICmpInst should not prevent IV widening
Max Kazantsev via llvm-commits
llvm-commits at lists.llvm.org
Sun Nov 29 20:17:37 PST 2020
Author: Max Kazantsev
Date: 2020-11-30T10:51:31+07:00
New Revision: 0c9c6ddf17bb01ae350a899b3395bb078aa0c62e
URL: https://github.com/llvm/llvm-project/commit/0c9c6ddf17bb01ae350a899b3395bb078aa0c62e
DIFF: https://github.com/llvm/llvm-project/commit/0c9c6ddf17bb01ae350a899b3395bb078aa0c62e.diff
LOG: [IndVars] ICmpInst should not prevent IV widening
If we decided to widen IV with zext, then unsigned comparisons
should not prevent widening (same for sext/sign comparisons).
The result of comparison in wider type does not change in this case.
Differential Revision: https://reviews.llvm.org/D92207
Reviewed By: nikic
Added:
Modified:
llvm/lib/Transforms/Utils/SimplifyIndVar.cpp
llvm/test/Transforms/IndVarSimplify/widen-loop-comp.ll
Removed:
################################################################################
diff --git a/llvm/lib/Transforms/Utils/SimplifyIndVar.cpp b/llvm/lib/Transforms/Utils/SimplifyIndVar.cpp
index d37fe74a0039..e281c66a4267 100644
--- a/llvm/lib/Transforms/Utils/SimplifyIndVar.cpp
+++ b/llvm/lib/Transforms/Utils/SimplifyIndVar.cpp
@@ -1541,10 +1541,14 @@ bool WidenIV::widenWithVariantUse(WidenIV::NarrowIVDefUse DU) {
bool CanZeroExtend = ExtKind == ZeroExtended && OBO->hasNoUnsignedWrap();
auto AnotherOpExtKind = ExtKind;
- // Check that all uses are either s/zext, or narrow def (in case of we are
- // widening the IV increment), or single-input LCSSA Phis.
+ // Check that all uses are either:
+ // - narrow def (in case of we are widening the IV increment);
+ // - single-input LCSSA Phis;
+ // - comparison of the chosen type;
+ // - extend of the chosen type (raison d'etre).
SmallVector<Instruction *, 4> ExtUsers;
SmallVector<PHINode *, 4> LCSSAPhiUsers;
+ SmallVector<ICmpInst *, 4> ICmpUsers;
for (Use &U : NarrowUse->uses()) {
Instruction *User = cast<Instruction>(U.getUser());
if (User == NarrowDef)
@@ -1558,6 +1562,19 @@ bool WidenIV::widenWithVariantUse(WidenIV::NarrowIVDefUse DU) {
LCSSAPhiUsers.push_back(LCSSAPhi);
continue;
}
+ if (auto *ICmp = dyn_cast<ICmpInst>(User)) {
+ auto Pred = ICmp->getPredicate();
+ // We have 3 types of predicates: signed, unsigned and equality
+ // predicates. For equality, it's legal to widen icmp for either sign and
+ // zero extend. For sign extend, we can also do so for signed predicates,
+ // likeweise for zero extend we can widen icmp for unsigned predicates.
+ if (ExtKind == ZeroExtended && ICmpInst::isSigned(Pred))
+ return false;
+ if (ExtKind == SignExtended && ICmpInst::isUnsigned(Pred))
+ return false;
+ ICmpUsers.push_back(ICmp);
+ continue;
+ }
if (ExtKind == SignExtended)
User = dyn_cast<SExtInst>(User);
else
@@ -1655,6 +1672,26 @@ bool WidenIV::widenWithVariantUse(WidenIV::NarrowIVDefUse DU) {
User->replaceAllUsesWith(TruncPN);
DeadInsts.emplace_back(User);
}
+
+ for (ICmpInst *User : ICmpUsers) {
+ Builder.SetInsertPoint(User);
+ auto ExtendedOp = [&](Value * V)->Value * {
+ if (V == NarrowUse)
+ return WideBO;
+ if (ExtKind == ZeroExtended)
+ return Builder.CreateZExt(V, WideBO->getType());
+ else
+ return Builder.CreateSExt(V, WideBO->getType());
+ };
+ auto Pred = User->getPredicate();
+ auto *LHS = ExtendedOp(User->getOperand(0));
+ auto *RHS = ExtendedOp(User->getOperand(1));
+ auto *WideCmp =
+ Builder.CreateICmp(Pred, LHS, RHS, User->getName() + ".wide");
+ User->replaceAllUsesWith(WideCmp);
+ DeadInsts.emplace_back(User);
+ }
+
return true;
}
diff --git a/llvm/test/Transforms/IndVarSimplify/widen-loop-comp.ll b/llvm/test/Transforms/IndVarSimplify/widen-loop-comp.ll
index dd095c008772..d3b117e73602 100644
--- a/llvm/test/Transforms/IndVarSimplify/widen-loop-comp.ll
+++ b/llvm/test/Transforms/IndVarSimplify/widen-loop-comp.ll
@@ -795,37 +795,36 @@ failure:
unreachable
}
-; TODO: We can widen here despite the icmp user of %foo in guarded block.
define i32 @test16_unsigned_pos1(i32 %start, i32* %p, i32* %q, i32 %x) {
; CHECK-LABEL: @test16_unsigned_pos1(
; CHECK-NEXT: entry:
; CHECK-NEXT: [[TMP0:%.*]] = zext i32 [[START:%.*]] to i64
-; CHECK-NEXT: [[TMP1:%.*]] = add i32 [[START]], -1
+; CHECK-NEXT: [[TMP1:%.*]] = add nsw i64 [[TMP0]], -1
+; CHECK-NEXT: [[TMP2:%.*]] = zext i32 [[X:%.*]] to i64
; CHECK-NEXT: br label [[LOOP:%.*]]
; CHECK: loop:
; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[INDVARS_IV_NEXT:%.*]], [[BACKEDGE:%.*]] ], [ [[TMP0]], [[ENTRY:%.*]] ]
; CHECK-NEXT: [[COND:%.*]] = icmp eq i64 [[INDVARS_IV]], 0
-; CHECK-NEXT: [[TMP2:%.*]] = trunc i64 [[INDVARS_IV]] to i32
-; CHECK-NEXT: [[FOO:%.*]] = add i32 [[TMP2]], -1
+; CHECK-NEXT: [[TMP3:%.*]] = add nsw i64 [[INDVARS_IV]], -1
; CHECK-NEXT: br i1 [[COND]], label [[EXIT:%.*]], label [[GUARDED:%.*]]
; CHECK: guarded:
-; CHECK-NEXT: [[ICMP_USER3:%.*]] = icmp ult i32 [[TMP1]], [[X:%.*]]
-; CHECK-NEXT: br i1 [[ICMP_USER3]], label [[BACKEDGE]], label [[SIDE_EXIT:%.*]]
+; CHECK-NEXT: [[ICMP_USER_WIDE4:%.*]] = icmp ult i64 [[TMP1]], [[TMP2]]
+; CHECK-NEXT: br i1 [[ICMP_USER_WIDE4]], label [[BACKEDGE]], label [[SIDE_EXIT:%.*]]
; CHECK: backedge:
-; CHECK-NEXT: [[INDEX:%.*]] = zext i32 [[FOO]] to i64
-; CHECK-NEXT: [[STORE_ADDR:%.*]] = getelementptr i32, i32* [[P:%.*]], i64 [[INDEX]]
+; CHECK-NEXT: [[STORE_ADDR:%.*]] = getelementptr i32, i32* [[P:%.*]], i64 [[TMP3]]
; CHECK-NEXT: store i32 1, i32* [[STORE_ADDR]], align 4
-; CHECK-NEXT: [[LOAD_ADDR:%.*]] = getelementptr i32, i32* [[Q:%.*]], i64 [[INDEX]]
-; CHECK-NEXT: [[STOP:%.*]] = load i32, i32* [[Q]], align 4
+; CHECK-NEXT: [[STOP:%.*]] = load i32, i32* [[Q:%.*]], align 4
; CHECK-NEXT: [[LOOP_COND:%.*]] = icmp eq i32 [[STOP]], 0
; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nsw i64 [[INDVARS_IV]], -1
; CHECK-NEXT: br i1 [[LOOP_COND]], label [[LOOP]], label [[FAILURE:%.*]]
; CHECK: exit:
-; CHECK-NEXT: call void @use(i32 -1)
-; CHECK-NEXT: ret i32 -1
+; CHECK-NEXT: [[TMP4:%.*]] = trunc i64 -1 to i32
+; CHECK-NEXT: call void @use(i32 [[TMP4]])
+; CHECK-NEXT: ret i32 [[TMP4]]
; CHECK: failure:
-; CHECK-NEXT: [[FOO_LCSSA2:%.*]] = phi i32 [ [[FOO]], [[BACKEDGE]] ]
-; CHECK-NEXT: call void @use(i32 [[FOO_LCSSA2]])
+; CHECK-NEXT: [[FOO_LCSSA2_WIDE:%.*]] = phi i64 [ [[TMP3]], [[BACKEDGE]] ]
+; CHECK-NEXT: [[TMP5:%.*]] = trunc i64 [[FOO_LCSSA2_WIDE]] to i32
+; CHECK-NEXT: call void @use(i32 [[TMP5]])
; CHECK-NEXT: unreachable
; CHECK: side_exit:
; CHECK-NEXT: ret i32 0
More information about the llvm-commits
mailing list