[LICM][MustExec] Make must-exec logic for IV condition commutative #93150
+147
−160
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
MustExec has special logic to determine whether the first loop iteration will always be executed, by simplifying the IV comparison with the start value. Currently, this code assumes that the IV is on the LHS of the comparison, but this is not guaranteed. Make sure it handles the commuted variant as well. The changed PhaseOrdering test previously performed peeling to make the loads dereferenceable -- as a side effect, this also reduced the exit count by one, avoiding the awkward <= MAX case. Now we know up-front the the loads are dereferenceable and can be simply hoisted. As such, we retain the original exit count and now have to handle it by widening the exit count calculation to i128. This is a regression, but at least it preserves the vectorization, which was the original goal. I'm not sure what else can be done about that test.
|
@llvm/pr-subscribers-llvm-analysis Author: Nikita Popov (nikic) ChangesMustExec has special logic to determine whether the first loop iteration will always be executed, by simplifying the IV comparison with the start value. Currently, this code assumes that the IV is on the LHS of the comparison, but this is not guaranteed. Make sure it handles the commuted variant as well. The changed PhaseOrdering test previously performed peeling to make the loads dereferenceable -- as a side effect, this also reduced the exit count by one, avoiding the awkward <= MAX case. Now we know up-front the the loads are dereferenceable and can be simply hoisted. As such, we retain the original exit count and now have to handle it by widening the exit count calculation to i128. This is a regression, but at least it preserves the vectorization, which was the original goal. I'm not sure what else can be done about that test. Patch is 27.98 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/93150.diff 3 Files Affected:
diff --git a/llvm/lib/Analysis/MustExecute.cpp b/llvm/lib/Analysis/MustExecute.cpp
index d4b31f2b00187..caed62679a683 100644
--- a/llvm/lib/Analysis/MustExecute.cpp
+++ b/llvm/lib/Analysis/MustExecute.cpp
@@ -135,16 +135,21 @@ static bool CanProveNotTakenFirstIteration(const BasicBlock *ExitBlock,
// todo: this would be a lot more powerful if we used scev, but all the
// plumbing is currently missing to pass a pointer in from the pass
// Check for cmp (phi [x, preheader] ...), y where (pred x, y is known
+ ICmpInst::Predicate Pred = Cond->getPredicate();
auto *LHS = dyn_cast<PHINode>(Cond->getOperand(0));
auto *RHS = Cond->getOperand(1);
- if (!LHS || LHS->getParent() != CurLoop->getHeader())
- return false;
+ if (!LHS || LHS->getParent() != CurLoop->getHeader()) {
+ Pred = Cond->getSwappedPredicate();
+ LHS = dyn_cast<PHINode>(Cond->getOperand(1));
+ RHS = Cond->getOperand(0);
+ if (!LHS || LHS->getParent() != CurLoop->getHeader())
+ return false;
+ }
+
auto DL = ExitBlock->getModule()->getDataLayout();
auto *IVStart = LHS->getIncomingValueForBlock(CurLoop->getLoopPreheader());
- auto *SimpleValOrNull = simplifyCmpInst(Cond->getPredicate(),
- IVStart, RHS,
- {DL, /*TLI*/ nullptr,
- DT, /*AC*/ nullptr, BI});
+ auto *SimpleValOrNull = simplifyCmpInst(
+ Pred, IVStart, RHS, {DL, /*TLI*/ nullptr, DT, /*AC*/ nullptr, BI});
auto *SimpleCst = dyn_cast_or_null<Constant>(SimpleValOrNull);
if (!SimpleCst)
return false;
diff --git a/llvm/test/Transforms/LICM/hoist-mustexec.ll b/llvm/test/Transforms/LICM/hoist-mustexec.ll
index 81e0815053ffe..a6f5a2be05ee4 100644
--- a/llvm/test/Transforms/LICM/hoist-mustexec.ll
+++ b/llvm/test/Transforms/LICM/hoist-mustexec.ll
@@ -218,7 +218,6 @@ fail:
}
; Same as previous case, with commuted icmp.
-; FIXME: The load should get hoisted here as well.
define i32 @test3_commuted(ptr noalias nocapture readonly %a) nounwind uwtable {
; CHECK-LABEL: define i32 @test3_commuted(
; CHECK-SAME: ptr noalias nocapture readonly [[A:%.*]]) #[[ATTR1]] {
@@ -227,6 +226,7 @@ define i32 @test3_commuted(ptr noalias nocapture readonly %a) nounwind uwtable {
; CHECK-NEXT: [[IS_ZERO:%.*]] = icmp eq i32 [[LEN]], 0
; CHECK-NEXT: br i1 [[IS_ZERO]], label [[FAIL:%.*]], label [[PREHEADER:%.*]]
; CHECK: preheader:
+; CHECK-NEXT: [[I1:%.*]] = load i32, ptr [[A]], align 4
; CHECK-NEXT: br label [[FOR_BODY:%.*]]
; CHECK: for.body:
; CHECK-NEXT: [[IV:%.*]] = phi i32 [ 0, [[PREHEADER]] ], [ [[INC:%.*]], [[CONTINUE:%.*]] ]
@@ -234,7 +234,6 @@ define i32 @test3_commuted(ptr noalias nocapture readonly %a) nounwind uwtable {
; CHECK-NEXT: [[R_CHK:%.*]] = icmp uge i32 [[LEN]], [[IV]]
; CHECK-NEXT: br i1 [[R_CHK]], label [[CONTINUE]], label [[FAIL_LOOPEXIT:%.*]]
; CHECK: continue:
-; CHECK-NEXT: [[I1:%.*]] = load i32, ptr [[A]], align 4
; CHECK-NEXT: [[ADD]] = add nsw i32 [[I1]], [[ACC]]
; CHECK-NEXT: [[INC]] = add nuw nsw i32 [[IV]], 1
; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i32 [[INC]], 1000
diff --git a/llvm/test/Transforms/PhaseOrdering/AArch64/peel-multiple-unreachable-exits-for-vectorization.ll b/llvm/test/Transforms/PhaseOrdering/AArch64/peel-multiple-unreachable-exits-for-vectorization.ll
index 8fc5189e8bc79..cc4890e27f2bd 100644
--- a/llvm/test/Transforms/PhaseOrdering/AArch64/peel-multiple-unreachable-exits-for-vectorization.ll
+++ b/llvm/test/Transforms/PhaseOrdering/AArch64/peel-multiple-unreachable-exits-for-vectorization.ll
@@ -9,95 +9,87 @@
define i64 @sum_2_at_with_int_conversion(ptr %A, ptr %B, i64 %N) {
; CHECK-LABEL: @sum_2_at_with_int_conversion(
-; CHECK-NEXT: at_with_int_conversion.exit11.peel:
+; CHECK-NEXT: entry:
; CHECK-NEXT: [[START_I:%.*]] = load ptr, ptr [[A:%.*]], align 8
; CHECK-NEXT: [[GEP_END_I:%.*]] = getelementptr i8, ptr [[A]], i64 8
; CHECK-NEXT: [[END_I:%.*]] = load ptr, ptr [[GEP_END_I]], align 8
; CHECK-NEXT: [[START_INT_I:%.*]] = ptrtoint ptr [[START_I]] to i64
; CHECK-NEXT: [[END_INT_I:%.*]] = ptrtoint ptr [[END_I]] to i64
; CHECK-NEXT: [[SUB_I:%.*]] = sub i64 [[END_INT_I]], [[START_INT_I]]
+; CHECK-NEXT: [[START_I1:%.*]] = load ptr, ptr [[B:%.*]], align 8
+; CHECK-NEXT: [[GEP_END_I2:%.*]] = getelementptr i8, ptr [[B]], i64 8
+; CHECK-NEXT: [[END_I3:%.*]] = load ptr, ptr [[GEP_END_I2]], align 8
+; CHECK-NEXT: [[START_INT_I4:%.*]] = ptrtoint ptr [[START_I1]] to i64
+; CHECK-NEXT: [[END_INT_I5:%.*]] = ptrtoint ptr [[END_I3]] to i64
+; CHECK-NEXT: [[SUB_I6:%.*]] = sub i64 [[END_INT_I5]], [[START_INT_I4]]
+; CHECK-NEXT: [[TMP0:%.*]] = zext i64 [[SUB_I]] to i128
+; CHECK-NEXT: [[TMP1:%.*]] = add nuw nsw i128 [[TMP0]], 1
+; CHECK-NEXT: [[TMP2:%.*]] = zext i64 [[SUB_I6]] to i128
+; CHECK-NEXT: [[TMP3:%.*]] = add nuw nsw i128 [[TMP2]], 1
+; CHECK-NEXT: [[UMIN:%.*]] = tail call i128 @llvm.umin.i128(i128 [[TMP3]], i128 [[TMP1]])
; CHECK-NEXT: [[SMAX:%.*]] = tail call i64 @llvm.smax.i64(i64 [[N:%.*]], i64 0)
-; CHECK-NEXT: [[GEP_END_I2:%.*]] = getelementptr i8, ptr [[B:%.*]], i64 8
-; CHECK-NEXT: [[START_I1_PEEL:%.*]] = load ptr, ptr [[B]], align 8
-; CHECK-NEXT: [[END_I3_PEEL:%.*]] = load ptr, ptr [[GEP_END_I2]], align 8
-; CHECK-NEXT: [[START_INT_I4_PEEL:%.*]] = ptrtoint ptr [[START_I1_PEEL]] to i64
-; CHECK-NEXT: [[END_INT_I5_PEEL:%.*]] = ptrtoint ptr [[END_I3_PEEL]] to i64
-; CHECK-NEXT: [[SUB_I6_PEEL:%.*]] = sub i64 [[END_INT_I5_PEEL]], [[START_INT_I4_PEEL]]
-; CHECK-NEXT: [[LV_I_PEEL:%.*]] = load i64, ptr [[START_I]], align 8
-; CHECK-NEXT: [[LV_I9_PEEL:%.*]] = load i64, ptr [[START_I1_PEEL]], align 8
-; CHECK-NEXT: [[SUM_NEXT_PEEL:%.*]] = add i64 [[LV_I_PEEL]], [[LV_I9_PEEL]]
-; CHECK-NEXT: [[EXITCOND_PEEL_NOT:%.*]] = icmp slt i64 [[N]], 1
-; CHECK-NEXT: br i1 [[EXITCOND_PEEL_NOT]], label [[EXIT:%.*]], label [[LOOP_PREHEADER:%.*]]
+; CHECK-NEXT: [[TMP4:%.*]] = zext nneg i64 [[SMAX]] to i128
+; CHECK-NEXT: [[UMIN12:%.*]] = tail call i128 @llvm.umin.i128(i128 [[UMIN]], i128 [[TMP4]])
+; CHECK-NEXT: [[TMP5:%.*]] = icmp eq i128 [[TMP1]], [[UMIN12]]
+; CHECK-NEXT: br i1 [[TMP5]], label [[ERROR_I:%.*]], label [[ENTRY_SPLIT:%.*]]
+; CHECK: entry.split:
+; CHECK-NEXT: [[TMP6:%.*]] = icmp eq i128 [[TMP3]], [[UMIN12]]
+; CHECK-NEXT: br i1 [[TMP6]], label [[ERROR_I10:%.*]], label [[LOOP_PREHEADER:%.*]]
; CHECK: loop.preheader:
-; CHECK-NEXT: [[TMP0:%.*]] = add nsw i64 [[SMAX]], -1
-; CHECK-NEXT: [[UMIN:%.*]] = tail call i64 @llvm.umin.i64(i64 [[SUB_I6_PEEL]], i64 [[TMP0]])
-; CHECK-NEXT: [[TMP1:%.*]] = freeze i64 [[UMIN]]
-; CHECK-NEXT: [[UMIN15:%.*]] = tail call i64 @llvm.umin.i64(i64 [[TMP1]], i64 [[SUB_I]])
-; CHECK-NEXT: [[TMP2:%.*]] = add i64 [[UMIN15]], 1
-; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP2]], 5
-; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[LOOP_PREHEADER20:%.*]], label [[VECTOR_PH:%.*]]
+; CHECK-NEXT: [[TMP7:%.*]] = add nuw i64 [[SMAX]], 1
+; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp slt i64 [[N]], 3
+; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[LOOP_PREHEADER17:%.*]], label [[VECTOR_PH:%.*]]
; CHECK: vector.ph:
-; CHECK-NEXT: [[N_MOD_VF:%.*]] = and i64 [[TMP2]], 3
-; CHECK-NEXT: [[TMP3:%.*]] = icmp eq i64 [[N_MOD_VF]], 0
-; CHECK-NEXT: [[TMP4:%.*]] = select i1 [[TMP3]], i64 4, i64 [[N_MOD_VF]]
-; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[TMP2]], [[TMP4]]
-; CHECK-NEXT: [[IND_END:%.*]] = add i64 [[N_VEC]], 1
-; CHECK-NEXT: [[TMP5:%.*]] = insertelement <2 x i64> <i64 poison, i64 0>, i64 [[SUM_NEXT_PEEL]], i64 0
+; CHECK-NEXT: [[N_VEC:%.*]] = and i64 [[TMP7]], -4
; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]
; CHECK: vector.body:
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT: [[VEC_PHI:%.*]] = phi <2 x i64> [ [[TMP5]], [[VECTOR_PH]] ], [ [[TMP12:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT: [[VEC_PHI16:%.*]] = phi <2 x i64> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP13:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT: [[OFFSET_IDX:%.*]] = or disjoint i64 [[INDEX]], 1
-; CHECK-NEXT: [[TMP6:%.*]] = getelementptr i64, ptr [[START_I]], i64 [[OFFSET_IDX]]
-; CHECK-NEXT: [[TMP7:%.*]] = getelementptr i8, ptr [[TMP6]], i64 16
-; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x i64>, ptr [[TMP6]], align 8
-; CHECK-NEXT: [[WIDE_LOAD17:%.*]] = load <2 x i64>, ptr [[TMP7]], align 8
-; CHECK-NEXT: [[TMP8:%.*]] = getelementptr i64, ptr [[START_I1_PEEL]], i64 [[OFFSET_IDX]]
+; CHECK-NEXT: [[VEC_PHI:%.*]] = phi <2 x i64> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP14:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT: [[VEC_PHI13:%.*]] = phi <2 x i64> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP15:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT: [[TMP8:%.*]] = getelementptr i64, ptr [[START_I]], i64 [[INDEX]]
; CHECK-NEXT: [[TMP9:%.*]] = getelementptr i8, ptr [[TMP8]], i64 16
-; CHECK-NEXT: [[WIDE_LOAD18:%.*]] = load <2 x i64>, ptr [[TMP8]], align 8
-; CHECK-NEXT: [[WIDE_LOAD19:%.*]] = load <2 x i64>, ptr [[TMP9]], align 8
-; CHECK-NEXT: [[TMP10:%.*]] = add <2 x i64> [[WIDE_LOAD]], [[VEC_PHI]]
-; CHECK-NEXT: [[TMP11:%.*]] = add <2 x i64> [[WIDE_LOAD17]], [[VEC_PHI16]]
-; CHECK-NEXT: [[TMP12]] = add <2 x i64> [[TMP10]], [[WIDE_LOAD18]]
-; CHECK-NEXT: [[TMP13]] = add <2 x i64> [[TMP11]], [[WIDE_LOAD19]]
+; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x i64>, ptr [[TMP8]], align 8
+; CHECK-NEXT: [[WIDE_LOAD14:%.*]] = load <2 x i64>, ptr [[TMP9]], align 8
+; CHECK-NEXT: [[TMP10:%.*]] = getelementptr i64, ptr [[START_I1]], i64 [[INDEX]]
+; CHECK-NEXT: [[TMP11:%.*]] = getelementptr i8, ptr [[TMP10]], i64 16
+; CHECK-NEXT: [[WIDE_LOAD15:%.*]] = load <2 x i64>, ptr [[TMP10]], align 8
+; CHECK-NEXT: [[WIDE_LOAD16:%.*]] = load <2 x i64>, ptr [[TMP11]], align 8
+; CHECK-NEXT: [[TMP12:%.*]] = add <2 x i64> [[WIDE_LOAD]], [[VEC_PHI]]
+; CHECK-NEXT: [[TMP13:%.*]] = add <2 x i64> [[WIDE_LOAD14]], [[VEC_PHI13]]
+; CHECK-NEXT: [[TMP14]] = add <2 x i64> [[TMP12]], [[WIDE_LOAD15]]
+; CHECK-NEXT: [[TMP15]] = add <2 x i64> [[TMP13]], [[WIDE_LOAD16]]
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
-; CHECK-NEXT: [[TMP14:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
-; CHECK-NEXT: br i1 [[TMP14]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; CHECK-NEXT: [[TMP16:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-NEXT: br i1 [[TMP16]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
; CHECK: middle.block:
-; CHECK-NEXT: [[BIN_RDX:%.*]] = add <2 x i64> [[TMP13]], [[TMP12]]
-; CHECK-NEXT: [[TMP15:%.*]] = tail call i64 @llvm.vector.reduce.add.v2i64(<2 x i64> [[BIN_RDX]])
-; CHECK-NEXT: br label [[LOOP_PREHEADER20]]
-; CHECK: loop.preheader20:
-; CHECK-NEXT: [[IV_PH:%.*]] = phi i64 [ 1, [[LOOP_PREHEADER]] ], [ [[IND_END]], [[MIDDLE_BLOCK]] ]
-; CHECK-NEXT: [[SUM_PH:%.*]] = phi i64 [ [[SUM_NEXT_PEEL]], [[LOOP_PREHEADER]] ], [ [[TMP15]], [[MIDDLE_BLOCK]] ]
+; CHECK-NEXT: [[BIN_RDX:%.*]] = add <2 x i64> [[TMP15]], [[TMP14]]
+; CHECK-NEXT: [[TMP17:%.*]] = tail call i64 @llvm.vector.reduce.add.v2i64(<2 x i64> [[BIN_RDX]])
+; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[TMP7]], [[N_VEC]]
+; CHECK-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[LOOP_PREHEADER17]]
+; CHECK: loop.preheader17:
+; CHECK-NEXT: [[IV_PH:%.*]] = phi i64 [ 0, [[LOOP_PREHEADER]] ], [ [[N_VEC]], [[MIDDLE_BLOCK]] ]
+; CHECK-NEXT: [[SUM_PH:%.*]] = phi i64 [ 0, [[LOOP_PREHEADER]] ], [ [[TMP17]], [[MIDDLE_BLOCK]] ]
; CHECK-NEXT: br label [[LOOP:%.*]]
; CHECK: loop:
-; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[IV_NEXT:%.*]], [[AT_WITH_INT_CONVERSION_EXIT11:%.*]] ], [ [[IV_PH]], [[LOOP_PREHEADER20]] ]
-; CHECK-NEXT: [[SUM:%.*]] = phi i64 [ [[SUM_NEXT:%.*]], [[AT_WITH_INT_CONVERSION_EXIT11]] ], [ [[SUM_PH]], [[LOOP_PREHEADER20]] ]
-; CHECK-NEXT: [[INRANGE_I:%.*]] = icmp ult i64 [[SUB_I]], [[IV]]
-; CHECK-NEXT: br i1 [[INRANGE_I]], label [[ERROR_I:%.*]], label [[AT_WITH_INT_CONVERSION_EXIT:%.*]]
-; CHECK: error.i:
-; CHECK-NEXT: tail call void @error()
-; CHECK-NEXT: unreachable
-; CHECK: at_with_int_conversion.exit:
-; CHECK-NEXT: [[INRANGE_I7:%.*]] = icmp ult i64 [[SUB_I6_PEEL]], [[IV]]
-; CHECK-NEXT: br i1 [[INRANGE_I7]], label [[ERROR_I10:%.*]], label [[AT_WITH_INT_CONVERSION_EXIT11]]
-; CHECK: error.i10:
-; CHECK-NEXT: tail call void @error()
-; CHECK-NEXT: unreachable
-; CHECK: at_with_int_conversion.exit11:
+; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[IV_NEXT:%.*]], [[LOOP]] ], [ [[IV_PH]], [[LOOP_PREHEADER17]] ]
+; CHECK-NEXT: [[SUM:%.*]] = phi i64 [ [[SUM_NEXT:%.*]], [[LOOP]] ], [ [[SUM_PH]], [[LOOP_PREHEADER17]] ]
; CHECK-NEXT: [[GEP_IDX_I:%.*]] = getelementptr i64, ptr [[START_I]], i64 [[IV]]
; CHECK-NEXT: [[LV_I:%.*]] = load i64, ptr [[GEP_IDX_I]], align 8
-; CHECK-NEXT: [[GEP_IDX_I8:%.*]] = getelementptr i64, ptr [[START_I1_PEEL]], i64 [[IV]]
+; CHECK-NEXT: [[GEP_IDX_I8:%.*]] = getelementptr i64, ptr [[START_I1]], i64 [[IV]]
; CHECK-NEXT: [[LV_I9:%.*]] = load i64, ptr [[GEP_IDX_I8]], align 8
; CHECK-NEXT: [[ADD:%.*]] = add i64 [[LV_I]], [[SUM]]
; CHECK-NEXT: [[SUM_NEXT]] = add i64 [[ADD]], [[LV_I9]]
; CHECK-NEXT: [[IV_NEXT]] = add nuw i64 [[IV]], 1
; CHECK-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[IV]], [[SMAX]]
-; CHECK-NEXT: br i1 [[EXITCOND_NOT]], label [[EXIT]], label [[LOOP]], !llvm.loop [[LOOP4:![0-9]+]]
+; CHECK-NEXT: br i1 [[EXITCOND_NOT]], label [[EXIT]], label [[LOOP]], !llvm.loop [[LOOP3:![0-9]+]]
+; CHECK: error.i:
+; CHECK-NEXT: tail call void @error()
+; CHECK-NEXT: unreachable
+; CHECK: error.i10:
+; CHECK-NEXT: tail call void @error()
+; CHECK-NEXT: unreachable
; CHECK: exit:
-; CHECK-NEXT: [[SUM_NEXT_LCSSA:%.*]] = phi i64 [ [[SUM_NEXT_PEEL]], [[AT_WITH_INT_CONVERSION_EXIT11_PEEL:%.*]] ], [ [[SUM_NEXT]], [[AT_WITH_INT_CONVERSION_EXIT11]] ]
+; CHECK-NEXT: [[SUM_NEXT_LCSSA:%.*]] = phi i64 [ [[TMP17]], [[MIDDLE_BLOCK]] ], [ [[SUM_NEXT]], [[LOOP]] ]
; CHECK-NEXT: ret i64 [[SUM_NEXT_LCSSA]]
;
entry:
@@ -120,120 +112,111 @@ exit:
define i64 @sum_3_at_with_int_conversion(ptr %A, ptr %B, ptr %C, i64 %N) {
; CHECK-LABEL: @sum_3_at_with_int_conversion(
-; CHECK-NEXT: at_with_int_conversion.exit22.peel:
+; CHECK-NEXT: entry:
; CHECK-NEXT: [[START_I:%.*]] = load ptr, ptr [[A:%.*]], align 8
; CHECK-NEXT: [[GEP_END_I:%.*]] = getelementptr i8, ptr [[A]], i64 8
; CHECK-NEXT: [[END_I:%.*]] = load ptr, ptr [[GEP_END_I]], align 8
; CHECK-NEXT: [[START_INT_I:%.*]] = ptrtoint ptr [[START_I]] to i64
; CHECK-NEXT: [[END_INT_I:%.*]] = ptrtoint ptr [[END_I]] to i64
; CHECK-NEXT: [[SUB_I:%.*]] = sub i64 [[END_INT_I]], [[START_INT_I]]
-; CHECK-NEXT: [[GEP_END_I13:%.*]] = getelementptr i8, ptr [[C:%.*]], i64 8
+; CHECK-NEXT: [[START_I1:%.*]] = load ptr, ptr [[B:%.*]], align 8
+; CHECK-NEXT: [[GEP_END_I2:%.*]] = getelementptr i8, ptr [[B]], i64 8
+; CHECK-NEXT: [[END_I3:%.*]] = load ptr, ptr [[GEP_END_I2]], align 8
+; CHECK-NEXT: [[START_INT_I4:%.*]] = ptrtoint ptr [[START_I1]] to i64
+; CHECK-NEXT: [[END_INT_I5:%.*]] = ptrtoint ptr [[END_I3]] to i64
+; CHECK-NEXT: [[SUB_I6:%.*]] = sub i64 [[END_INT_I5]], [[START_INT_I4]]
+; CHECK-NEXT: [[START_I12:%.*]] = load ptr, ptr [[C:%.*]], align 8
+; CHECK-NEXT: [[GEP_END_I13:%.*]] = getelementptr i8, ptr [[C]], i64 8
+; CHECK-NEXT: [[END_I14:%.*]] = load ptr, ptr [[GEP_END_I13]], align 8
+; CHECK-NEXT: [[START_INT_I15:%.*]] = ptrtoint ptr [[START_I12]] to i64
+; CHECK-NEXT: [[END_INT_I16:%.*]] = ptrtoint ptr [[END_I14]] to i64
+; CHECK-NEXT: [[SUB_I17:%.*]] = sub i64 [[END_INT_I16]], [[START_INT_I15]]
+; CHECK-NEXT: [[TMP0:%.*]] = zext i64 [[SUB_I]] to i128
+; CHECK-NEXT: [[TMP1:%.*]] = add nuw nsw i128 [[TMP0]], 1
+; CHECK-NEXT: [[TMP2:%.*]] = zext i64 [[SUB_I6]] to i128
+; CHECK-NEXT: [[TMP3:%.*]] = add nuw nsw i128 [[TMP2]], 1
+; CHECK-NEXT: [[UMIN:%.*]] = tail call i128 @llvm.umin.i128(i128 [[TMP3]], i128 [[TMP1]])
+; CHECK-NEXT: [[TMP4:%.*]] = zext i64 [[SUB_I17]] to i128
+; CHECK-NEXT: [[TMP5:%.*]] = add nuw nsw i128 [[TMP4]], 1
+; CHECK-NEXT: [[UMIN23:%.*]] = tail call i128 @llvm.umin.i128(i128 [[UMIN]], i128 [[TMP5]])
; CHECK-NEXT: [[SMAX:%.*]] = tail call i64 @llvm.smax.i64(i64 [[N:%.*]], i64 0)
-; CHECK-NEXT: [[GEP_END_I2:%.*]] = getelementptr i8, ptr [[B:%.*]], i64 8
-; CHECK-NEXT: [[LV_I_PEEL:%.*]] = load i64, ptr [[START_I]], align 8
-; CHECK-NEXT: [[START_I1_PEEL:%.*]] = load ptr, ptr [[B]], align 8
-; CHECK-NEXT: [[END_I3_PEEL:%.*]] = load ptr, ptr [[GEP_END_I2]], align 8
-; CHECK-NEXT: [[START_INT_I4_PEEL:%.*]] = ptrtoint ptr [[START_I1_PEEL]] to i64
-; CHECK-NEXT: [[END_I3_PEEL_FR:%.*]] = freeze ptr [[END_I3_PEEL]]
-; CHECK-NEXT: [[END_INT_I5_PEEL:%.*]] = ptrtoint ptr [[END_I3_PEEL_FR]] to i64
-; CHECK-NEXT: [[SUB_I6_PEEL:%.*]] = sub i64 [[END_INT_I5_PEEL]], [[START_INT_I4_PEEL]]
-; CHECK-NEXT: [[START_I12_PEEL:%.*]] = load ptr, ptr [[C]], align 8
-; CHECK-NEXT: [[END_I14_PEEL:%.*]] = load ptr, ptr [[GEP_END_I13]], align 8
-; CHECK-NEXT: [[START_INT_I15_PEEL:%.*]] = ptrtoint ptr [[START_I12_PEEL]] to i64
-; CHECK-NEXT: [[END_INT_I16_PEEL:%.*]] = ptrtoint ptr [[END_I14_PEEL]] to i64
-; CHECK-NEXT: [[SUB_I17_PEEL:%.*]] = sub i64 [[END_INT_I16_PEEL]], [[START_INT_I15_PEEL]]
-; CHECK-NEXT: [[LV_I9_PEEL:%.*]] = load i64, ptr [[START_I1_PEEL]], align 8
-; CHECK-NEXT: [[LV_I20_PEEL:%.*]] = load i64, ptr [[START_I12_PEEL]], align 8
-; CHECK-NEXT: [[ADD_2_PEEL:%.*]] = add i64 [[LV_I_PEEL]], [[LV_I9_PEEL]]
-; CHECK-NEXT: [[SUM_NEXT_PEEL:%.*]] = add i64 [[ADD_2_PEEL]], [[LV_I20_PEEL]]
-; CHECK-NEXT: [[EXITCOND_PEEL_NOT:%.*]] = icmp slt i64 [[N]], 1
-; CHECK-NEXT: br i1 [[EXITCOND_PEEL_NOT]], label [[EXIT:%.*]], label [[LOOP_PREHEADER:%.*]]
+; CHECK-NEXT: [[TMP6:%.*]] = zext nneg i64 [[SMAX]] to i128
+; CHECK-NEXT: [[UMIN24:%.*]] = tail call i128 @llvm.umin.i128(i128 [[UMIN23]], i128 [[TMP6]])
+; CHECK-NEXT: [[TMP7:%.*]] = icmp eq i128 [[TMP1]], [[UMIN24]]
+; CHECK-NEXT: [[TMP8:%.*]] = icmp eq i128 [[TMP5]], [[UMIN24]]
+; CHECK-NEXT: br i1 [[TMP7]], label [[ERROR_I:%.*]], label [[ENTRY_SPLIT:%.*]]
+; CHECK: entry.split:
+; CHECK-NEXT: [[TMP9:%.*]] = icmp eq i128 [[TMP3]], [[UMIN24]]
+; CHECK-NEXT: br i1 [[TMP9]], label [[ERROR_I10:%.*]], label [[ENTRY_SPLIT_SPLIT:%.*]]
+; CHECK: entry.split.split:
+; CHECK-NEXT: br i1 [[TMP8]], label [[ERROR_I21:%.*]], label [[LOOP_PREHEADER:%.*]]
; CHECK: loop.preheader:
-; CHECK-NEXT: [[TMP0:%.*]] = add nsw i64 [[SMAX]], -1
-; CHECK-NEXT: [[UMIN:%.*]] = tail call i64 @llvm.umin.i64(i64 [[SUB_I17_PEEL]], i64 [[TMP0]])
-; CHECK-NEXT: [[TMP1:%.*]] = freeze i64 [[UMIN]]
-; CHECK-NEXT: [[UMIN26:%.*]] = tail call i64 @llvm.umin.i64(i64 [[TMP1]], i64 [[SUB_I6_PEEL]])
-; CHECK-NEXT: [[UMIN27:%.*]] = tail call i64 @llvm.umin.i64(i64 [[UMIN26]], i64 [[SUB_I]])
-; CHECK-NEXT: [[TMP2:%.*]] = add i64 [[UMIN27]], 1
-; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP2]], 5
-; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[LOOP_PREHEADER34:%.*]], label [[VECTOR_PH:%.*]]
+; CHECK-NEXT: [[TMP10:%.*]] = add nuw i64 [[SMAX]], 1
+; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp slt i64 [[N]], 3
+; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[LOOP_PREHEADER31:%.*]], label [[VECTOR_PH:%.*]]
; CHECK: vector.ph:
-; CHECK-NEXT: [[N_MOD_VF:%.*]] = and i64 [[TMP2]], 3
-; CHECK-NEXT: [[TMP3:%.*]] = icmp eq i64 [[N_MOD_VF]], 0
-; CHECK-NEXT: [[TMP4:%.*]] = select i1 [[TMP3]], i64 4, i64 [[N_MOD_VF]]
-; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[TMP2]], [[TMP4]]
-; CHECK-NEXT: [[IND_END:%.*]] = add i64 [[N_VEC]], 1
-; CHECK-NEXT: [[TMP5:%.*]] = insertelement <2 x i64> <i64 poison, i64 0>, i64 [[SUM_NEXT_PEEL]], i64 0
+; CHECK-NEXT: [[N_VEC:%.*]] = and i64 [[TMP10]], -4
; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]
; CHECK: vector.body:
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT: [[VEC_PHI:%....
[truncated]
|
|
@llvm/pr-subscribers-llvm-transforms Author: Nikita Popov (nikic) ChangesMustExec has special logic to determine whether the first loop iteration will always be executed, by simplifying the IV comparison with the start value. Currently, this code assumes that the IV is on the LHS of the comparison, but this is not guaranteed. Make sure it handles the commuted variant as well. The changed PhaseOrdering test previously performed peeling to make the loads dereferenceable -- as a side effect, this also reduced the exit count by one, avoiding the awkward <= MAX case. Now we know up-front the the loads are dereferenceable and can be simply hoisted. As such, we retain the original exit count and now have to handle it by widening the exit count calculation to i128. This is a regression, but at least it preserves the vectorization, which was the original goal. I'm not sure what else can be done about that test. Patch is 27.98 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/93150.diff 3 Files Affected:
diff --git a/llvm/lib/Analysis/MustExecute.cpp b/llvm/lib/Analysis/MustExecute.cpp
index d4b31f2b00187..caed62679a683 100644
--- a/llvm/lib/Analysis/MustExecute.cpp
+++ b/llvm/lib/Analysis/MustExecute.cpp
@@ -135,16 +135,21 @@ static bool CanProveNotTakenFirstIteration(const BasicBlock *ExitBlock,
// todo: this would be a lot more powerful if we used scev, but all the
// plumbing is currently missing to pass a pointer in from the pass
// Check for cmp (phi [x, preheader] ...), y where (pred x, y is known
+ ICmpInst::Predicate Pred = Cond->getPredicate();
auto *LHS = dyn_cast<PHINode>(Cond->getOperand(0));
auto *RHS = Cond->getOperand(1);
- if (!LHS || LHS->getParent() != CurLoop->getHeader())
- return false;
+ if (!LHS || LHS->getParent() != CurLoop->getHeader()) {
+ Pred = Cond->getSwappedPredicate();
+ LHS = dyn_cast<PHINode>(Cond->getOperand(1));
+ RHS = Cond->getOperand(0);
+ if (!LHS || LHS->getParent() != CurLoop->getHeader())
+ return false;
+ }
+
auto DL = ExitBlock->getModule()->getDataLayout();
auto *IVStart = LHS->getIncomingValueForBlock(CurLoop->getLoopPreheader());
- auto *SimpleValOrNull = simplifyCmpInst(Cond->getPredicate(),
- IVStart, RHS,
- {DL, /*TLI*/ nullptr,
- DT, /*AC*/ nullptr, BI});
+ auto *SimpleValOrNull = simplifyCmpInst(
+ Pred, IVStart, RHS, {DL, /*TLI*/ nullptr, DT, /*AC*/ nullptr, BI});
auto *SimpleCst = dyn_cast_or_null<Constant>(SimpleValOrNull);
if (!SimpleCst)
return false;
diff --git a/llvm/test/Transforms/LICM/hoist-mustexec.ll b/llvm/test/Transforms/LICM/hoist-mustexec.ll
index 81e0815053ffe..a6f5a2be05ee4 100644
--- a/llvm/test/Transforms/LICM/hoist-mustexec.ll
+++ b/llvm/test/Transforms/LICM/hoist-mustexec.ll
@@ -218,7 +218,6 @@ fail:
}
; Same as previous case, with commuted icmp.
-; FIXME: The load should get hoisted here as well.
define i32 @test3_commuted(ptr noalias nocapture readonly %a) nounwind uwtable {
; CHECK-LABEL: define i32 @test3_commuted(
; CHECK-SAME: ptr noalias nocapture readonly [[A:%.*]]) #[[ATTR1]] {
@@ -227,6 +226,7 @@ define i32 @test3_commuted(ptr noalias nocapture readonly %a) nounwind uwtable {
; CHECK-NEXT: [[IS_ZERO:%.*]] = icmp eq i32 [[LEN]], 0
; CHECK-NEXT: br i1 [[IS_ZERO]], label [[FAIL:%.*]], label [[PREHEADER:%.*]]
; CHECK: preheader:
+; CHECK-NEXT: [[I1:%.*]] = load i32, ptr [[A]], align 4
; CHECK-NEXT: br label [[FOR_BODY:%.*]]
; CHECK: for.body:
; CHECK-NEXT: [[IV:%.*]] = phi i32 [ 0, [[PREHEADER]] ], [ [[INC:%.*]], [[CONTINUE:%.*]] ]
@@ -234,7 +234,6 @@ define i32 @test3_commuted(ptr noalias nocapture readonly %a) nounwind uwtable {
; CHECK-NEXT: [[R_CHK:%.*]] = icmp uge i32 [[LEN]], [[IV]]
; CHECK-NEXT: br i1 [[R_CHK]], label [[CONTINUE]], label [[FAIL_LOOPEXIT:%.*]]
; CHECK: continue:
-; CHECK-NEXT: [[I1:%.*]] = load i32, ptr [[A]], align 4
; CHECK-NEXT: [[ADD]] = add nsw i32 [[I1]], [[ACC]]
; CHECK-NEXT: [[INC]] = add nuw nsw i32 [[IV]], 1
; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i32 [[INC]], 1000
diff --git a/llvm/test/Transforms/PhaseOrdering/AArch64/peel-multiple-unreachable-exits-for-vectorization.ll b/llvm/test/Transforms/PhaseOrdering/AArch64/peel-multiple-unreachable-exits-for-vectorization.ll
index 8fc5189e8bc79..cc4890e27f2bd 100644
--- a/llvm/test/Transforms/PhaseOrdering/AArch64/peel-multiple-unreachable-exits-for-vectorization.ll
+++ b/llvm/test/Transforms/PhaseOrdering/AArch64/peel-multiple-unreachable-exits-for-vectorization.ll
@@ -9,95 +9,87 @@
define i64 @sum_2_at_with_int_conversion(ptr %A, ptr %B, i64 %N) {
; CHECK-LABEL: @sum_2_at_with_int_conversion(
-; CHECK-NEXT: at_with_int_conversion.exit11.peel:
+; CHECK-NEXT: entry:
; CHECK-NEXT: [[START_I:%.*]] = load ptr, ptr [[A:%.*]], align 8
; CHECK-NEXT: [[GEP_END_I:%.*]] = getelementptr i8, ptr [[A]], i64 8
; CHECK-NEXT: [[END_I:%.*]] = load ptr, ptr [[GEP_END_I]], align 8
; CHECK-NEXT: [[START_INT_I:%.*]] = ptrtoint ptr [[START_I]] to i64
; CHECK-NEXT: [[END_INT_I:%.*]] = ptrtoint ptr [[END_I]] to i64
; CHECK-NEXT: [[SUB_I:%.*]] = sub i64 [[END_INT_I]], [[START_INT_I]]
+; CHECK-NEXT: [[START_I1:%.*]] = load ptr, ptr [[B:%.*]], align 8
+; CHECK-NEXT: [[GEP_END_I2:%.*]] = getelementptr i8, ptr [[B]], i64 8
+; CHECK-NEXT: [[END_I3:%.*]] = load ptr, ptr [[GEP_END_I2]], align 8
+; CHECK-NEXT: [[START_INT_I4:%.*]] = ptrtoint ptr [[START_I1]] to i64
+; CHECK-NEXT: [[END_INT_I5:%.*]] = ptrtoint ptr [[END_I3]] to i64
+; CHECK-NEXT: [[SUB_I6:%.*]] = sub i64 [[END_INT_I5]], [[START_INT_I4]]
+; CHECK-NEXT: [[TMP0:%.*]] = zext i64 [[SUB_I]] to i128
+; CHECK-NEXT: [[TMP1:%.*]] = add nuw nsw i128 [[TMP0]], 1
+; CHECK-NEXT: [[TMP2:%.*]] = zext i64 [[SUB_I6]] to i128
+; CHECK-NEXT: [[TMP3:%.*]] = add nuw nsw i128 [[TMP2]], 1
+; CHECK-NEXT: [[UMIN:%.*]] = tail call i128 @llvm.umin.i128(i128 [[TMP3]], i128 [[TMP1]])
; CHECK-NEXT: [[SMAX:%.*]] = tail call i64 @llvm.smax.i64(i64 [[N:%.*]], i64 0)
-; CHECK-NEXT: [[GEP_END_I2:%.*]] = getelementptr i8, ptr [[B:%.*]], i64 8
-; CHECK-NEXT: [[START_I1_PEEL:%.*]] = load ptr, ptr [[B]], align 8
-; CHECK-NEXT: [[END_I3_PEEL:%.*]] = load ptr, ptr [[GEP_END_I2]], align 8
-; CHECK-NEXT: [[START_INT_I4_PEEL:%.*]] = ptrtoint ptr [[START_I1_PEEL]] to i64
-; CHECK-NEXT: [[END_INT_I5_PEEL:%.*]] = ptrtoint ptr [[END_I3_PEEL]] to i64
-; CHECK-NEXT: [[SUB_I6_PEEL:%.*]] = sub i64 [[END_INT_I5_PEEL]], [[START_INT_I4_PEEL]]
-; CHECK-NEXT: [[LV_I_PEEL:%.*]] = load i64, ptr [[START_I]], align 8
-; CHECK-NEXT: [[LV_I9_PEEL:%.*]] = load i64, ptr [[START_I1_PEEL]], align 8
-; CHECK-NEXT: [[SUM_NEXT_PEEL:%.*]] = add i64 [[LV_I_PEEL]], [[LV_I9_PEEL]]
-; CHECK-NEXT: [[EXITCOND_PEEL_NOT:%.*]] = icmp slt i64 [[N]], 1
-; CHECK-NEXT: br i1 [[EXITCOND_PEEL_NOT]], label [[EXIT:%.*]], label [[LOOP_PREHEADER:%.*]]
+; CHECK-NEXT: [[TMP4:%.*]] = zext nneg i64 [[SMAX]] to i128
+; CHECK-NEXT: [[UMIN12:%.*]] = tail call i128 @llvm.umin.i128(i128 [[UMIN]], i128 [[TMP4]])
+; CHECK-NEXT: [[TMP5:%.*]] = icmp eq i128 [[TMP1]], [[UMIN12]]
+; CHECK-NEXT: br i1 [[TMP5]], label [[ERROR_I:%.*]], label [[ENTRY_SPLIT:%.*]]
+; CHECK: entry.split:
+; CHECK-NEXT: [[TMP6:%.*]] = icmp eq i128 [[TMP3]], [[UMIN12]]
+; CHECK-NEXT: br i1 [[TMP6]], label [[ERROR_I10:%.*]], label [[LOOP_PREHEADER:%.*]]
; CHECK: loop.preheader:
-; CHECK-NEXT: [[TMP0:%.*]] = add nsw i64 [[SMAX]], -1
-; CHECK-NEXT: [[UMIN:%.*]] = tail call i64 @llvm.umin.i64(i64 [[SUB_I6_PEEL]], i64 [[TMP0]])
-; CHECK-NEXT: [[TMP1:%.*]] = freeze i64 [[UMIN]]
-; CHECK-NEXT: [[UMIN15:%.*]] = tail call i64 @llvm.umin.i64(i64 [[TMP1]], i64 [[SUB_I]])
-; CHECK-NEXT: [[TMP2:%.*]] = add i64 [[UMIN15]], 1
-; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP2]], 5
-; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[LOOP_PREHEADER20:%.*]], label [[VECTOR_PH:%.*]]
+; CHECK-NEXT: [[TMP7:%.*]] = add nuw i64 [[SMAX]], 1
+; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp slt i64 [[N]], 3
+; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[LOOP_PREHEADER17:%.*]], label [[VECTOR_PH:%.*]]
; CHECK: vector.ph:
-; CHECK-NEXT: [[N_MOD_VF:%.*]] = and i64 [[TMP2]], 3
-; CHECK-NEXT: [[TMP3:%.*]] = icmp eq i64 [[N_MOD_VF]], 0
-; CHECK-NEXT: [[TMP4:%.*]] = select i1 [[TMP3]], i64 4, i64 [[N_MOD_VF]]
-; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[TMP2]], [[TMP4]]
-; CHECK-NEXT: [[IND_END:%.*]] = add i64 [[N_VEC]], 1
-; CHECK-NEXT: [[TMP5:%.*]] = insertelement <2 x i64> <i64 poison, i64 0>, i64 [[SUM_NEXT_PEEL]], i64 0
+; CHECK-NEXT: [[N_VEC:%.*]] = and i64 [[TMP7]], -4
; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]
; CHECK: vector.body:
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT: [[VEC_PHI:%.*]] = phi <2 x i64> [ [[TMP5]], [[VECTOR_PH]] ], [ [[TMP12:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT: [[VEC_PHI16:%.*]] = phi <2 x i64> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP13:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT: [[OFFSET_IDX:%.*]] = or disjoint i64 [[INDEX]], 1
-; CHECK-NEXT: [[TMP6:%.*]] = getelementptr i64, ptr [[START_I]], i64 [[OFFSET_IDX]]
-; CHECK-NEXT: [[TMP7:%.*]] = getelementptr i8, ptr [[TMP6]], i64 16
-; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x i64>, ptr [[TMP6]], align 8
-; CHECK-NEXT: [[WIDE_LOAD17:%.*]] = load <2 x i64>, ptr [[TMP7]], align 8
-; CHECK-NEXT: [[TMP8:%.*]] = getelementptr i64, ptr [[START_I1_PEEL]], i64 [[OFFSET_IDX]]
+; CHECK-NEXT: [[VEC_PHI:%.*]] = phi <2 x i64> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP14:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT: [[VEC_PHI13:%.*]] = phi <2 x i64> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP15:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT: [[TMP8:%.*]] = getelementptr i64, ptr [[START_I]], i64 [[INDEX]]
; CHECK-NEXT: [[TMP9:%.*]] = getelementptr i8, ptr [[TMP8]], i64 16
-; CHECK-NEXT: [[WIDE_LOAD18:%.*]] = load <2 x i64>, ptr [[TMP8]], align 8
-; CHECK-NEXT: [[WIDE_LOAD19:%.*]] = load <2 x i64>, ptr [[TMP9]], align 8
-; CHECK-NEXT: [[TMP10:%.*]] = add <2 x i64> [[WIDE_LOAD]], [[VEC_PHI]]
-; CHECK-NEXT: [[TMP11:%.*]] = add <2 x i64> [[WIDE_LOAD17]], [[VEC_PHI16]]
-; CHECK-NEXT: [[TMP12]] = add <2 x i64> [[TMP10]], [[WIDE_LOAD18]]
-; CHECK-NEXT: [[TMP13]] = add <2 x i64> [[TMP11]], [[WIDE_LOAD19]]
+; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x i64>, ptr [[TMP8]], align 8
+; CHECK-NEXT: [[WIDE_LOAD14:%.*]] = load <2 x i64>, ptr [[TMP9]], align 8
+; CHECK-NEXT: [[TMP10:%.*]] = getelementptr i64, ptr [[START_I1]], i64 [[INDEX]]
+; CHECK-NEXT: [[TMP11:%.*]] = getelementptr i8, ptr [[TMP10]], i64 16
+; CHECK-NEXT: [[WIDE_LOAD15:%.*]] = load <2 x i64>, ptr [[TMP10]], align 8
+; CHECK-NEXT: [[WIDE_LOAD16:%.*]] = load <2 x i64>, ptr [[TMP11]], align 8
+; CHECK-NEXT: [[TMP12:%.*]] = add <2 x i64> [[WIDE_LOAD]], [[VEC_PHI]]
+; CHECK-NEXT: [[TMP13:%.*]] = add <2 x i64> [[WIDE_LOAD14]], [[VEC_PHI13]]
+; CHECK-NEXT: [[TMP14]] = add <2 x i64> [[TMP12]], [[WIDE_LOAD15]]
+; CHECK-NEXT: [[TMP15]] = add <2 x i64> [[TMP13]], [[WIDE_LOAD16]]
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
-; CHECK-NEXT: [[TMP14:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
-; CHECK-NEXT: br i1 [[TMP14]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; CHECK-NEXT: [[TMP16:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-NEXT: br i1 [[TMP16]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
; CHECK: middle.block:
-; CHECK-NEXT: [[BIN_RDX:%.*]] = add <2 x i64> [[TMP13]], [[TMP12]]
-; CHECK-NEXT: [[TMP15:%.*]] = tail call i64 @llvm.vector.reduce.add.v2i64(<2 x i64> [[BIN_RDX]])
-; CHECK-NEXT: br label [[LOOP_PREHEADER20]]
-; CHECK: loop.preheader20:
-; CHECK-NEXT: [[IV_PH:%.*]] = phi i64 [ 1, [[LOOP_PREHEADER]] ], [ [[IND_END]], [[MIDDLE_BLOCK]] ]
-; CHECK-NEXT: [[SUM_PH:%.*]] = phi i64 [ [[SUM_NEXT_PEEL]], [[LOOP_PREHEADER]] ], [ [[TMP15]], [[MIDDLE_BLOCK]] ]
+; CHECK-NEXT: [[BIN_RDX:%.*]] = add <2 x i64> [[TMP15]], [[TMP14]]
+; CHECK-NEXT: [[TMP17:%.*]] = tail call i64 @llvm.vector.reduce.add.v2i64(<2 x i64> [[BIN_RDX]])
+; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[TMP7]], [[N_VEC]]
+; CHECK-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[LOOP_PREHEADER17]]
+; CHECK: loop.preheader17:
+; CHECK-NEXT: [[IV_PH:%.*]] = phi i64 [ 0, [[LOOP_PREHEADER]] ], [ [[N_VEC]], [[MIDDLE_BLOCK]] ]
+; CHECK-NEXT: [[SUM_PH:%.*]] = phi i64 [ 0, [[LOOP_PREHEADER]] ], [ [[TMP17]], [[MIDDLE_BLOCK]] ]
; CHECK-NEXT: br label [[LOOP:%.*]]
; CHECK: loop:
-; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[IV_NEXT:%.*]], [[AT_WITH_INT_CONVERSION_EXIT11:%.*]] ], [ [[IV_PH]], [[LOOP_PREHEADER20]] ]
-; CHECK-NEXT: [[SUM:%.*]] = phi i64 [ [[SUM_NEXT:%.*]], [[AT_WITH_INT_CONVERSION_EXIT11]] ], [ [[SUM_PH]], [[LOOP_PREHEADER20]] ]
-; CHECK-NEXT: [[INRANGE_I:%.*]] = icmp ult i64 [[SUB_I]], [[IV]]
-; CHECK-NEXT: br i1 [[INRANGE_I]], label [[ERROR_I:%.*]], label [[AT_WITH_INT_CONVERSION_EXIT:%.*]]
-; CHECK: error.i:
-; CHECK-NEXT: tail call void @error()
-; CHECK-NEXT: unreachable
-; CHECK: at_with_int_conversion.exit:
-; CHECK-NEXT: [[INRANGE_I7:%.*]] = icmp ult i64 [[SUB_I6_PEEL]], [[IV]]
-; CHECK-NEXT: br i1 [[INRANGE_I7]], label [[ERROR_I10:%.*]], label [[AT_WITH_INT_CONVERSION_EXIT11]]
-; CHECK: error.i10:
-; CHECK-NEXT: tail call void @error()
-; CHECK-NEXT: unreachable
-; CHECK: at_with_int_conversion.exit11:
+; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[IV_NEXT:%.*]], [[LOOP]] ], [ [[IV_PH]], [[LOOP_PREHEADER17]] ]
+; CHECK-NEXT: [[SUM:%.*]] = phi i64 [ [[SUM_NEXT:%.*]], [[LOOP]] ], [ [[SUM_PH]], [[LOOP_PREHEADER17]] ]
; CHECK-NEXT: [[GEP_IDX_I:%.*]] = getelementptr i64, ptr [[START_I]], i64 [[IV]]
; CHECK-NEXT: [[LV_I:%.*]] = load i64, ptr [[GEP_IDX_I]], align 8
-; CHECK-NEXT: [[GEP_IDX_I8:%.*]] = getelementptr i64, ptr [[START_I1_PEEL]], i64 [[IV]]
+; CHECK-NEXT: [[GEP_IDX_I8:%.*]] = getelementptr i64, ptr [[START_I1]], i64 [[IV]]
; CHECK-NEXT: [[LV_I9:%.*]] = load i64, ptr [[GEP_IDX_I8]], align 8
; CHECK-NEXT: [[ADD:%.*]] = add i64 [[LV_I]], [[SUM]]
; CHECK-NEXT: [[SUM_NEXT]] = add i64 [[ADD]], [[LV_I9]]
; CHECK-NEXT: [[IV_NEXT]] = add nuw i64 [[IV]], 1
; CHECK-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[IV]], [[SMAX]]
-; CHECK-NEXT: br i1 [[EXITCOND_NOT]], label [[EXIT]], label [[LOOP]], !llvm.loop [[LOOP4:![0-9]+]]
+; CHECK-NEXT: br i1 [[EXITCOND_NOT]], label [[EXIT]], label [[LOOP]], !llvm.loop [[LOOP3:![0-9]+]]
+; CHECK: error.i:
+; CHECK-NEXT: tail call void @error()
+; CHECK-NEXT: unreachable
+; CHECK: error.i10:
+; CHECK-NEXT: tail call void @error()
+; CHECK-NEXT: unreachable
; CHECK: exit:
-; CHECK-NEXT: [[SUM_NEXT_LCSSA:%.*]] = phi i64 [ [[SUM_NEXT_PEEL]], [[AT_WITH_INT_CONVERSION_EXIT11_PEEL:%.*]] ], [ [[SUM_NEXT]], [[AT_WITH_INT_CONVERSION_EXIT11]] ]
+; CHECK-NEXT: [[SUM_NEXT_LCSSA:%.*]] = phi i64 [ [[TMP17]], [[MIDDLE_BLOCK]] ], [ [[SUM_NEXT]], [[LOOP]] ]
; CHECK-NEXT: ret i64 [[SUM_NEXT_LCSSA]]
;
entry:
@@ -120,120 +112,111 @@ exit:
define i64 @sum_3_at_with_int_conversion(ptr %A, ptr %B, ptr %C, i64 %N) {
; CHECK-LABEL: @sum_3_at_with_int_conversion(
-; CHECK-NEXT: at_with_int_conversion.exit22.peel:
+; CHECK-NEXT: entry:
; CHECK-NEXT: [[START_I:%.*]] = load ptr, ptr [[A:%.*]], align 8
; CHECK-NEXT: [[GEP_END_I:%.*]] = getelementptr i8, ptr [[A]], i64 8
; CHECK-NEXT: [[END_I:%.*]] = load ptr, ptr [[GEP_END_I]], align 8
; CHECK-NEXT: [[START_INT_I:%.*]] = ptrtoint ptr [[START_I]] to i64
; CHECK-NEXT: [[END_INT_I:%.*]] = ptrtoint ptr [[END_I]] to i64
; CHECK-NEXT: [[SUB_I:%.*]] = sub i64 [[END_INT_I]], [[START_INT_I]]
-; CHECK-NEXT: [[GEP_END_I13:%.*]] = getelementptr i8, ptr [[C:%.*]], i64 8
+; CHECK-NEXT: [[START_I1:%.*]] = load ptr, ptr [[B:%.*]], align 8
+; CHECK-NEXT: [[GEP_END_I2:%.*]] = getelementptr i8, ptr [[B]], i64 8
+; CHECK-NEXT: [[END_I3:%.*]] = load ptr, ptr [[GEP_END_I2]], align 8
+; CHECK-NEXT: [[START_INT_I4:%.*]] = ptrtoint ptr [[START_I1]] to i64
+; CHECK-NEXT: [[END_INT_I5:%.*]] = ptrtoint ptr [[END_I3]] to i64
+; CHECK-NEXT: [[SUB_I6:%.*]] = sub i64 [[END_INT_I5]], [[START_INT_I4]]
+; CHECK-NEXT: [[START_I12:%.*]] = load ptr, ptr [[C:%.*]], align 8
+; CHECK-NEXT: [[GEP_END_I13:%.*]] = getelementptr i8, ptr [[C]], i64 8
+; CHECK-NEXT: [[END_I14:%.*]] = load ptr, ptr [[GEP_END_I13]], align 8
+; CHECK-NEXT: [[START_INT_I15:%.*]] = ptrtoint ptr [[START_I12]] to i64
+; CHECK-NEXT: [[END_INT_I16:%.*]] = ptrtoint ptr [[END_I14]] to i64
+; CHECK-NEXT: [[SUB_I17:%.*]] = sub i64 [[END_INT_I16]], [[START_INT_I15]]
+; CHECK-NEXT: [[TMP0:%.*]] = zext i64 [[SUB_I]] to i128
+; CHECK-NEXT: [[TMP1:%.*]] = add nuw nsw i128 [[TMP0]], 1
+; CHECK-NEXT: [[TMP2:%.*]] = zext i64 [[SUB_I6]] to i128
+; CHECK-NEXT: [[TMP3:%.*]] = add nuw nsw i128 [[TMP2]], 1
+; CHECK-NEXT: [[UMIN:%.*]] = tail call i128 @llvm.umin.i128(i128 [[TMP3]], i128 [[TMP1]])
+; CHECK-NEXT: [[TMP4:%.*]] = zext i64 [[SUB_I17]] to i128
+; CHECK-NEXT: [[TMP5:%.*]] = add nuw nsw i128 [[TMP4]], 1
+; CHECK-NEXT: [[UMIN23:%.*]] = tail call i128 @llvm.umin.i128(i128 [[UMIN]], i128 [[TMP5]])
; CHECK-NEXT: [[SMAX:%.*]] = tail call i64 @llvm.smax.i64(i64 [[N:%.*]], i64 0)
-; CHECK-NEXT: [[GEP_END_I2:%.*]] = getelementptr i8, ptr [[B:%.*]], i64 8
-; CHECK-NEXT: [[LV_I_PEEL:%.*]] = load i64, ptr [[START_I]], align 8
-; CHECK-NEXT: [[START_I1_PEEL:%.*]] = load ptr, ptr [[B]], align 8
-; CHECK-NEXT: [[END_I3_PEEL:%.*]] = load ptr, ptr [[GEP_END_I2]], align 8
-; CHECK-NEXT: [[START_INT_I4_PEEL:%.*]] = ptrtoint ptr [[START_I1_PEEL]] to i64
-; CHECK-NEXT: [[END_I3_PEEL_FR:%.*]] = freeze ptr [[END_I3_PEEL]]
-; CHECK-NEXT: [[END_INT_I5_PEEL:%.*]] = ptrtoint ptr [[END_I3_PEEL_FR]] to i64
-; CHECK-NEXT: [[SUB_I6_PEEL:%.*]] = sub i64 [[END_INT_I5_PEEL]], [[START_INT_I4_PEEL]]
-; CHECK-NEXT: [[START_I12_PEEL:%.*]] = load ptr, ptr [[C]], align 8
-; CHECK-NEXT: [[END_I14_PEEL:%.*]] = load ptr, ptr [[GEP_END_I13]], align 8
-; CHECK-NEXT: [[START_INT_I15_PEEL:%.*]] = ptrtoint ptr [[START_I12_PEEL]] to i64
-; CHECK-NEXT: [[END_INT_I16_PEEL:%.*]] = ptrtoint ptr [[END_I14_PEEL]] to i64
-; CHECK-NEXT: [[SUB_I17_PEEL:%.*]] = sub i64 [[END_INT_I16_PEEL]], [[START_INT_I15_PEEL]]
-; CHECK-NEXT: [[LV_I9_PEEL:%.*]] = load i64, ptr [[START_I1_PEEL]], align 8
-; CHECK-NEXT: [[LV_I20_PEEL:%.*]] = load i64, ptr [[START_I12_PEEL]], align 8
-; CHECK-NEXT: [[ADD_2_PEEL:%.*]] = add i64 [[LV_I_PEEL]], [[LV_I9_PEEL]]
-; CHECK-NEXT: [[SUM_NEXT_PEEL:%.*]] = add i64 [[ADD_2_PEEL]], [[LV_I20_PEEL]]
-; CHECK-NEXT: [[EXITCOND_PEEL_NOT:%.*]] = icmp slt i64 [[N]], 1
-; CHECK-NEXT: br i1 [[EXITCOND_PEEL_NOT]], label [[EXIT:%.*]], label [[LOOP_PREHEADER:%.*]]
+; CHECK-NEXT: [[TMP6:%.*]] = zext nneg i64 [[SMAX]] to i128
+; CHECK-NEXT: [[UMIN24:%.*]] = tail call i128 @llvm.umin.i128(i128 [[UMIN23]], i128 [[TMP6]])
+; CHECK-NEXT: [[TMP7:%.*]] = icmp eq i128 [[TMP1]], [[UMIN24]]
+; CHECK-NEXT: [[TMP8:%.*]] = icmp eq i128 [[TMP5]], [[UMIN24]]
+; CHECK-NEXT: br i1 [[TMP7]], label [[ERROR_I:%.*]], label [[ENTRY_SPLIT:%.*]]
+; CHECK: entry.split:
+; CHECK-NEXT: [[TMP9:%.*]] = icmp eq i128 [[TMP3]], [[UMIN24]]
+; CHECK-NEXT: br i1 [[TMP9]], label [[ERROR_I10:%.*]], label [[ENTRY_SPLIT_SPLIT:%.*]]
+; CHECK: entry.split.split:
+; CHECK-NEXT: br i1 [[TMP8]], label [[ERROR_I21:%.*]], label [[LOOP_PREHEADER:%.*]]
; CHECK: loop.preheader:
-; CHECK-NEXT: [[TMP0:%.*]] = add nsw i64 [[SMAX]], -1
-; CHECK-NEXT: [[UMIN:%.*]] = tail call i64 @llvm.umin.i64(i64 [[SUB_I17_PEEL]], i64 [[TMP0]])
-; CHECK-NEXT: [[TMP1:%.*]] = freeze i64 [[UMIN]]
-; CHECK-NEXT: [[UMIN26:%.*]] = tail call i64 @llvm.umin.i64(i64 [[TMP1]], i64 [[SUB_I6_PEEL]])
-; CHECK-NEXT: [[UMIN27:%.*]] = tail call i64 @llvm.umin.i64(i64 [[UMIN26]], i64 [[SUB_I]])
-; CHECK-NEXT: [[TMP2:%.*]] = add i64 [[UMIN27]], 1
-; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP2]], 5
-; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[LOOP_PREHEADER34:%.*]], label [[VECTOR_PH:%.*]]
+; CHECK-NEXT: [[TMP10:%.*]] = add nuw i64 [[SMAX]], 1
+; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp slt i64 [[N]], 3
+; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[LOOP_PREHEADER31:%.*]], label [[VECTOR_PH:%.*]]
; CHECK: vector.ph:
-; CHECK-NEXT: [[N_MOD_VF:%.*]] = and i64 [[TMP2]], 3
-; CHECK-NEXT: [[TMP3:%.*]] = icmp eq i64 [[N_MOD_VF]], 0
-; CHECK-NEXT: [[TMP4:%.*]] = select i1 [[TMP3]], i64 4, i64 [[N_MOD_VF]]
-; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[TMP2]], [[TMP4]]
-; CHECK-NEXT: [[IND_END:%.*]] = add i64 [[N_VEC]], 1
-; CHECK-NEXT: [[TMP5:%.*]] = insertelement <2 x i64> <i64 poison, i64 0>, i64 [[SUM_NEXT_PEEL]], i64 0
+; CHECK-NEXT: [[N_VEC:%.*]] = and i64 [[TMP10]], -4
; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]
; CHECK: vector.body:
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT: [[VEC_PHI:%....
[truncated]
|
Can we truncate the count? I mean, the math in howManyLessThans doesn't fit into i64, but does the actual backedge-taken count overflow 64 bits? |
efriedma-quic
approved these changes
May 23, 2024
I think so. The BECount for |
|
Suppose a loop exits via a given exit, and that loop exit is guarded by a comparison of a loop-invariant value against an affine AddRec, and the comparison has bitwidth iN. Since the AddRec has width iN, it can take on at most 2^N possible values. Under these conditions, the AddRec can't repeat values, and there must be at least one value of the AddRec that causes the loop to exit. Therefore, there are at most (2^N)-1 values which don't cause the loop to exit. So the exit-not-taken count is at most (2^N)-1, which fits into an iN. Arguably not worth implementing, though. |
|
@efriedma-quic I think "Suppose a loop exits via a given exit" is the critical bit here. I don't think we can make this assumption when computing exit counts. The contract for the exit count is not just that "if this exit is taken, it will be after N backedges taken" but also "the backedge will not be taken more than N times before the loop exits through some exit". As such, if you have a |
|
We can still use the information, I think? In the simplest case, we have a loop with a single exit, that we can prove is finite some other way (mustprogress etc.). Then we can just report N+1, because the overflow case would be undefined. The given testcase is a bit more complicated: you have two exits. But still, we know the loop is finite. The backedge-taken count is something like Even if we can't prove the loop is finite, recording the information could be useful... see #83052 (review) . |
We already do this (the "exit controls finite loop" case).
Yes, this formulation using |
We can use umax(N, N+1) as the count as long as the AddRec is nowrap, I think? And if the AddRec isn't nowrap, we can add that as a predicate. That's basically the same reasoning we're using to return zext(N)+1 anyway. |
I don't think this is the case. Consider two exits, first |
|
Oh, I see what you mean: the overall backedge-taken count fits into an iN, but the backedge-taken count for the specific edge doesn't fit. That's unfortunate; I can't think of any way to solve it. I suspect it's still slightly profitable to try to truncate the overall backedge-taken count to iN, but maybe not by as much as I was thinking. |
Yes, exactly!
I just realized that the specific IR in the test case is being introduced by the "exit predication" transform in IndVars, which currently doesn't have a high cost expansion guard. I think if that is added, then we avoid that transform and get something more similar to the previous IR in LoopVectorize, where we only have to expand the BECount, not also the exit counts. And at that point having a narrower BECount would probably become worthwhile. |
Hm, this is trickier than I thought because I do think we want to allow some high cost expansions for that transform -- as it can enable branches to be unswitched out of the loop, doing it can make sense even if trip counts are somewhat expensive. Even in the case here, it may have been a profitable transform if LoopVectorize didn't have to the option of picking an even better alternative. I'm not sure I really want to follow this rabbit hole all the way to the end. @fhahn As you added the test, are you fine with this landing as-is? |
|
Another ping for @fhahn |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
MustExec has special logic to determine whether the first loop iteration will always be executed, by simplifying the IV comparison with the start value. Currently, this code assumes that the IV is on the LHS of the comparison, but this is not guaranteed. Make sure it handles the commuted variant as well.
The changed PhaseOrdering test previously performed peeling to make the loads dereferenceable -- as a side effect, this also reduced the exit count by one, avoiding the awkward <= MAX case.
Now we know up-front the the loads are dereferenceable and can be simply hoisted. As such, we retain the original exit count and now have to handle it by widening the exit count calculation to i128. This is a regression, but at least it preserves the vectorization, which was the original goal. I'm not sure what else can be done about that test.