[llvm] cb64c3c - [LoopInterchange] Precommit tests for scalar dependencies. NFC. (#115900)

via llvm-commits llvm-commits at lists.llvm.org
Thu Nov 14 01:13:13 PST 2024


Author: Sjoerd Meijer
Date: 2024-11-14T09:13:10Z
New Revision: cb64c3c573d7239036d46addb3ea09f954ca3a55

URL: https://github.com/llvm/llvm-project/commit/cb64c3c573d7239036d46addb3ea09f954ca3a55
DIFF: https://github.com/llvm/llvm-project/commit/cb64c3c573d7239036d46addb3ea09f954ca3a55.diff

LOG: [LoopInterchange] Precommit tests for scalar dependencies. NFC. (#115900)

We are miscompiling and incorrectly interchanging loops with scalar
dependencies that are live-out and conditionally set. This precommits
some tests demonstrating this. This is based on the tests in
https://reviews.llvm.org/D87879 by `mdchen`.

Added: 
    llvm/test/Transforms/LoopInterchange/gh54176-scalar-deps.ll

Modified: 
    

Removed: 
    


################################################################################
diff  --git a/llvm/test/Transforms/LoopInterchange/gh54176-scalar-deps.ll b/llvm/test/Transforms/LoopInterchange/gh54176-scalar-deps.ll
new file mode 100644
index 00000000000000..b3383655668980
--- /dev/null
+++ b/llvm/test/Transforms/LoopInterchange/gh54176-scalar-deps.ll
@@ -0,0 +1,225 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt < %s -passes=loop-interchange -S | FileCheck %s
+
+ at f = dso_local local_unnamed_addr global [4 x [9 x i32]] [[9 x i32] [i32 5, i32 3, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0], [9 x i32] zeroinitializer, [9 x i32] zeroinitializer, [9 x i32] zeroinitializer], align 4
+ at g = common dso_local local_unnamed_addr global i32 0, align 4
+
+; int32_t f[3][3];
+; int32_t g;
+;
+; int32_t test1(_Bool cond) {
+;     for (int64_t i = 0; i < 3; i++) {
+;         for (int64_t j = 0; j < 3; j++) {
+;            int32_t val = f[i][j];
+;            if (val == 0)
+;              if (!cond)
+;                g++;
+;            else
+;              g = 3;
+;              if (!cond)
+;                g++;
+;         }
+;     }
+;     return g;
+; }
+;
+; FIXME: if there's an output dependency inside the loop and Src doesn't
+; dominate Dst, we should not interchange. Thus, this currently miscompiles.
+;
+define dso_local i32 @test1(i1 %cond) {
+; CHECK-LABEL: define dso_local i32 @test1(
+; CHECK-SAME: i1 [[COND:%.*]]) {
+; CHECK-NEXT:  [[FOR_PREHEADER:.*:]]
+; CHECK-NEXT:    br label %[[INNERLOOP_PREHEADER:.*]]
+; CHECK:       [[OUTERLOOP_PREHEADER:.*]]:
+; CHECK-NEXT:    br label %[[OUTERLOOP:.*]]
+; CHECK:       [[OUTERLOOP]]:
+; CHECK-NEXT:    [[I:%.*]] = phi i64 [ [[INDVARS_IV_NEXT21_I:%.*]], %[[FOR_LATCH:.*]] ], [ 0, %[[OUTERLOOP_PREHEADER]] ]
+; CHECK-NEXT:    br label %[[INNERLOOP_SPLIT:.*]]
+; CHECK:       [[INNERLOOP_PREHEADER]]:
+; CHECK-NEXT:    br label %[[INNERLOOP:.*]]
+; CHECK:       [[INNERLOOP]]:
+; CHECK-NEXT:    [[J:%.*]] = phi i64 [ [[TMP0:%.*]], %[[IF_END_SPLIT:.*]] ], [ 0, %[[INNERLOOP_PREHEADER]] ]
+; CHECK-NEXT:    br label %[[OUTERLOOP_PREHEADER]]
+; CHECK:       [[INNERLOOP_SPLIT]]:
+; CHECK-NEXT:    [[ARRAYIDX6_I:%.*]] = getelementptr inbounds [4 x [9 x i32]], ptr @f, i64 0, i64 [[J]], i64 [[I]]
+; CHECK-NEXT:    [[I1:%.*]] = load i32, ptr [[ARRAYIDX6_I]], align 4
+; CHECK-NEXT:    [[TOBOOL_I:%.*]] = icmp eq i32 [[I1]], 0
+; CHECK-NEXT:    br i1 [[TOBOOL_I]], label %[[LAND_END:.*]], label %[[LAND_RHS:.*]]
+; CHECK:       [[LAND_RHS]]:
+; CHECK-NEXT:    store i32 3, ptr @g, align 4
+; CHECK-NEXT:    br label %[[LAND_END]]
+; CHECK:       [[LAND_END]]:
+; CHECK-NEXT:    br i1 [[COND]], label %[[IF_END:.*]], label %[[IF_THEN:.*]]
+; CHECK:       [[IF_THEN]]:
+; CHECK-NEXT:    [[I2:%.*]] = load i32, ptr @g, align 4
+; CHECK-NEXT:    [[INC_I:%.*]] = add i32 [[I2]], 1
+; CHECK-NEXT:    store i32 [[INC_I]], ptr @g, align 4
+; CHECK-NEXT:    br label %[[IF_END]]
+; CHECK:       [[IF_END]]:
+; CHECK-NEXT:    [[J_NEXT:%.*]] = add nuw nsw i64 [[J]], 1
+; CHECK-NEXT:    [[EXITCOND_I:%.*]] = icmp eq i64 [[J_NEXT]], 3
+; CHECK-NEXT:    br label %[[FOR_LATCH]]
+; CHECK:       [[IF_END_SPLIT]]:
+; CHECK-NEXT:    [[TMP0]] = add nuw nsw i64 [[J]], 1
+; CHECK-NEXT:    [[TMP1:%.*]] = icmp eq i64 [[TMP0]], 3
+; CHECK-NEXT:    br i1 [[TMP1]], label %[[EXIT:.*]], label %[[INNERLOOP]]
+; CHECK:       [[FOR_LATCH]]:
+; CHECK-NEXT:    [[INDVARS_IV_NEXT21_I]] = add nsw i64 [[I]], 1
+; CHECK-NEXT:    [[CMP_I:%.*]] = icmp slt i64 [[I]], 2
+; CHECK-NEXT:    br i1 [[CMP_I]], label %[[OUTERLOOP]], label %[[IF_END_SPLIT]]
+; CHECK:       [[EXIT]]:
+; CHECK-NEXT:    [[I3:%.*]] = load i32, ptr @g, align 4
+; CHECK-NEXT:    ret i32 [[I3]]
+;
+for.preheader:
+  br label %outerloop
+
+outerloop:
+  %i = phi i64 [ 0, %for.preheader ], [ %indvars.iv.next21.i, %for.latch ]
+  br label %innerloop
+
+innerloop:
+  %j = phi i64 [ 0, %outerloop ], [ %j.next, %if.end ]
+  %arrayidx6.i = getelementptr inbounds [4 x [9 x i32]], [4 x [9 x i32]]* @f, i64 0, i64 %j, i64 %i
+  %i1 = load i32, i32* %arrayidx6.i, align 4
+  %tobool.i = icmp eq i32 %i1, 0
+  br i1 %tobool.i, label %land.end, label %land.rhs
+
+land.rhs:
+  store i32 3, i32* @g, align 4
+  br label %land.end
+
+land.end:
+  br i1 %cond, label %if.end, label %if.then
+
+if.then:
+  %i2 = load i32, i32* @g, align 4
+  %inc.i = add i32 %i2, 1
+  store i32 %inc.i, i32* @g, align 4
+  br label %if.end
+
+if.end:
+  %j.next = add nuw nsw i64 %j, 1
+  %exitcond.i = icmp eq i64 %j.next, 3
+  br i1 %exitcond.i, label %for.latch, label %innerloop
+
+for.latch:
+  %indvars.iv.next21.i = add nsw i64 %i, 1
+  %cmp.i = icmp slt i64 %i, 2
+  br i1 %cmp.i, label %outerloop, label %exit
+
+exit:
+  %i3 = load i32, i32* @g, align 4
+  ret i32 %i3
+}
+
+; int32_t f[3][3];
+; int32_t g;
+;
+; int32_t test2(_Bool cond) {
+;     for (int64_t i = 0; i < 3; i++) {
+;       for (int64_t j = 0; j < 3; j++) {
+;          int32_t val = f[i][j];
+;          g = 3;
+;          if (val == 0)
+;            if (!cond)
+;              g++;
+;          else
+;            if (!cond)
+;              g++;
+;         }
+;     }
+;     return g;
+; }
+;
+; FIXME: if there's an output dependency inside the loop and Src doesn't
+; dominate Dst, we should not interchange. Thus, this currently miscompiles.
+;
+define dso_local i32 @test2(i1 %cond) {
+; CHECK-LABEL: define dso_local i32 @test2(
+; CHECK-SAME: i1 [[COND:%.*]]) {
+; CHECK-NEXT:  [[FOR_PREHEADER:.*:]]
+; CHECK-NEXT:    br label %[[INNERLOOP_PREHEADER:.*]]
+; CHECK:       [[OUTERLOOP_PREHEADER:.*]]:
+; CHECK-NEXT:    br label %[[OUTERLOOP:.*]]
+; CHECK:       [[OUTERLOOP]]:
+; CHECK-NEXT:    [[I:%.*]] = phi i64 [ [[INDVARS_IV_NEXT21_I:%.*]], %[[FOR_LATCH:.*]] ], [ 0, %[[OUTERLOOP_PREHEADER]] ]
+; CHECK-NEXT:    br label %[[INNERLOOP_SPLIT:.*]]
+; CHECK:       [[INNERLOOP_PREHEADER]]:
+; CHECK-NEXT:    br label %[[INNERLOOP:.*]]
+; CHECK:       [[INNERLOOP]]:
+; CHECK-NEXT:    [[J:%.*]] = phi i64 [ [[TMP0:%.*]], %[[IF_END_SPLIT:.*]] ], [ 0, %[[INNERLOOP_PREHEADER]] ]
+; CHECK-NEXT:    br label %[[OUTERLOOP_PREHEADER]]
+; CHECK:       [[INNERLOOP_SPLIT]]:
+; CHECK-NEXT:    [[ARRAYIDX6_I:%.*]] = getelementptr inbounds [4 x [9 x i32]], ptr @f, i64 0, i64 [[J]], i64 [[I]]
+; CHECK-NEXT:    [[I1:%.*]] = load i32, ptr [[ARRAYIDX6_I]], align 4
+; CHECK-NEXT:    [[TOBOOL_I:%.*]] = icmp eq i32 [[I1]], 0
+; CHECK-NEXT:    store i32 3, ptr @g, align 4
+; CHECK-NEXT:    br i1 [[TOBOOL_I]], label %[[LAND_END:.*]], label %[[LAND_RHS:.*]]
+; CHECK:       [[LAND_RHS]]:
+; CHECK-NEXT:    br label %[[LAND_END]]
+; CHECK:       [[LAND_END]]:
+; CHECK-NEXT:    br i1 [[COND]], label %[[IF_END:.*]], label %[[IF_THEN:.*]]
+; CHECK:       [[IF_THEN]]:
+; CHECK-NEXT:    [[I2:%.*]] = load i32, ptr @g, align 4
+; CHECK-NEXT:    [[INC_I:%.*]] = add i32 [[I2]], 1
+; CHECK-NEXT:    store i32 [[INC_I]], ptr @g, align 4
+; CHECK-NEXT:    br label %[[IF_END]]
+; CHECK:       [[IF_END]]:
+; CHECK-NEXT:    [[J_NEXT:%.*]] = add nuw nsw i64 [[J]], 1
+; CHECK-NEXT:    [[EXITCOND_I:%.*]] = icmp eq i64 [[J_NEXT]], 3
+; CHECK-NEXT:    br label %[[FOR_LATCH]]
+; CHECK:       [[IF_END_SPLIT]]:
+; CHECK-NEXT:    [[TMP0]] = add nuw nsw i64 [[J]], 1
+; CHECK-NEXT:    [[TMP1:%.*]] = icmp eq i64 [[TMP0]], 3
+; CHECK-NEXT:    br i1 [[TMP1]], label %[[EXIT:.*]], label %[[INNERLOOP]]
+; CHECK:       [[FOR_LATCH]]:
+; CHECK-NEXT:    [[INDVARS_IV_NEXT21_I]] = add nsw i64 [[I]], 1
+; CHECK-NEXT:    [[CMP_I:%.*]] = icmp slt i64 [[I]], 2
+; CHECK-NEXT:    br i1 [[CMP_I]], label %[[OUTERLOOP]], label %[[IF_END_SPLIT]]
+; CHECK:       [[EXIT]]:
+; CHECK-NEXT:    [[I3:%.*]] = load i32, ptr @g, align 4
+; CHECK-NEXT:    ret i32 [[I3]]
+;
+for.preheader:
+  br label %outerloop
+
+outerloop:
+  %i = phi i64 [ 0, %for.preheader ], [ %indvars.iv.next21.i, %for.latch ]
+  br label %innerloop
+
+innerloop:
+  %j = phi i64 [ 0, %outerloop ], [ %j.next, %if.end ]
+  %arrayidx6.i = getelementptr inbounds [4 x [9 x i32]], [4 x [9 x i32]]* @f, i64 0, i64 %j, i64 %i
+  %i1 = load i32, i32* %arrayidx6.i, align 4
+  %tobool.i = icmp eq i32 %i1, 0
+  store i32 3, i32* @g, align 4
+  br i1 %tobool.i, label %land.end, label %land.rhs
+
+land.rhs:
+  br label %land.end
+
+land.end:
+  br i1 %cond, label %if.end, label %if.then
+
+if.then:
+  %i2 = load i32, i32* @g, align 4
+  %inc.i = add i32 %i2, 1
+  store i32 %inc.i, i32* @g, align 4
+  br label %if.end
+
+if.end:
+  %j.next = add nuw nsw i64 %j, 1
+  %exitcond.i = icmp eq i64 %j.next, 3
+  br i1 %exitcond.i, label %for.latch, label %innerloop
+
+for.latch:
+  %indvars.iv.next21.i = add nsw i64 %i, 1
+  %cmp.i = icmp slt i64 %i, 2
+  br i1 %cmp.i, label %outerloop, label %exit
+
+exit:
+  %i3 = load i32, i32* @g, align 4
+  ret i32 %i3
+}


        


More information about the llvm-commits mailing list