[llvm] [AMDGPU][SplitKit] Skip any uses that are placed before the current block (PR #177270)

[Dmitry Sidorov via llvm-commits]

================
@@ -0,0 +1,86 @@
+; REQUIRES: asserts
+;
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1201 -O1 < %s -o /dev/null
+;
+; Test that SplitAnalysis doesn't crash when UseSlots contains uses in blocks that
+; are not covered by the live interval.
+
+%struct.S1 = type { i8, %struct.S0 }
+%struct.S0 = type { i64 }
+
+define i16 @backsmith_pure_2(ptr addrspace(5) %BS_VAR_0, i1 %tobool.not, i64 %add919, <32 x i8> %shuffle57, <32 x i8> %vecinit827, <16 x i16> %0) {
+entry:
+  %BS_VAR_01 = alloca [8 x <32 x i8>], align 32, addrspace(5)
+  %call2 = call i8 @backsmith_snippet_44()
+  br i1 %tobool.not, label %if.end, label %if.then, !prof !0
+
+if.then:                                          ; preds = %entry
+  br i1 %tobool.not, label %for.cond.cleanup, label %for.body
+
+for.cond.cleanup:                                 ; preds = %for.body, %if.then
+  %call976 = call <4 x i64> @backsmith_pure_0(<16 x i16> zeroinitializer)
+  br label %if.end
+
+for.body:                                         ; preds = %for.body, %if.then
+  store <32 x i8> %shuffle57, ptr addrspace(5) %BS_VAR_0, align 32
+  %call394 = call <4 x i16> @backsmith_snippet_679()
+  %call393 = call <4 x i16> @backsmith_snippet_679()
+  %cmp104 = icmp ne <32 x i8> %vecinit827, zeroinitializer
+  %sext = sext <32 x i1> %cmp104 to <32 x i8>
+  %vecext347 = extractelement <32 x i8> %sext, i64 29
+  %cond893.v = select i1 %tobool.not, i8 0, i8 %vecext347
+  %vecinit895 = insertelement <32 x i8> %vecinit827, i8 %cond893.v, i64 29
+  store <32 x i8> %vecinit895, ptr addrspace(5) null, align 32
+  %exitcond.not = icmp eq i64 0, %add919
+  br i1 %exitcond.not, label %for.cond.cleanup, label %for.body
+
+if.end:                                           ; preds = %for.cond.cleanup, %entry
+  %call1066 = call <4 x i64> @backsmith_pure_0(<16 x i16> %0)
+  ret i16 0
+}
+
+declare i8 @backsmith_snippet_44()
+
+declare <4 x i16> @backsmith_snippet_679()
+
+declare <4 x i64> @backsmith_pure_0(<16 x i16>)
+
+define %struct.S1 @func_103(i8 %0, i1 %tobool2.not30) {
+entry:
+  %add = shufflevector <4 x i16> splat (i16 1), <4 x i16> zeroinitializer, <4 x i32> zeroinitializer
+  br label %for.cond1.preheader
+
+for.cond.loopexit:                                ; preds = %for.end, %for.cond1.preheader
+  %tobool.not = icmp eq i8 %0, 0
+  br i1 %tobool.not, label %for.end25, label %for.cond1.preheader
+
+for.cond1.preheader:                              ; preds = %for.cond.loopexit, %entry
+  br i1 %tobool2.not30, label %for.cond.loopexit, label %for.body3
+
+for.body3:                                        ; preds = %for.end, %for.cond1.preheader
+  br label %for.cond4
+
+for.cond4:                                        ; preds = %for.cond4, %for.body3
+  %tobool7.not = icmp eq i64 0, 0
+  br i1 %tobool7.not, label %for.end, label %for.cond4
+
+for.end:                                          ; preds = %for.cond4
+  %call121 = call i16 @backsmith_pure_2(ptr addrspace(5) null, i1 false, i64 0, <32 x i8> zeroinitializer, <32 x i8> zeroinitializer, <16 x i16> zeroinitializer)
+  store <4 x i16> %add, ptr addrspace(5) null, align 8
+  tail call void @backsmith_pure_7(<16 x i64> zeroinitializer, i32 -352674805, <2 x i32> splat (i32 -247168742))
+  tail call void @backsmith_pure_7(<16 x i64> zeroinitializer, i32 -352674805, <2 x i32> splat (i32 -247168742))
+  %tobool2.not = icmp eq ptr addrspace(5) null, null
+  br i1 %tobool2.not, label %for.cond.loopexit, label %for.body3
+
+for.end25:                                        ; preds = %for.cond.loopexit
+  ret %struct.S1 zeroinitializer
+}
+
+declare void @backsmith_pure_7(<16 x i64>, i32, <2 x i32>)
+
+; uselistorder directives
+uselistorder ptr @backsmith_snippet_679, { 1, 0 }
+uselistorder ptr @backsmith_pure_0, { 1, 0 }
+uselistorder ptr @backsmith_pure_7, { 1, 0 }
+
+!0 = !{!"branch_weights", !"expected", i32 2000, i32 1}
----------------
MrSidims wrote:

Apparently - yes. Removing the metadata makes the sample passing the backend.

Thinking out loud: I'm failing the create a .mir reproducer as well, even if I dump mir just before greedy - such sample passes the following tool chain. It means, that actually some pipeline state affects the bug. Meanwhile, this metadata doesn't affect generated mir and apparently it affects the state, cause the bug. And it kind of put under a question not only the approach I have in the PR but also a bit improved, where I remove blocks that are not covered the live interval:
```
  if (!CurLI->empty()) {
    UseSlots.erase(std::remove_if(UseSlots.begin(), UseSlots.end(),
                                  [this](SlotIndex Idx) {
                                    SlotIndex Start, Stop;
                                    std::tie(Start, Stop) =
                                        LIS.getSlotIndexes()->getMBBRange(
                                            LIS.getMBBFromIndex(Idx));
                                    return !CurLI->overlaps(Start, Stop);
                                  }),
                   UseSlots.end());
  }
```

May be the edit on top of this comment is fine, but let me move the PR to draft until I sure, where is the bug. Probably instead of patching split I should check and fix the BranchProbabilityInfo and/or BlockFrequencyInfo instead of regalloc.

https://github.com/llvm/llvm-project/pull/177270