[llvm] Preserve range metadata when load is narrowed (PR #128144)

via llvm-commits llvm-commits at lists.llvm.org
Thu Feb 20 23:44:26 PST 2025


llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-llvm-selectiondag

@llvm/pr-subscribers-backend-amdgpu

Author: None (LU-JOHN)

<details>
<summary>Changes</summary>

In DAGCombiner.cpp preserve range metadata when load is narrowed to load LSBs if original range metadata bounds can fit in the narrower type.

Utilize preserved range metadata to reduce 64-bit shl to 32-bit shl.

---
Full diff: https://github.com/llvm/llvm-project/pull/128144.diff


2 Files Affected:

- (modified) llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp (+33-6) 
- (modified) llvm/test/CodeGen/AMDGPU/shl64_reduce.ll (+34-4) 


``````````diff
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index bc7cdf38dbc2a..ba59bf269ba86 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -14903,12 +14903,39 @@ SDValue DAGCombiner::reduceLoadWidth(SDNode *N) {
   AddToWorklist(NewPtr.getNode());
 
   SDValue Load;
-  if (ExtType == ISD::NON_EXTLOAD)
-    Load = DAG.getLoad(VT, DL, LN0->getChain(), NewPtr,
-                       LN0->getPointerInfo().getWithOffset(PtrOff),
-                       LN0->getOriginalAlign(),
-                       LN0->getMemOperand()->getFlags(), LN0->getAAInfo());
-  else
+  if (ExtType == ISD::NON_EXTLOAD) {
+    const MDNode *OldRanges = LN0->getRanges();
+    const MDNode *NewRanges = nullptr;
+    /* If LSBs are loaded and all bounds in the OldRanges metadata fit in
+       the narrower size, preserve the range information by translating
+       to the the new narrower type, NewTy */
+    if (ShAmt == 0 && OldRanges) {
+      Type *NewTy = VT.getTypeForEVT(*DAG.getContext());
+      const unsigned NumOperands = OldRanges->getNumOperands();
+      const unsigned NewWidth = NewTy->getIntegerBitWidth();
+      bool InRange = true;
+      SmallVector<Metadata *, 4> Bounds;
+      Bounds.reserve(NumOperands);
+
+      for (unsigned i = 0; i < NumOperands; ++i) {
+        const APInt &BoundValue =
+            mdconst::extract<ConstantInt>(OldRanges->getOperand(i))->getValue();
+        if (BoundValue.getBitWidth() - BoundValue.getNumSignBits() >=
+            NewWidth) {
+          InRange = false;
+          break;
+        }
+        Bounds.push_back(ConstantAsMetadata::get(
+            ConstantInt::get(NewTy, BoundValue.trunc(NewWidth))));
+      }
+      if (InRange)
+        NewRanges = MDNode::get(*DAG.getContext(), Bounds);
+    }
+    Load = DAG.getLoad(
+        VT, DL, LN0->getChain(), NewPtr,
+        LN0->getPointerInfo().getWithOffset(PtrOff), LN0->getOriginalAlign(),
+        LN0->getMemOperand()->getFlags(), LN0->getAAInfo(), NewRanges);
+  } else
     Load = DAG.getExtLoad(ExtType, DL, VT, LN0->getChain(), NewPtr,
                           LN0->getPointerInfo().getWithOffset(PtrOff), ExtVT,
                           LN0->getOriginalAlign(),
diff --git a/llvm/test/CodeGen/AMDGPU/shl64_reduce.ll b/llvm/test/CodeGen/AMDGPU/shl64_reduce.ll
index 05430213c17d2..55bfc079cb1c4 100644
--- a/llvm/test/CodeGen/AMDGPU/shl64_reduce.ll
+++ b/llvm/test/CodeGen/AMDGPU/shl64_reduce.ll
@@ -21,11 +21,39 @@ define i64 @shl_metadata(i64 %arg0, ptr %arg1.ptr) {
 ; CHECK-LABEL: shl_metadata:
 ; CHECK:       ; %bb.0:
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    flat_load_dword v1, v[2:3]
+; CHECK-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_lshlrev_b32_e32 v1, v1, v0
+; CHECK-NEXT:    v_mov_b32_e32 v0, 0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %shift.amt = load i64, ptr %arg1.ptr, !range !0, !noundef !{}
+  %shl = shl i64 %arg0, %shift.amt
+  ret i64 %shl
+}
+
+define i64 @shl_metadata_two_ranges(i64 %arg0, ptr %arg1.ptr) {
+; CHECK-LABEL: shl_metadata_two_ranges:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    flat_load_dword v1, v[2:3]
+; CHECK-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_lshlrev_b32_e32 v1, v1, v0
+; CHECK-NEXT:    v_mov_b32_e32 v0, 0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %shift.amt = load i64, ptr %arg1.ptr, !range !1, !noundef !{}
+  %shl = shl i64 %arg0, %shift.amt
+  ret i64 %shl
+}
+
+define i64 @shl_metadata_out_of_range(i64 %arg0, ptr %arg1.ptr) {
+; CHECK-LABEL: shl_metadata_out_of_range:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; CHECK-NEXT:    flat_load_dword v2, v[2:3]
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; CHECK-NEXT:    v_lshlrev_b64 v[0:1], v2, v[0:1]
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
-  %shift.amt = load i64, ptr %arg1.ptr, !range !0
+  %shift.amt = load i64, ptr %arg1.ptr, !range !2, !noundef !{}
   %shl = shl i64 %arg0, %shift.amt
   ret i64 %shl
 }
@@ -39,7 +67,7 @@ define <2 x i64> @shl_v2_metadata(<2 x i64> %arg0, ptr %arg1.ptr) {
 ; CHECK-NEXT:    v_lshlrev_b64 v[0:1], v4, v[0:1]
 ; CHECK-NEXT:    v_lshlrev_b64 v[2:3], v6, v[2:3]
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
-  %shift.amt = load <2 x i64>, ptr %arg1.ptr, !range !0
+  %shift.amt = load <2 x i64>, ptr %arg1.ptr, !range !0, !noundef !{}
   %shl = shl <2 x i64> %arg0, %shift.amt
   ret <2 x i64> %shl
 }
@@ -55,7 +83,7 @@ define <3 x i64> @shl_v3_metadata(<3 x i64> %arg0, ptr %arg1.ptr) {
 ; CHECK-NEXT:    v_lshlrev_b64 v[0:1], v8, v[0:1]
 ; CHECK-NEXT:    v_lshlrev_b64 v[2:3], v10, v[2:3]
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
-  %shift.amt = load <3 x i64>, ptr %arg1.ptr, !range !0
+  %shift.amt = load <3 x i64>, ptr %arg1.ptr, !range !0, !noundef !{}
   %shl = shl <3 x i64> %arg0, %shift.amt
   ret <3 x i64> %shl
 }
@@ -74,12 +102,14 @@ define <4 x i64> @shl_v4_metadata(<4 x i64> %arg0, ptr %arg1.ptr) {
 ; CHECK-NEXT:    v_lshlrev_b64 v[4:5], v13, v[4:5]
 ; CHECK-NEXT:    v_lshlrev_b64 v[6:7], v15, v[6:7]
 ; CHECK-NEXT:    s_setpc_b64 s[30:31]
-  %shift.amt = load <4 x i64>, ptr %arg1.ptr, !range !0
+  %shift.amt = load <4 x i64>, ptr %arg1.ptr, !range !0, !noundef !{}
   %shl = shl <4 x i64> %arg0, %shift.amt
   ret <4 x i64> %shl
 }
 
 !0 = !{i64 32, i64 64}
+!1 = !{i64 32, i64 38, i64 42, i64 48}
+!2 = !{i64 31, i64 38, i64 42, i64 48}
 
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ; Test range with an "or X, 16"

``````````

</details>


https://github.com/llvm/llvm-project/pull/128144


More information about the llvm-commits mailing list