[DAGCombiner][RISCV] CSE zext nneg and sext.

[topperc]

If we have a sext and a zext nneg with the same types and operand
we should combine them into the sext. We can't go the other way
because the nneg flag may only be valid in the context of the uses
of the zext nneg.

The test case load, compare, branch, and zext are based on real patterns I've seen.

[llvmbot]

@llvm/pr-subscribers-llvm-selectiondag

Author: Craig Topper (topperc)

Changes

If we have a sext and a zext nneg with the same types and operand
we should combine them into the sext. We can't go the other way
because the nneg flag may only be valid in the context of the uses
of the zext nneg.

The test case load, compare, branch, and zext are based on real patterns I've seen.

---

Full diff: https://github.com/llvm/llvm-project/pull/82597.diff

2 Files Affected:

• (modified) llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp (+7)
• (modified) llvm/test/CodeGen/RISCV/sext-zext-trunc.ll (+61)

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 89ef648ee7d7ed..ed43dd7f528821 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -13997,6 +13997,13 @@ SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
   if (SDValue Res = tryToFoldExtendSelectLoad(N, TLI, DAG, Level))
     return Res;
 
+  // CSE zext nneg with sext if the zext is not free.
+  if (N->getFlags().hasNonNeg() && !TLI.isZExtFree(N0.getValueType(), VT)) {
+    SDNode *CSENode = DAG.getNodeIfExists(ISD::SIGN_EXTEND, N->getVTList(), N0);
+    if (CSENode)
+      return SDValue(CSENode, 0);
+  }
+
   return SDValue();
 }
 
diff --git a/llvm/test/CodeGen/RISCV/sext-zext-trunc.ll b/llvm/test/CodeGen/RISCV/sext-zext-trunc.ll
index a2a953ca882bad..255a1257ef43ad 100644
--- a/llvm/test/CodeGen/RISCV/sext-zext-trunc.ll
+++ b/llvm/test/CodeGen/RISCV/sext-zext-trunc.ll
@@ -871,3 +871,64 @@ define void @zext_nneg_dominating_icmp_i32_zeroext(i16 signext %0) {
 5:
   ret void
 }
+
+; The load is used extended and non-extended in the successor basic block. The
+; signed compare will cause the non-extended value to exported out of the first
+; basic block using a sext. We need to CSE the zext nneg with the sext so that
+; we can form a sextload.
+define void @load_zext_nneg_sext_cse(ptr %p) nounwind {
+; RV32I-LABEL: load_zext_nneg_sext_cse:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    addi sp, sp, -16
+; RV32I-NEXT:    sw ra, 12(sp) # 4-byte Folded Spill
+; RV32I-NEXT:    sw s0, 8(sp) # 4-byte Folded Spill
+; RV32I-NEXT:    lh s0, 0(a0)
+; RV32I-NEXT:    bltz s0, .LBB50_2
+; RV32I-NEXT:  # %bb.1: # %bb1
+; RV32I-NEXT:    mv a0, s0
+; RV32I-NEXT:    call bar_i16
+; RV32I-NEXT:    mv a0, s0
+; RV32I-NEXT:    lw ra, 12(sp) # 4-byte Folded Reload
+; RV32I-NEXT:    lw s0, 8(sp) # 4-byte Folded Reload
+; RV32I-NEXT:    addi sp, sp, 16
+; RV32I-NEXT:    tail bar_i32
+; RV32I-NEXT:  .LBB50_2: # %bb2
+; RV32I-NEXT:    lw ra, 12(sp) # 4-byte Folded Reload
+; RV32I-NEXT:    lw s0, 8(sp) # 4-byte Folded Reload
+; RV32I-NEXT:    addi sp, sp, 16
+; RV32I-NEXT:    ret
+;
+; RV64-LABEL: load_zext_nneg_sext_cse:
+; RV64:       # %bb.0:
+; RV64-NEXT:    addi sp, sp, -16
+; RV64-NEXT:    sd ra, 8(sp) # 8-byte Folded Spill
+; RV64-NEXT:    sd s0, 0(sp) # 8-byte Folded Spill
+; RV64-NEXT:    lh s0, 0(a0)
+; RV64-NEXT:    bltz s0, .LBB50_2
+; RV64-NEXT:  # %bb.1: # %bb1
+; RV64-NEXT:    mv a0, s0
+; RV64-NEXT:    call bar_i16
+; RV64-NEXT:    mv a0, s0
+; RV64-NEXT:    ld ra, 8(sp) # 8-byte Folded Reload
+; RV64-NEXT:    ld s0, 0(sp) # 8-byte Folded Reload
+; RV64-NEXT:    addi sp, sp, 16
+; RV64-NEXT:    tail bar_i32
+; RV64-NEXT:  .LBB50_2: # %bb2
+; RV64-NEXT:    ld ra, 8(sp) # 8-byte Folded Reload
+; RV64-NEXT:    ld s0, 0(sp) # 8-byte Folded Reload
+; RV64-NEXT:    addi sp, sp, 16
+; RV64-NEXT:    ret
+  %load = load i16, ptr %p
+  %zext = zext nneg i16 %load to i32
+  %cmp = icmp sgt i16 %load, -1
+  br i1 %cmp, label %bb1, label %bb2
+
+bb1:
+  tail call void @bar_i16(i16 signext %load)
+  tail call void @bar_i32(i32 signext %zext)
+  br label %bb2
+
+bb2:
+  ret void
+}
+declare void @bar_i16(i16);

[RKSimon]

LGTM

[nikic]

Makes me wonder whether we should also CSE these in IR...