[SelectionDAG][RISCV] Mask constants to narrow size in TargetLowering::expandUnalignedStore.

[topperc]

If the SRL for Hi constant folds, but we don't remove those bits from the Lo, we can end up with strange constant folding. I've only seen this with constants being lowered to constant pools during lowering on RISC-V.

On the first split we create two i32 trunc stores and a srl to shift the high part down. The srl gets constant folded, but to produce a new i32 constant. But the truncstore for the low store still uses the original constant.

This original constant then gets converted to a constant pool before we revisit the stores to further split them. The constant pool prevents further constant folding of the additional srls.

After legalization is done, we run DAGCombiner and get some constant folding of srl via computeKnownBits which can peek through the constant pool load. This can create new constants that also need a constant pool.

[llvmbot]

@llvm/pr-subscribers-llvm-selectiondag

@llvm/pr-subscribers-backend-risc-v

Changes

If the SRL for Hi constant folds, but we don't remoe those bits from the Lo, we can end up with strange constant folding. I've only seen this with constants being lowered to constant pools during lowering on RISC-V.

On the first split we create two i32 trunc stores and a srl to shift the high part down. The srl gets constant folded, but to produce a new i32 constant. But the truncstore for the low store still uses the original constant.

This original constant then gets converted to a constant pool before we revisit the stores to further split them. The constant pool prevents further constant folding of the additional srls.

After legalization is done, we run DAGCombiner and get some constant folding of srl via computeKnownBits which can peek through the constant pool load. This can create new constants that also need a constant pool.

---

Full diff: https://github.com/llvm/llvm-project/pull/66567.diff

2 Files Affected:

• (modified) llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp (+8)
• (modified) llvm/test/CodeGen/RISCV/unaligned-load-store.ll (+41)

diff --git a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
index bd1940994a87f0f..1a7799816711c03 100644
--- a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -9558,6 +9558,14 @@ SDValue TargetLowering::expandUnalignedStore(StoreSDNode *ST,
   SDValue ShiftAmount = DAG.getConstant(
       NumBits, dl, getShiftAmountTy(Val.getValueType(), DAG.getDataLayout()));
   SDValue Lo = Val;
+  // If Val is a constant, replace the upper bits with 0. The SRL will constant
+  // fold and not use the upper bits. A smaller constant may be easier to
+  // materialize.
+  if (auto *C = dyn_cast<ConstantSDNode>(Lo); C && !C->isOpaque())
+    Lo = DAG.getNode(
+        ISD::AND, dl, VT, Lo,
+        DAG.getConstant(APInt::getLowBitsSet(VT.getSizeInBits(), NumBits), dl,
+                        VT));
   SDValue Hi = DAG.getNode(ISD::SRL, dl, VT, Val, ShiftAmount);
 
   // Store the two parts
diff --git a/llvm/test/CodeGen/RISCV/unaligned-load-store.ll b/llvm/test/CodeGen/RISCV/unaligned-load-store.ll
index 429f0543b41b3db..ce0d8fedbfb88f2 100644
--- a/llvm/test/CodeGen/RISCV/unaligned-load-store.ll
+++ b/llvm/test/CodeGen/RISCV/unaligned-load-store.ll
@@ -415,3 +415,44 @@ define void @merge_stores_i32_i64(ptr %p) {
   store i32 0, ptr %p2
   ret void
 }
+
+define void @store_large_constant(ptr %x) {
+; SLOW-LABEL: store_large_constant:
+; SLOW:       # %bb.0:
+; SLOW-NEXT:    li a1, 254
+; SLOW-NEXT:    sb a1, 7(a0)
+; SLOW-NEXT:    li a1, 220
+; SLOW-NEXT:    sb a1, 6(a0)
+; SLOW-NEXT:    li a1, 186
+; SLOW-NEXT:    sb a1, 5(a0)
+; SLOW-NEXT:    li a1, 152
+; SLOW-NEXT:    sb a1, 4(a0)
+; SLOW-NEXT:    li a1, 118
+; SLOW-NEXT:    sb a1, 3(a0)
+; SLOW-NEXT:    li a1, 84
+; SLOW-NEXT:    sb a1, 2(a0)
+; SLOW-NEXT:    li a1, 50
+; SLOW-NEXT:    sb a1, 1(a0)
+; SLOW-NEXT:    li a1, 16
+; SLOW-NEXT:    sb a1, 0(a0)
+; SLOW-NEXT:    ret
+;
+; RV32I-FAST-LABEL: store_large_constant:
+; RV32I-FAST:       # %bb.0:
+; RV32I-FAST-NEXT:    lui a1, 1043916
+; RV32I-FAST-NEXT:    addi a1, a1, -1384
+; RV32I-FAST-NEXT:    sw a1, 4(a0)
+; RV32I-FAST-NEXT:    lui a1, 484675
+; RV32I-FAST-NEXT:    addi a1, a1, 528
+; RV32I-FAST-NEXT:    sw a1, 0(a0)
+; RV32I-FAST-NEXT:    ret
+;
+; RV64I-FAST-LABEL: store_large_constant:
+; RV64I-FAST:       # %bb.0:
+; RV64I-FAST-NEXT:    lui a1, %hi(.LCPI16_0)
+; RV64I-FAST-NEXT:    ld a1, %lo(.LCPI16_0)(a1)
+; RV64I-FAST-NEXT:    sd a1, 0(a0)
+; RV64I-FAST-NEXT:    ret
+  store i64 18364758544493064720, ptr %x, align 1
+  ret void
+}

[RKSimon]

LGTM

[asb]

LGTM

[topperc]

committed 17a12a2 and 8f04d81