[llvm] [DAG] isAllOnesOrAllOnesSplat - handle implicitly truncated constants (PR #200209)

[via llvm-commits]

llvmorg-github-actions[bot] wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-loongarch

Author: Simon Pilgrim (RKSimon)

<details>
<summary>Changes</summary>

Allow implicit truncation in BUILD_VECTOR/SPLAT_VECTOR splat nodes - we only need 1's in the lower bits of every element

More yak shaving for #<!-- -->198162 where we often have <8 x i1> <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1> used in XOR patterns.

---
Full diff: https://github.com/llvm/llvm-project/pull/200209.diff


2 Files Affected:

- (modified) llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp (+3-2) 
- (modified) llvm/test/CodeGen/LoongArch/lsx/vmskcond.ll (+6-12) 


``````````diff
diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index b92ba79564f56..75d550801315b 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -13795,8 +13795,9 @@ bool llvm::isOneOrOneSplatFP(SDValue N, bool AllowUndefs) {
 bool llvm::isAllOnesOrAllOnesSplat(SDValue N, bool AllowUndefs) {
   N = peekThroughBitcasts(N);
   unsigned BitWidth = N.getScalarValueSizeInBits();
-  ConstantSDNode *C = isConstOrConstSplat(N, AllowUndefs);
-  return C && C->isAllOnes() && C->getValueSizeInBits(0) == BitWidth;
+  ConstantSDNode *C =
+      isConstOrConstSplat(N, AllowUndefs, /*AllowTruncation=*/true);
+  return C && C->getAPIntValue().countTrailingOnes() >= BitWidth;
 }
 
 bool llvm::isOnesOrOnesSplat(SDValue N, bool AllowUndefs) {
diff --git a/llvm/test/CodeGen/LoongArch/lsx/vmskcond.ll b/llvm/test/CodeGen/LoongArch/lsx/vmskcond.ll
index 2963c40e8d1b4..f6176c7479a37 100644
--- a/llvm/test/CodeGen/LoongArch/lsx/vmskcond.ll
+++ b/llvm/test/CodeGen/LoongArch/lsx/vmskcond.ll
@@ -658,11 +658,10 @@ entry:
 define i32 @vmsk2_sgt_allones_i8(<32 x i8> %a) {
 ; LA32-LABEL: vmsk2_sgt_allones_i8:
 ; LA32:       # %bb.0: # %entry
-; LA32-NEXT:    vrepli.b $vr2, -1
-; LA32-NEXT:    vslt.b $vr0, $vr2, $vr0
+; LA32-NEXT:    vxori.b $vr0, $vr0, 255
 ; LA32-NEXT:    vmskltz.b $vr0, $vr0
 ; LA32-NEXT:    vpickve2gr.hu $a0, $vr0, 0
-; LA32-NEXT:    vslt.b $vr0, $vr2, $vr1
+; LA32-NEXT:    vxori.b $vr0, $vr1, 255
 ; LA32-NEXT:    vmskltz.b $vr0, $vr0
 ; LA32-NEXT:    vpickve2gr.hu $a1, $vr0, 0
 ; LA32-NEXT:    slli.w $a1, $a1, 16
@@ -671,11 +670,10 @@ define i32 @vmsk2_sgt_allones_i8(<32 x i8> %a) {
 ;
 ; LA64-LABEL: vmsk2_sgt_allones_i8:
 ; LA64:       # %bb.0: # %entry
-; LA64-NEXT:    vrepli.b $vr2, -1
-; LA64-NEXT:    vslt.b $vr0, $vr2, $vr0
+; LA64-NEXT:    vxori.b $vr0, $vr0, 255
 ; LA64-NEXT:    vmskltz.b $vr0, $vr0
 ; LA64-NEXT:    vpickve2gr.hu $a0, $vr0, 0
-; LA64-NEXT:    vslt.b $vr0, $vr2, $vr1
+; LA64-NEXT:    vxori.b $vr0, $vr1, 255
 ; LA64-NEXT:    vmskltz.b $vr0, $vr0
 ; LA64-NEXT:    vpickve2gr.hu $a1, $vr0, 0
 ; LA64-NEXT:    slli.d $a1, $a1, 16
@@ -776,11 +774,9 @@ entry:
 define i32 @vmsk2_sle_allones_i8(<32 x i8> %a) {
 ; LA32-LABEL: vmsk2_sle_allones_i8:
 ; LA32:       # %bb.0: # %entry
-; LA32-NEXT:    vslei.b $vr0, $vr0, -1
 ; LA32-NEXT:    vmskltz.b $vr0, $vr0
 ; LA32-NEXT:    vpickve2gr.hu $a0, $vr0, 0
-; LA32-NEXT:    vslei.b $vr0, $vr1, -1
-; LA32-NEXT:    vmskltz.b $vr0, $vr0
+; LA32-NEXT:    vmskltz.b $vr0, $vr1
 ; LA32-NEXT:    vpickve2gr.hu $a1, $vr0, 0
 ; LA32-NEXT:    slli.w $a1, $a1, 16
 ; LA32-NEXT:    or $a0, $a0, $a1
@@ -788,11 +784,9 @@ define i32 @vmsk2_sle_allones_i8(<32 x i8> %a) {
 ;
 ; LA64-LABEL: vmsk2_sle_allones_i8:
 ; LA64:       # %bb.0: # %entry
-; LA64-NEXT:    vslei.b $vr0, $vr0, -1
 ; LA64-NEXT:    vmskltz.b $vr0, $vr0
 ; LA64-NEXT:    vpickve2gr.hu $a0, $vr0, 0
-; LA64-NEXT:    vslei.b $vr0, $vr1, -1
-; LA64-NEXT:    vmskltz.b $vr0, $vr0
+; LA64-NEXT:    vmskltz.b $vr0, $vr1
 ; LA64-NEXT:    vpickve2gr.hu $a1, $vr0, 0
 ; LA64-NEXT:    slli.d $a1, $a1, 16
 ; LA64-NEXT:    or $a0, $a0, $a1

``````````

</details>


https://github.com/llvm/llvm-project/pull/200209