[llvm] 3a7c82e - [DAG] isGuaranteedNotToBeUndefOrPoison - handle ISD::BUILD_VECTOR nodes

[Simon Pilgrim via llvm-commits]

Author: Simon Pilgrim
Date: 2021-07-31T15:08:25+01:00
New Revision: 3a7c82efb8db57f0bf1cfbbd681b3905556bd049

URL: https://github.com/llvm/llvm-project/commit/3a7c82efb8db57f0bf1cfbbd681b3905556bd049
DIFF: https://github.com/llvm/llvm-project/commit/3a7c82efb8db57f0bf1cfbbd681b3905556bd049.diff

LOG: [DAG] isGuaranteedNotToBeUndefOrPoison - handle ISD::BUILD_VECTOR nodes

If all demanded elements of the BUILD_VECTOR pass a isGuaranteedNotToBeUndefOrPoison check, then we can treat this specific demanded use of the BUILD_VECTOR as guaranteed not to be undef or poison either.

Differential Revision: https://reviews.llvm.org/D107174

Added: 
    

Modified: 
    llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
    llvm/test/CodeGen/X86/freeze-constant-fold.ll
    llvm/test/CodeGen/X86/freeze-legalize.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index 2a98464425c40..1450bf0eafca3 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -4285,7 +4285,17 @@ bool SelectionDAG::isGuaranteedNotToBeUndefOrPoison(SDValue Op,
   case ISD::UNDEF:
     return PoisonOnly;
 
-  // TODO: ISD::BUILD_VECTOR handling
+  case ISD::BUILD_VECTOR:
+    // NOTE: BUILD_VECTOR has implicit truncation of wider scalar elements -
+    // this shouldn't affect the result.
+    for (unsigned i = 0, e = Op.getNumOperands(); i < e; ++i) {
+      if (!DemandedElts[i])
+        continue;
+      if (!isGuaranteedNotToBeUndefOrPoison(Op.getOperand(i), PoisonOnly,
+                                            Depth + 1))
+        return false;
+    }
+    return true;
 
   // TODO: Search for noundef attributes from library functions.
 

diff  --git a/llvm/test/CodeGen/X86/freeze-constant-fold.ll b/llvm/test/CodeGen/X86/freeze-constant-fold.ll
index 552921918709c..6c22d855e79fe 100644
--- a/llvm/test/CodeGen/X86/freeze-constant-fold.ll
+++ b/llvm/test/CodeGen/X86/freeze-constant-fold.ll
@@ -15,7 +15,6 @@ define i32 @fold_add_freeze_i32() {
   ret i32 %5
 }
 
-; FIXME: X86 scalarization sees through the freeze, but vector types don't.
 define <4 x i32> @fold_add_freeze_v4i32() {
 ; X86-LABEL: fold_add_freeze_v4i32:
 ; X86:       # %bb.0:
@@ -28,9 +27,7 @@ define <4 x i32> @fold_add_freeze_v4i32() {
 ;
 ; X64-LABEL: fold_add_freeze_v4i32:
 ; X64:       # %bb.0:
-; X64-NEXT:    pxor %xmm1, %xmm1
 ; X64-NEXT:    pcmpeqd %xmm0, %xmm0
-; X64-NEXT:    paddd %xmm1, %xmm0
 ; X64-NEXT:    retq
   %1 = insertelement <4 x i32> poison, i32 0, i32 0
   %2 = shufflevector <4 x i32> %1, <4 x i32> poison, <4 x i32> zeroinitializer

diff  --git a/llvm/test/CodeGen/X86/freeze-legalize.ll b/llvm/test/CodeGen/X86/freeze-legalize.ll
index 4fe60bc4cce6c..788ce154dfdce 100644
--- a/llvm/test/CodeGen/X86/freeze-legalize.ll
+++ b/llvm/test/CodeGen/X86/freeze-legalize.ll
@@ -46,7 +46,7 @@ define i10 @promote() {
 define <2 x i10> @promote_vec() {
 ; CHECK-LABEL: promote_vec:
 ; CHECK:       ## %bb.0:
-; CHECK-NEXT:    movw $1674, %ax ## imm = 0x68A
+; CHECK-NEXT:    movw $650, %ax ## imm = 0x28A
 ; CHECK-NEXT:    movw $518, %dx ## imm = 0x206
 ; CHECK-NEXT:    retl
   %a = freeze <2 x i10> <i10 682, i10 125>