[SDAG] Lower range attribute to AssertZext

[andjo403]

split out as requested in #94847 (comment)
CC @nikic

[llvmbot]

@llvm/pr-subscribers-backend-aarch64
@llvm/pr-subscribers-backend-x86

@llvm/pr-subscribers-llvm-selectiondag

Author: Andreas Jonson (andjo403)

Changes

split out as requested in #94847 (comment)
CC @nikic

---

Full diff: https://github.com/llvm/llvm-project/pull/95450.diff

2 Files Affected:

• (modified) llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp (+16-7)
• (modified) llvm/test/CodeGen/X86/legalize-vec-assertzext.ll (+18)

diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
index be5e0f6ef058b..5d21a905c316c 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
@@ -4487,6 +4487,18 @@ static const MDNode *getRangeMetadata(const Instruction &I) {
   return I.getMetadata(LLVMContext::MD_range);
 }
 
+static std::optional<ConstantRange> getRange(const Instruction &I) {
+  if (const auto *CB = dyn_cast<CallBase>(&I)) {
+    // see comment in getRangeMetadata about this check
+    if (CB->hasRetAttr(Attribute::NoUndef))
+      return CB->getRange();
+  }
+  if (const MDNode *Range = getRangeMetadata(I)) {
+    return getConstantRangeFromMetadata(*Range);
+  }
+  return std::nullopt;
+}
+
 void SelectionDAGBuilder::visitLoad(const LoadInst &I) {
   if (I.isAtomic())
     return visitAtomicLoad(I);
@@ -10229,19 +10241,16 @@ void SelectionDAGBuilder::visitVACopy(const CallInst &I) {
 SDValue SelectionDAGBuilder::lowerRangeToAssertZExt(SelectionDAG &DAG,
                                                     const Instruction &I,
                                                     SDValue Op) {
-  const MDNode *Range = getRangeMetadata(I);
-  if (!Range)
-    return Op;
+  std::optional<ConstantRange> CR = getRange(I);
 
-  ConstantRange CR = getConstantRangeFromMetadata(*Range);
-  if (CR.isFullSet() || CR.isEmptySet() || CR.isUpperWrapped())
+  if (!CR || CR->isFullSet() || CR->isEmptySet() || CR->isUpperWrapped())
     return Op;
 
-  APInt Lo = CR.getUnsignedMin();
+  APInt Lo = CR->getUnsignedMin();
   if (!Lo.isMinValue())
     return Op;
 
-  APInt Hi = CR.getUnsignedMax();
+  APInt Hi = CR->getUnsignedMax();
   unsigned Bits = std::max(Hi.getActiveBits(),
                            static_cast<unsigned>(IntegerType::MIN_INT_BITS));
 
diff --git a/llvm/test/CodeGen/X86/legalize-vec-assertzext.ll b/llvm/test/CodeGen/X86/legalize-vec-assertzext.ll
index 1c595b7fb5e1e..2cf37c68b8b40 100644
--- a/llvm/test/CodeGen/X86/legalize-vec-assertzext.ll
+++ b/llvm/test/CodeGen/X86/legalize-vec-assertzext.ll
@@ -34,6 +34,24 @@ define i64 @widen_assertzext(ptr %x) nounwind {
   ret i64 %d
 }
 
+define i64 @widen_assertzext_range_attr(ptr %x) nounwind {
+; CHECK-LABEL: widen_assertzext_range_attr:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    pushq %rax
+; CHECK-NEXT:    callq test2@PLT
+; CHECK-NEXT:    movb $127, %al
+; CHECK-NEXT:    kmovw %eax, %k1
+; CHECK-NEXT:    vpexpandq %zmm0, %zmm0 {%k1} {z}
+; CHECK-NEXT:    vextracti32x4 $3, %zmm0, %xmm0
+; CHECK-NEXT:    vmovq %xmm0, %rax
+; CHECK-NEXT:    popq %rcx
+; CHECK-NEXT:    vzeroupper
+; CHECK-NEXT:    retq
+  %e = call noundef range(i64 0, 2) <7 x i64> @test2()
+  %d = extractelement <7 x i64> %e, i32 6
+  ret i64 %d
+}
+
 declare  <16 x i64> @test()
 declare  <7 x i64> @test2()
 !0 = !{ i64 0, i64 2 }

[andjo403]

tests added

[nikic]

LGTM

It looks like MachineMemOperand represents ranges as const MDNode *, so supporting range attributes for other getRangeMetadata() calls in this file may be tricky :(

[andjo403]

fixed the braces. have not received merge access (but have asked for it now at least)

yes the MachineMemOperand representation have me a bit stumped have tried to replace it with std::optional<ConstantRange> but have hit problems with the MIParser that I think that I need to update

[nikic]

@andjo403 The thing I would be concerned about is memory usage. Currently this is one pointer, while ConstantRange is four pointers large (though it only actually holds two and a half pointers worth of data...)