Added sm_101 exception logic

Author: rajatbajpai

llvm/lib/Target/NVPTX/NVPTXInstrInfo.td

@@ -109,13 +109,15 @@ class AnyPred<list<Predicate> Preds>
                 ")">;
 
 // Checks PTX version and family-specific and architecture-specific SM versions.
-// For example, sm_100{f/a} and any future variants in the same family will match.
+// For example, sm_100{f/a} and any future variants in the same family will match
+// for any PTX version greater than or equal to `PTXVersion`.
 class PTXWithFamilySMs<int PTXVersion, list<int> SMVersions> :
   Predicate<"Subtarget->hasPTXWithFamilySMs(" # PTXVersion # ", {" #
             !interleave(SMVersions, ", ") # "})">;
 
 // Checks PTX version and architecture-specific SM versions.
-// For example, sm_100{a} will match.
+// For example, sm_100{a} will match for any PTX version
+// greater than or equal to `PTXVersion`.
 class PTXWithAccelSMs<int PTXVersion, list<int> SMVersions> :
   Predicate<"Subtarget->hasPTXWithAccelSMs(" # PTXVersion # ", {" #
             !interleave(SMVersions, ", ") # "})">;

llvm/lib/Target/NVPTX/NVPTXSubtarget.cpp

@@ -74,21 +74,36 @@ const SelectionDAGTargetInfo *NVPTXSubtarget::getSelectionDAGInfo() const {
 
 bool NVPTXSubtarget::hasPTXWithFamilySMs(unsigned PTXVersion,
                                          ArrayRef<unsigned> SMVersions) const {
-  if (!hasFamilySpecificFeatures() || getPTXVersion() < PTXVersion)
+  unsigned PTXVer = getPTXVersion();
+  if (!hasFamilySpecificFeatures() || PTXVer < PTXVersion)
     return false;
 
+  unsigned SMVer = getSmVersion();
   return llvm::any_of(SMVersions, [&](unsigned SM) {
-    return getSmFamilyVersion() == SM / 10 && getSmVersion() >= SM;
+    // sm_101 is a different family, never group it with sm_10x.
+    if (SMVer == 101 || SM == 101)
+      return SMVer == SM &&
+             // PTX 9.0 and later renamed sm_101 to sm_110, so sm_101 is not
+             // supported.
+             !(PTXVer >= 90 && SMVer == 101);
+
+    return getSmFamilyVersion() == SM / 10 && SMVer >= SM;
   });
 }
 
 bool NVPTXSubtarget::hasPTXWithAccelSMs(unsigned PTXVersion,
                                         ArrayRef<unsigned> SMVersions) const {
-  if (!hasArchAccelFeatures() || getPTXVersion() < PTXVersion)
+  unsigned PTXVer = getPTXVersion();
+  if (!hasArchAccelFeatures() || PTXVer < PTXVersion)
     return false;
 
-  return llvm::any_of(SMVersions,
-                      [&](unsigned SM) { return getSmVersion() == SM; });
+  unsigned SMVer = getSmVersion();
+  return llvm::any_of(SMVersions, [&](unsigned SM) {
+    return SMVer == SM &&
+           // PTX 9.0 and later renamed sm_101 to sm_110, so sm_101 is not
+           // supported.
+           !(PTXVer >= 90 && SMVer == 101);
+  });
 }
 
 bool NVPTXSubtarget::allowFP16Math() const {

llvm/lib/Target/NVPTX/NVPTXSubtarget.h

@@ -75,11 +75,13 @@ class NVPTXSubtarget : public NVPTXGenSubtargetInfo {
 
   // Checks PTX version and family-specific and architecture-specific SM
   // versions. For example, sm_100{f/a} and any future variants in the same
-  // family will match.
+  // family will match for any PTX version greater than or equal to
+  // `PTXVersion`.
   bool hasPTXWithFamilySMs(unsigned PTXVersion,
                            ArrayRef<unsigned> SMVersions) const;
   // Checks PTX version and architecture-specific SM versions.
-  // For example, sm_100{a} will match.
+  // For example, sm_100{a} will match for any PTX version greater than or equal
+  // to `PTXVersion`.
   bool hasPTXWithAccelSMs(unsigned PTXVersion,
                           ArrayRef<unsigned> SMVersions) const;