[llvm] [clang] [clang-tools-extra] [X86] Use plain load/store instead of cmpxchg16b for atomics with AVX (PR #74275)

[Matt Arsenault via cfe-commits]

================
@@ -30113,32 +30120,40 @@ TargetLoweringBase::AtomicExpansionKind
 X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
   Type *MemType = SI->getValueOperand()->getType();
 
-  bool NoImplicitFloatOps =
-      SI->getFunction()->hasFnAttribute(Attribute::NoImplicitFloat);
-  if (MemType->getPrimitiveSizeInBits() == 64 && !Subtarget.is64Bit() &&
-      !Subtarget.useSoftFloat() && !NoImplicitFloatOps &&
-      (Subtarget.hasSSE1() || Subtarget.hasX87()))
-    return AtomicExpansionKind::None;
+  if (!SI->getFunction()->hasFnAttribute(Attribute::NoImplicitFloat) &&
+      !Subtarget.useSoftFloat()) {
+    if (MemType->getPrimitiveSizeInBits() == 64 && !Subtarget.is64Bit() &&
+        (Subtarget.hasSSE1() || Subtarget.hasX87()))
+      return AtomicExpansionKind::None;
+
+    if (MemType->getPrimitiveSizeInBits() == 128 && Subtarget.is64Bit() &&
+        Subtarget.hasAVX())
+      return AtomicExpansionKind::None;
+  }
 
   return needsCmpXchgNb(MemType) ? AtomicExpansionKind::Expand
                                  : AtomicExpansionKind::None;
 }
 
 // Note: this turns large loads into lock cmpxchg8b/16b.
-// TODO: In 32-bit mode, use MOVLPS when SSE1 is available?
 TargetLowering::AtomicExpansionKind
 X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
   Type *MemType = LI->getType();
 
-  // If this a 64 bit atomic load on a 32-bit target and SSE2 is enabled, we
-  // can use movq to do the load. If we have X87 we can load into an 80-bit
-  // X87 register and store it to a stack temporary.
-  bool NoImplicitFloatOps =
-      LI->getFunction()->hasFnAttribute(Attribute::NoImplicitFloat);
-  if (MemType->getPrimitiveSizeInBits() == 64 && !Subtarget.is64Bit() &&
-      !Subtarget.useSoftFloat() && !NoImplicitFloatOps &&
-      (Subtarget.hasSSE1() || Subtarget.hasX87()))
-    return AtomicExpansionKind::None;
+  if (!LI->getFunction()->hasFnAttribute(Attribute::NoImplicitFloat) &&
+      !Subtarget.useSoftFloat()) {
+    // If this a 64 bit atomic load on a 32-bit target and SSE2 is enabled, we
+    // can use movq to do the load. If we have X87 we can load into an 80-bit
+    // X87 register and store it to a stack temporary.
+    if (MemType->getPrimitiveSizeInBits() == 64 && !Subtarget.is64Bit() &&
----------------
arsenm wrote:

getPrimitiveSizeInBits doesn't work correctly for pointer typed atomics; you need to use the DataLayout. Is there appropriate pointer typed atomic load test coverage for this?

https://github.com/llvm/llvm-project/pull/74275