[llvm] [SDAG] Add missing float type legalizations for FMODF (PR #128055)

[Eli Friedman via llvm-commits]

================
@@ -2616,6 +2624,18 @@ bool SelectionDAG::expandMultipleResultFPLibCall(
 
   auto [Call, CallChain] = TLI->LowerCallTo(CLI);
 
+  if (CallRetResNo && !Node->hasAnyUseOfValue(*CallRetResNo)) {
+    // This is needed for x87, which uses a floating-point stack. If (for
----------------
efriedma-quic wrote:

The "generic" issue is that libcalls inserted after SelectionDAGBuilder are weird.  Normal calls have a unique ordering: the end of one call is attached to the beginning of the next.  libcalls are sustained by a series of hacks across the code.  We have a bunch of code to make this work, including code in the scheduler to ensure call sequences don't overlap.

The x86-specific issue is that the "CopyFromReg" has a side-effect.  Well, not really a side-effect, but the COPY needs to be present so that later code can correctly compute the depth of the x87 stack.  This is usually not true for CopyFromReg operations: for almost all targets, for almost all register classes, copying a register doesn't affect any other state, so it's safe to delete a copy.

I was thinking X86ISD::PopX87Reg would just lower to the same COPY we currently generate for CopyFromReg; whether that actually lowers to a pop would be decided later.

The other possibility I was thinking about is whether we can pull the CopyFromReg inside the call sequence, so it's before the adjstackup instead of after it.  That would give it the same scheduling protections as other bits of code inside call sequences.  But I'm not sure if that would have other effects.

https://github.com/llvm/llvm-project/pull/128055