[SelectionDAGBuilder] Bugfix in visitInlineAsm()

[jonpa]

Summary

In case of a virtual register tied to a phys-def, the register class needs to be computed. Make sure that this works generally also with fast regalloc by using TLI.getRegClassFor() whenever possible, and make only the case of 'Untyped' use getMinimalPhysRegClass().

Fixes https://bugs.llvm.org/show_bug.cgi?id=51699

[uweigand]

As you discuss in the bugzilla, this seems a good compromise between using the (larger) RC associated with the type when available, and falling back to the (minimal) RC associated with the physreg if we have a "weird" type. LGTM.

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

case InlineAsm::isInput: {
 
           SmallVector<unsigned, 4> Regs;
           MachineFunction &MF = DAG.getMachineFunction();
           MachineRegisterInfo &MRI = MF.getRegInfo();
           const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
           auto *R = cast<RegisterSDNode>(AsmNodeOperands[CurOp+1]);
           Register TiedReg = R->getReg();
           MVT RegVT = R->getSimpleValueType(0);
-          const TargetRegisterClass *RC = TiedReg.isVirtual() ?
-            MRI.getRegClass(TiedReg) : TRI.getMinimalPhysRegClass(TiedReg);
+          const TargetRegisterClass *RC =
+              TiedReg.isVirtual()     ? MRI.getRegClass(TiedReg)
+              : RegVT != MVT::Untyped ? TLI.getRegClassFor(RegVT)
+                                      : TRI.getMinimalPhysRegClass(TiedReg);
           unsigned NumRegs = InlineAsm::getNumOperandRegisters(OpFlag);
           for (unsigned i = 0; i != NumRegs; ++i)
             Regs.push_back(MRI.createVirtualRegister(RC));
 
           RegsForValue MatchedRegs(Regs, RegVT, InOperandVal.getValueType());
 
           SDLoc dl = getCurSDLoc();
           // Use the produced MatchedRegs object to

llvm/test/CodeGen/X86/20210831-inlineasm.ll

+; RUN: llc < %s -O0 -mtriple=x86_64-unknown-linux-gnu
+; https://bugs.llvm.org/show_bug.cgi?id=51699
+
+%"[]u8" = type { i8*, i64 }
+%std.mem.Allocator = type { void ({ %"[]u8", i16 }*, %std.builtin.StackTrace*, %std.mem.Allocator*, i64, i29, i29, i64)*, void ({ i64, i16 }*, %std.builtin.StackTrace*, %std.mem.Allocator*, %"[]u8"*, i29, i64, i29, i64)* }
+%std.builtin.StackTrace = type { i64, %"[]usize" }
+%"[]usize" = type { i64*, i64 }
+
+define  void @fun(%"[]u8"* %0) #0 {
+Entry:
+  %1 = alloca [6 x i64], align 8
+  br label %ErrRetContinue
+
+ErrRetContinue:                                   ; preds = %Entry
+  %2 = call i64 asm sideeffect "rolq $$3,  %rdi ; rolq $$13, %rdi\0Arolq $$61, %rdi ; rolq $$51, %rdi\0Axchgq %rbx,%rbx\0A", "={rdx},{rax},0,~{cc},~{memory}"(i64 undef, i64 0)
+  %3 = call fastcc i64 undef(%std.mem.Allocator* undef, %"[]u8"* %0, i29 undef, i64 0, i29 0, i64 undef)
+  ret void
+}
+
+attributes #0 = { sspstrong }