[ARM] Fix FP16 stack loads/stores for Thumb2 with frame pointer

[olista01]

Summary

The new addressing mode added for the v8.2A FP16 instructions uses bit 8 of the immediate to encode the sign of the offset, like the other FP loads/stores, so need to be treated the same way.

[SjoerdMeijer]

Looks like a straightforward fix to me.

test/CodeGen/ARM/fp16-frame-lowering.ll

4 ↗

(On Diff #188874)

nit: perhaps omit "that"?

llvm/trunk/lib/Target/ARM/Thumb2InstrInfo.cpp

if (Opcode == ARM::t2ADDri || Opcode == ARM::t2ADDri12) {
     // Attempt to fold address computation
     // Common case: small offset, fits into instruction.
     int ImmedOffset = Offset / Scale;
     unsigned Mask = (1 << NumBits) - 1;
     if ((unsigned)Offset <= Mask * Scale) {
       // Replace the FrameIndex with fp/sp
       MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
       if (isSub) {
-        if (AddrMode == ARMII::AddrMode5)
+        if (AddrMode == ARMII::AddrMode5 || AddrMode == ARMII::AddrMode5FP16)
           // FIXME: Not consistent.
           ImmedOffset |= 1 << NumBits;
         else
           ImmedOffset = -ImmedOffset;
       }
       ImmOp.ChangeToImmediate(ImmedOffset);
       Offset = 0;
       return true;
     }
 
     // Otherwise, offset doesn't fit. Pull in what we can to simplify
     ImmedOffset = ImmedOffset & Mask;
     if (isSub) {
-      if (AddrMode == ARMII::AddrMode5)
+      if (AddrMode == ARMII::AddrMode5 || AddrMode == ARMII::AddrMode5FP16)
         // FIXME: Not consistent.
         ImmedOffset |= 1 << NumBits;
       else {
         ImmedOffset = -ImmedOffset;
         if (ImmedOffset == 0)
           // Change the opcode back if the encoded offset is zero.
           MI.setDesc(TII.get(positiveOffsetOpcode(NewOpc)));
       }

llvm/trunk/test/CodeGen/ARM/fp16-frame-lowering.ll

+; RUN: llc < %s -mtriple armv8a--none-eabi -mattr=+fullfp16             | FileCheck %s
+; RUN: llc < %s -mtriple armv8a--none-eabi -mattr=+fullfp16,+thumb-mode | FileCheck %s
+
+; Check that frame lowering for the fp16 instructions works correctly with
+; negative offsets (which happens when using the frame pointer).
+
+define void @foo(i32 %count) {
+entry:
+  %half_alloca = alloca half, align 2
+; CHECK: vstr.16 {{s[0-9]+}}, [{{r[0-9]+}}, #-10]
+  store half 0.0, half* %half_alloca
+  call void @bar(half* %half_alloca)
+
+  ; A variable-sized alloca to force the above store to use the frame pointer
+  ; instead of the stack pointer, and so need a negative offset.
+  %var_alloca = alloca i32, i32 %count
+  call void @baz(i32* %var_alloca)
+  ret void
+}
+
+declare void @bar(half*)
+declare void @baz(i32*)