[GISel] Fix unmerging of constants for big endian target

[Kai]

Summary

Currently, unmerging of constants is always done in little endian style,
which is wrong for big endian targets. This change fixes this.

I need some advice how to test this, as no in-tree big endian target uses
G_UNMERGE_VALUES. I discovered this in my m88k backend.

[arsenm]

Shouldn't M68K need and be testable with this?

[aemerson]

Could you upload with more context (-U9999).

AArch64 supports big endian, so if you're testing individual passes you can specify a big endian target. (If you try to run a full IR compile it'll fall back to SDAG in the translator).

[arsenm]

Reverse ping

[arsenm]

LGTM, but an m68k test would be nice

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

          "Not enough operands to replace all defs");
   unsigned NumElems = MI.getNumOperands() - 1;
   Builder.setInstrAndDebugLoc(MI);
+  bool IsBigEndianTarget =
+      MI.getParent()->getParent()->getDataLayout().isBigEndian();
   for (unsigned Idx = 0; Idx < NumElems; ++Idx) {
     Register DstReg = MI.getOperand(Idx).getReg();
-    Builder.buildConstant(DstReg, Csts[Idx]);
+    unsigned CstsIdx = IsBigEndianTarget ? bigEndianByteAt(NumElems, Idx)
+                                         : littleEndianByteAt(NumElems, Idx);
+    Builder.buildConstant(DstReg, Csts[CstsIdx]);
   }
 
   MI.eraseFromParent();