[llvm] [InstCombine] Canonicalize more geps with constant gep bases and constant offsets. (PR #110033)

[Nikita Popov via llvm-commits]

================
@@ -2822,11 +2822,17 @@ Instruction *InstCombinerImpl::visitGetElementPtrInst(GetElementPtrInst &GEP) {
   //    This has better support in BasicAA.
   //  - gep i32 p, mul(O, C) -> gep i8, p, mul(O, C*4) to fold the two
   //    multiplies together.
-  if (GEPEltType->isScalableTy() ||
-      (!GEPEltType->isIntegerTy(8) && GEP.getNumIndices() == 1 &&
-       match(GEP.getOperand(1),
-             m_OneUse(m_CombineOr(m_Mul(m_Value(), m_ConstantInt()),
-                                  m_Shl(m_Value(), m_ConstantInt())))))) {
+  //  - gep (gep @global, C1), %x, C2 is expanded so the two constants can
+  //    possibly be merged together.
+  if (!GEPEltType->isIntegerTy(8) &&
+      (GEPEltType->isScalableTy() ||
+       (GEP.getNumIndices() == 1 &&
+        match(GEP.getOperand(1),
+              m_OneUse(m_CombineOr(m_Mul(m_Value(), m_ConstantInt()),
+                                   m_Shl(m_Value(), m_ConstantInt()))))) ||
+       (isa<GEPOperator>(PtrOp) && isa<ConstantExpr>(PtrOp) &&
+        any_of(drop_begin(GEP.indices()),
+               [](Value *V) { return isa<Constant>(V); })))) {
----------------
nikic wrote:

I think this condition has reached the point where it wants to live in its own function :)

I'm also wondering whether zero constants should be excluded from this heuristic, as they don't add any actual offset. And then you wouldn't need the drop_begin? Or at least I don't fully get why the drop_begin() is there.

https://github.com/llvm/llvm-project/pull/110033