[InstCombine] Lower multi-dimensional GEP to ptradd

[usha1830]

This change will help canonicalize multi-dimensional array geps with exactly one variable index and all other indices constant into a flat, single-index gep over the element type. This would cover cases such as:

%gep = getelementptr [9 x [9 x [9 x i32]]], ptr @arr, i64 0, i64 %i, i64 2, i64 3

https://alive2.llvm.org/ce/z/SWRKZd

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: Usha Gupta (usha1830)

Changes

This change will help canonicalize multi-dimensional array geps with exactly one variable index and all other indices constant into a flat, single-index gep over the element type. This would cover cases such as:

%gep = getelementptr [9 x [9 x [9 x i32]]], ptr @<!-- -->arr, i64 0, i64 %i, i64 2, i64 3

---

Full diff: https://github.com/llvm/llvm-project/pull/150383.diff

1 Files Affected:

• (modified) llvm/lib/Transforms/InstCombine/InstructionCombining.cpp (+9)

diff --git a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
index e2a9255ca9c6e..9b148e523b7a7 100644
--- a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
@@ -2995,6 +2995,15 @@ static bool shouldCanonicalizeGEPToPtrAdd(GetElementPtrInst &GEP) {
                                  m_Shl(m_Value(), m_ConstantInt())))))
     return true;
 
+  // Flatten multidimensional GEPs with one variable index.
+  unsigned NumVarIndices = 0;
+  for (unsigned i = 1; i < GEP.getNumOperands(); ++i) {
+    if (!isa<ConstantInt>(GEP.getOperand(i)))
+      ++NumVarIndices;
+  }
+  if (NumVarIndices == 1)
+    return true;
+
   // gep (gep %p, C1), %x, C2 is expanded so the two constants can
   // possibly be merged together.
   auto PtrOpGep = dyn_cast<GEPOperator>(PtrOp);

[nikic]

I think this change is going to have a very big impact on other transforms and will require lots of mitigation before it's viable. Especially after #137297 this is going to convert all GEPs into ptradd representation.

[usha1830]

I think this change is going to have a very big impact on other transforms and will require lots of mitigation before it's viable. Especially after #137297 this is going to convert all GEPs into ptradd representation.

This patch helps in fixing the regression seen in this issue #143219.
And yes, I see a lot of test failures and am looking into addressing those.