[llvm] 2128fca - [InstCombine] Canonicalize `gep T* X, V / sizeof(T)` to `gep i8* X, V` (#76458)

via llvm-commits llvm-commits at lists.llvm.org
Thu Dec 28 19:30:04 PST 2023


Author: Yingwei Zheng
Date: 2023-12-29T11:30:00+08:00
New Revision: 2128fca6c1e3a0ba4d105f62ad0f6a841e992cfb

URL: https://github.com/llvm/llvm-project/commit/2128fca6c1e3a0ba4d105f62ad0f6a841e992cfb
DIFF: https://github.com/llvm/llvm-project/commit/2128fca6c1e3a0ba4d105f62ad0f6a841e992cfb.diff

LOG: [InstCombine] Canonicalize `gep T* X, V / sizeof(T)` to `gep i8* X, V` (#76458)

This patch canonicalize `gep T* X, V / sizeof(T)` to `gep i8* X, V`.
Alive2: https://alive2.llvm.org/ce/z/7XGjiB

As this pattern has been handled by the backends, the motivation of this
patch is to reduce the ref count of sdiv, which will enable more
optimizations.

Added: 
    

Modified: 
    llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
    llvm/test/Transforms/InstCombine/getelementptr.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
index 7f5a7b666903db..df393d72a85bf2 100644
--- a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
@@ -2469,31 +2469,31 @@ Instruction *InstCombinerImpl::visitGetElementPtrInst(GetElementPtrInst &GEP) {
         DL.getIndexSizeInBits(AS)) {
       uint64_t TyAllocSize = DL.getTypeAllocSize(GEPEltType).getFixedValue();
 
-      bool Matched = false;
-      uint64_t C;
-      Value *V = nullptr;
       if (TyAllocSize == 1) {
-        V = GEP.getOperand(1);
-        Matched = true;
-      } else if (match(GEP.getOperand(1),
-                       m_AShr(m_Value(V), m_ConstantInt(C)))) {
-        if (TyAllocSize == 1ULL << C)
-          Matched = true;
-      } else if (match(GEP.getOperand(1),
-                       m_SDiv(m_Value(V), m_ConstantInt(C)))) {
-        if (TyAllocSize == C)
-          Matched = true;
+        // Canonicalize (gep i8* X, (ptrtoint Y)-(ptrtoint X)) to (bitcast Y),
+        // but only if both point to the same underlying object (otherwise
+        // provenance is not necessarily retained).
+        Value *X = GEP.getPointerOperand();
+        Value *Y;
+        if (match(GEP.getOperand(1),
+                  m_Sub(m_PtrToInt(m_Value(Y)), m_PtrToInt(m_Specific(X)))) &&
+            getUnderlyingObject(X) == getUnderlyingObject(Y))
+          return CastInst::CreatePointerBitCastOrAddrSpaceCast(Y, GEPType);
+      } else {
+        // Canonicalize (gep T* X, V / sizeof(T)) to (gep i8* X, V)
+        Value *V;
+        if ((has_single_bit(TyAllocSize) &&
+             match(GEP.getOperand(1),
+                   m_Exact(m_AShr(m_Value(V),
+                                  m_SpecificInt(countr_zero(TyAllocSize)))))) ||
+            match(GEP.getOperand(1),
+                  m_Exact(m_SDiv(m_Value(V), m_SpecificInt(TyAllocSize))))) {
+          GetElementPtrInst *NewGEP = GetElementPtrInst::Create(
+              Builder.getInt8Ty(), GEP.getPointerOperand(), V);
+          NewGEP->setIsInBounds(GEP.isInBounds());
+          return NewGEP;
+        }
       }
-
-      // Canonicalize (gep i8* X, (ptrtoint Y)-(ptrtoint X)) to (bitcast Y), but
-      // only if both point to the same underlying object (otherwise provenance
-      // is not necessarily retained).
-      Value *Y;
-      Value *X = GEP.getOperand(0);
-      if (Matched &&
-          match(V, m_Sub(m_PtrToInt(m_Value(Y)), m_PtrToInt(m_Specific(X)))) &&
-          getUnderlyingObject(X) == getUnderlyingObject(Y))
-        return CastInst::CreatePointerBitCastOrAddrSpaceCast(Y, GEPType);
     }
   }
   // We do not handle pointer-vector geps here.

diff  --git a/llvm/test/Transforms/InstCombine/getelementptr.ll b/llvm/test/Transforms/InstCombine/getelementptr.ll
index bc7fdc9352df6c..7d67f2583aa24d 100644
--- a/llvm/test/Transforms/InstCombine/getelementptr.ll
+++ b/llvm/test/Transforms/InstCombine/getelementptr.ll
@@ -1453,4 +1453,88 @@ define ptr @const_gep_chain(ptr %p, i64 %a) {
   ret ptr %p4
 }
 
+define ptr @gep_sdiv(ptr %p, i64 %off) {
+; CHECK-LABEL: @gep_sdiv(
+; CHECK-NEXT:    [[PTR:%.*]] = getelementptr i8, ptr [[P:%.*]], i64 [[OFF:%.*]]
+; CHECK-NEXT:    ret ptr [[PTR]]
+;
+  %index = sdiv exact i64 %off, 7
+  %ptr = getelementptr %struct.C, ptr %p, i64 %index
+  ret ptr %ptr
+}
+
+define <2 x ptr> @gep_sdiv_vec(<2 x ptr> %p, <2 x i64> %off) {
+; CHECK-LABEL: @gep_sdiv_vec(
+; CHECK-NEXT:    [[PTR:%.*]] = getelementptr i8, <2 x ptr> [[P:%.*]], <2 x i64> [[OFF:%.*]]
+; CHECK-NEXT:    ret <2 x ptr> [[PTR]]
+;
+  %index = sdiv exact <2 x i64> %off, <i64 7, i64 7>
+  %ptr = getelementptr %struct.C, <2 x ptr> %p, <2 x i64> %index
+  ret <2 x ptr> %ptr
+}
+
+define ptr @gep_sdiv_inbounds(ptr %p, i64 %off) {
+; CHECK-LABEL: @gep_sdiv_inbounds(
+; CHECK-NEXT:    [[PTR:%.*]] = getelementptr inbounds i8, ptr [[P:%.*]], i64 [[OFF:%.*]]
+; CHECK-NEXT:    ret ptr [[PTR]]
+;
+  %index = sdiv exact i64 %off, 7
+  %ptr = getelementptr inbounds %struct.C, ptr %p, i64 %index
+  ret ptr %ptr
+}
+
+define ptr @gep_ashr(ptr %p, i64 %off) {
+; CHECK-LABEL: @gep_ashr(
+; CHECK-NEXT:    [[PTR:%.*]] = getelementptr i8, ptr [[P:%.*]], i64 [[OFF:%.*]]
+; CHECK-NEXT:    ret ptr [[PTR]]
+;
+  %index = ashr exact i64 %off, 2
+  %ptr = getelementptr i32, ptr %p, i64 %index
+  ret ptr %ptr
+}
+
+; Negative tests
+
+define ptr @gep_i8(ptr %p, i64 %off) {
+; CHECK-LABEL: @gep_i8(
+; CHECK-NEXT:    [[PTR:%.*]] = getelementptr i8, ptr [[P:%.*]], i64 [[OFF:%.*]]
+; CHECK-NEXT:    ret ptr [[PTR]]
+;
+  %ptr = getelementptr i8, ptr %p, i64 %off
+  ret ptr %ptr
+}
+
+define ptr @gep_sdiv_mismatched_size(ptr %p, i64 %off) {
+; CHECK-LABEL: @gep_sdiv_mismatched_size(
+; CHECK-NEXT:    [[INDEX:%.*]] = sdiv exact i64 [[OFF:%.*]], 20
+; CHECK-NEXT:    [[PTR:%.*]] = getelementptr [[STRUCT_C:%.*]], ptr [[P:%.*]], i64 [[INDEX]]
+; CHECK-NEXT:    ret ptr [[PTR]]
+;
+  %index = sdiv exact i64 %off, 20
+  %ptr = getelementptr %struct.C, ptr %p, i64 %index
+  ret ptr %ptr
+}
+
+define ptr @gep_sdiv_without_exact(ptr %p, i64 %off) {
+; CHECK-LABEL: @gep_sdiv_without_exact(
+; CHECK-NEXT:    [[INDEX:%.*]] = sdiv i64 [[OFF:%.*]], 7
+; CHECK-NEXT:    [[PTR:%.*]] = getelementptr [[STRUCT_C:%.*]], ptr [[P:%.*]], i64 [[INDEX]]
+; CHECK-NEXT:    ret ptr [[PTR]]
+;
+  %index = sdiv i64 %off, 7
+  %ptr = getelementptr %struct.C, ptr %p, i64 %index
+  ret ptr %ptr
+}
+
+define ptr @gep_ashr_without_exact(ptr %p, i64 %off) {
+; CHECK-LABEL: @gep_ashr_without_exact(
+; CHECK-NEXT:    [[INDEX:%.*]] = ashr i64 [[OFF:%.*]], 2
+; CHECK-NEXT:    [[PTR:%.*]] = getelementptr i32, ptr [[P:%.*]], i64 [[INDEX]]
+; CHECK-NEXT:    ret ptr [[PTR]]
+;
+  %index = ashr i64 %off, 2
+  %ptr = getelementptr i32, ptr %p, i64 %index
+  ret ptr %ptr
+}
+
 !0 = !{!"branch_weights", i32 2, i32 10}


        


More information about the llvm-commits mailing list