[llvm] [InstCombine] Canonicalize more geps with constant gep bases and constant offsets. (PR #110033)
David Green via llvm-commits
llvm-commits at lists.llvm.org
Sun Oct 6 02:43:36 PDT 2024
https://github.com/davemgreen updated https://github.com/llvm/llvm-project/pull/110033
>From 149efb8681ee98a86241383fb5fcdd1dc1706228 Mon Sep 17 00:00:00 2001
From: David Green <david.green at arm.com>
Date: Sun, 6 Oct 2024 10:43:25 +0100
Subject: [PATCH] [InstCombine] Canonicalize more geps with constant gep bases
and constant offsets.
This is another small but hopefully not performance negative step to
canonicalizing towards i8 geps. We looks for geps with a constant offset base
pointer of the form `gep (gep @glob, C1), x, C2` and expand the gep
instruction, so that the constant can hopefully be combined together (or the
offset can be computed in common).
---
.../InstCombine/InstructionCombining.cpp | 46 ++++++++++++++-----
.../InstCombine/canonicalize-gep-constglob.ll | 14 ++++--
.../InstCombine/gep-merge-constant-indices.ll | 13 +++---
3 files changed, 50 insertions(+), 23 deletions(-)
diff --git a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
index 1ea3efb27937d2..1f4a6f793404cf 100644
--- a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
@@ -2731,6 +2731,36 @@ Value *InstCombiner::getFreelyInvertedImpl(Value *V, bool WillInvertAllUses,
return nullptr;
}
+/// Return true if we should canonicalize the gep to an i8 ptradd.
+static bool shouldCanonicalizeGEPToPtrAdd(GetElementPtrInst &GEP) {
+ Value *PtrOp = GEP.getOperand(0);
+ Type *GEPEltType = GEP.getSourceElementType();
+ if (GEPEltType->isIntegerTy(8))
+ return false;
+
+ // Canonicalize scalable GEPs to an explicit offset using the llvm.vscale
+ // intrinsic. This has better support in BasicAA.
+ if (GEPEltType->isScalableTy())
+ return true;
+
+ // gep i32 p, mul(O, C) -> gep i8, p, mul(O, C*4) to fold the two multiplies
+ // together.
+ if (GEP.getNumIndices() == 1 &&
+ match(GEP.getOperand(1),
+ m_OneUse(m_CombineOr(m_Mul(m_Value(), m_ConstantInt()),
+ m_Shl(m_Value(), m_ConstantInt())))))
+ 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);
+ return PtrOpGep && PtrOpGep->hasAllConstantIndices() &&
+ any_of(GEP.indices(), [](Value *V) {
+ const APInt *C;
+ return match(V, m_APInt(C)) && !C->isZero();
+ });
+}
+
Instruction *InstCombinerImpl::visitGetElementPtrInst(GetElementPtrInst &GEP) {
Value *PtrOp = GEP.getOperand(0);
SmallVector<Value *, 8> Indices(GEP.indices());
@@ -2812,19 +2842,11 @@ Instruction *InstCombinerImpl::visitGetElementPtrInst(GetElementPtrInst &GEP) {
GEP.getNoWrapFlags()));
}
- // Canonicalize
- // - scalable GEPs to an explicit offset using the llvm.vscale intrinsic.
- // 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())))))) {
+ if (shouldCanonicalizeGEPToPtrAdd(GEP)) {
Value *Offset = EmitGEPOffset(cast<GEPOperator>(&GEP));
- return replaceInstUsesWith(
- GEP, Builder.CreatePtrAdd(PtrOp, Offset, "", GEP.getNoWrapFlags()));
+ Value *NewGEP =
+ Builder.CreatePtrAdd(PtrOp, Offset, "", GEP.getNoWrapFlags());
+ return replaceInstUsesWith(GEP, NewGEP);
}
// Check to see if the inputs to the PHI node are getelementptr instructions.
diff --git a/llvm/test/Transforms/InstCombine/canonicalize-gep-constglob.ll b/llvm/test/Transforms/InstCombine/canonicalize-gep-constglob.ll
index f1f01a1884a22e..53585dfde48b22 100644
--- a/llvm/test/Transforms/InstCombine/canonicalize-gep-constglob.ll
+++ b/llvm/test/Transforms/InstCombine/canonicalize-gep-constglob.ll
@@ -7,7 +7,8 @@ define ptr @x12(i64 %x) {
; CHECK-LABEL: define ptr @x12(
; CHECK-SAME: i64 [[X:%.*]]) {
; CHECK-NEXT: entry:
-; CHECK-NEXT: [[GEP:%.*]] = getelementptr inbounds [10 x [10 x [10 x i32]]], ptr getelementptr inbounds (i8, ptr @glob, i64 36), i64 0, i64 [[X]], i64 1, i64 2
+; CHECK-NEXT: [[GEP_IDX:%.*]] = mul nsw i64 [[X]], 400
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr i8, ptr getelementptr inbounds (i8, ptr @glob, i64 84), i64 [[GEP_IDX]]
; CHECK-NEXT: ret ptr [[GEP]]
;
entry:
@@ -19,7 +20,10 @@ define ptr @x1y(i64 %x, i64 %y) {
; CHECK-LABEL: define ptr @x1y(
; CHECK-SAME: i64 [[X:%.*]], i64 [[Y:%.*]]) {
; CHECK-NEXT: entry:
-; CHECK-NEXT: [[GEP:%.*]] = getelementptr inbounds [10 x [10 x [10 x i32]]], ptr getelementptr inbounds (i8, ptr @glob, i64 36), i64 0, i64 [[X]], i64 2, i64 [[Y]]
+; CHECK-NEXT: [[GEP_IDX:%.*]] = mul nsw i64 [[X]], 400
+; CHECK-NEXT: [[GEP_IDX1:%.*]] = shl nsw i64 [[Y]], 2
+; CHECK-NEXT: [[TMP0:%.*]] = getelementptr i8, ptr getelementptr inbounds (i8, ptr @glob, i64 116), i64 [[GEP_IDX]]
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr i8, ptr [[TMP0]], i64 [[GEP_IDX1]]
; CHECK-NEXT: ret ptr [[GEP]]
;
entry:
@@ -55,8 +59,10 @@ define i32 @twoloads(i64 %x) {
; CHECK-LABEL: define i32 @twoloads(
; CHECK-SAME: i64 [[X:%.*]]) {
; CHECK-NEXT: entry:
-; CHECK-NEXT: [[GEP1:%.*]] = getelementptr inbounds [10 x [10 x [10 x i32]]], ptr getelementptr inbounds (i8, ptr @glob, i64 50), i64 0, i64 [[X]], i64 2, i64 1
-; CHECK-NEXT: [[GEP2:%.*]] = getelementptr inbounds [10 x [10 x [10 x i32]]], ptr getelementptr inbounds (i8, ptr @glob, i64 36), i64 0, i64 [[X]], i64 2, i64 4
+; CHECK-NEXT: [[GEP1_IDX:%.*]] = mul nsw i64 [[X]], 400
+; CHECK-NEXT: [[GEP1:%.*]] = getelementptr i8, ptr getelementptr inbounds (i8, ptr @glob, i64 134), i64 [[GEP1_IDX]]
+; CHECK-NEXT: [[GEP2_IDX:%.*]] = mul nsw i64 [[X]], 400
+; CHECK-NEXT: [[GEP2:%.*]] = getelementptr i8, ptr getelementptr inbounds (i8, ptr @glob, i64 132), i64 [[GEP2_IDX]]
; CHECK-NEXT: [[A:%.*]] = load i32, ptr [[GEP1]], align 4
; CHECK-NEXT: [[B:%.*]] = load i32, ptr [[GEP2]], align 4
; CHECK-NEXT: [[C:%.*]] = add i32 [[A]], [[B]]
diff --git a/llvm/test/Transforms/InstCombine/gep-merge-constant-indices.ll b/llvm/test/Transforms/InstCombine/gep-merge-constant-indices.ll
index 509d19c8c00c7e..edec54495c1ae6 100644
--- a/llvm/test/Transforms/InstCombine/gep-merge-constant-indices.ll
+++ b/llvm/test/Transforms/InstCombine/gep-merge-constant-indices.ll
@@ -140,12 +140,11 @@ define ptr @notDivisible(ptr %p) {
ret ptr %2
}
-; Negative test. Two GEP should not be merged if not both offsets are constant
-; or divisible by the other's size.
define ptr @partialConstant2(ptr %p, i64 %a) {
; CHECK-LABEL: @partialConstant2(
-; CHECK-NEXT: [[TMP1:%.*]] = getelementptr inbounds i8, ptr [[P:%.*]], i64 4
-; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds [4 x i64], ptr [[TMP1]], i64 [[A:%.*]], i64 2
+; CHECK-NEXT: [[DOTIDX:%.*]] = shl nsw i64 [[A:%.*]], 5
+; CHECK-NEXT: [[TMP1:%.*]] = getelementptr i8, ptr [[P:%.*]], i64 20
+; CHECK-NEXT: [[TMP2:%.*]] = getelementptr i8, ptr [[TMP1]], i64 [[DOTIDX]]
; CHECK-NEXT: ret ptr [[TMP2]]
;
%1 = getelementptr inbounds i32, ptr %p, i64 1
@@ -153,13 +152,13 @@ define ptr @partialConstant2(ptr %p, i64 %a) {
ret ptr %2
}
-; Negative test. Two GEP should not be merged if there is another use of the
-; first GEP by the second GEP.
define ptr @partialConstant3(ptr %p) {
; CHECK-LABEL: @partialConstant3(
; CHECK-NEXT: [[TMP1:%.*]] = getelementptr inbounds i8, ptr [[P:%.*]], i64 4
; CHECK-NEXT: [[TMP2:%.*]] = ptrtoint ptr [[TMP1]] to i64
-; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds [4 x i64], ptr [[TMP1]], i64 [[TMP2]], i64 2
+; CHECK-NEXT: [[DOTIDX:%.*]] = shl nsw i64 [[TMP2]], 5
+; CHECK-NEXT: [[DOTOFFS:%.*]] = or disjoint i64 [[DOTIDX]], 16
+; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds i8, ptr [[TMP1]], i64 [[DOTOFFS]]
; CHECK-NEXT: ret ptr [[TMP3]]
;
%1 = getelementptr inbounds i32, ptr %p, i64 1
More information about the llvm-commits
mailing list