[llvm] [InstCombine] Extend ADD+GEP->GEP+GEP combine to disjoint or. (PR #76981)

[Yingwei Zheng via llvm-commits]

dtcxzyw wrote:

Yeah, I see some similar changes in `bench/ocio/optimized/ImagePacking.cpp.ll`:
```
diff --git a/bench/ocio/optimized/ImagePacking.cpp.ll b/bench/ocio/optimized/ImagePacking.cpp.ll
index 490d6aec..b632e889 100644
--- a/bench/ocio/optimized/ImagePacking.cpp.ll
+++ b/bench/ocio/optimized/ImagePacking.cpp.ll
@@ -116,12 +116,15 @@ while.body.preheader:                             ; preds = %if.end7
   %add.ptr = getelementptr inbounds i8, ptr %9, i64 %mul8
   %add.ptr14 = getelementptr inbounds i8, ptr %add.ptr, i64 %mul13
   %wide.trip.count = zext nneg i32 %outputBufferSize to i64
-  %spec.select62 = select i1 %tobool.not, ptr null, ptr %add.ptr24
+  %spec.select66 = select i1 %tobool.not, ptr null, ptr %add.ptr24
+  %invariant.gep = getelementptr i8, ptr %inBitDepthBuffer, i64 1
+  %invariant.gep62 = getelementptr i8, ptr %inBitDepthBuffer, i64 2
+  %invariant.gep64 = getelementptr i8, ptr %inBitDepthBuffer, i64 3
   br label %while.body
 
 while.body:                                       ; preds = %while.body.preheader, %cond.end
   %indvars.iv = phi i64 [ %indvars.iv.next, %cond.end ], [ 0, %while.body.preheader ]
-  %aPtr.154 = phi ptr [ %spec.select, %cond.end ], [ %spec.select62, %while.body.preheader ]
+  %aPtr.154 = phi ptr [ %spec.select, %cond.end ], [ %spec.select66, %while.body.preheader ]
   %bPtr.053 = phi ptr [ %add.ptr44, %cond.end ], [ %add.ptr18, %while.body.preheader ]
   %gPtr.052 = phi ptr [ %add.ptr43, %cond.end ], [ %add.ptr16, %while.body.preheader ]
   %rPtr.051 = phi ptr [ %add.ptr42, %cond.end ], [ %add.ptr14, %while.body.preheader ]
@@ -130,25 +133,22 @@ while.body:                                       ; preds = %while.body.preheade
   %arrayidx = getelementptr inbounds i8, ptr %inBitDepthBuffer, i64 %11
   store i8 %10, ptr %arrayidx, align 1
   %12 = load i8, ptr %gPtr.052, align 1
-  %13 = or disjoint i64 %11, 1
-  %arrayidx31 = getelementptr inbounds i8, ptr %inBitDepthBuffer, i64 %13
-  store i8 %12, ptr %arrayidx31, align 1
-  %14 = load i8, ptr %bPtr.053, align 1
-  %15 = or disjoint i64 %11, 2
-  %arrayidx35 = getelementptr inbounds i8, ptr %inBitDepthBuffer, i64 %15
-  store i8 %14, ptr %arrayidx35, align 1
+  %gep = getelementptr i8, ptr %invariant.gep, i64 %11
+  store i8 %12, ptr %gep, align 1
+  %13 = load i8, ptr %bPtr.053, align 1
+  %gep63 = getelementptr i8, ptr %invariant.gep62, i64 %11
+  store i8 %13, ptr %gep63, align 1
   %tobool36.not = icmp eq ptr %aPtr.154, null
   br i1 %tobool36.not, label %cond.end, label %cond.true
 
 cond.true:                                        ; preds = %while.body
-  %16 = load i8, ptr %aPtr.154, align 1
+  %14 = load i8, ptr %aPtr.154, align 1
   br label %cond.end
 
 cond.end:                                         ; preds = %while.body, %cond.true
-  %cond = phi i8 [ %16, %cond.true ], [ 0, %while.body ]
-  %17 = or disjoint i64 %11, 3
-  %arrayidx40 = getelementptr inbounds i8, ptr %inBitDepthBuffer, i64 %17
-  store i8 %cond, ptr %arrayidx40, align 1
+  %cond = phi i8 [ %14, %cond.true ], [ 0, %while.body ]
+  %gep65 = getelementptr i8, ptr %invariant.gep64, i64 %11
+  store i8 %cond, ptr %gep65, align 1
   %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
   %add.ptr42 = getelementptr inbounds i8, ptr %rPtr.051, i64 %4
   %add.ptr43 = getelementptr inbounds i8, ptr %gPtr.052, i64 %4
```

@nikic Did you mean it will cause register allocator to introduce more spills?


https://github.com/llvm/llvm-project/pull/76981