[AArch64]Redundant store instructions should be removed as dead code

[David.Xu]

Summary

If there is a store followed by a store with the same value to the same location, then the store is dead/noop. It can be removed.

This problem is found in spec2006-197.parser.

For example,
stur w10, [x11, #-4]
stur w10, [x11, #-4]
Then one of the two stur instructions can be removed.

[jmolloy]

Hi David,

Thanks for working on this. This patch looks good to me, thanks!

Cheers,

James

[mcrosier]

LGTM, with a few comments.

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

9797

What if the 2nd store is a larger value type? Wouldn't the first store still be dead?

(Just a comment that, if true, doesn't need to be addresses in this patch.)

9800

Just move this line up to maximize horizontal space and reduce vertical space (i.e., "ST1->isunindexed() && !ST1->isVolatie() &&").

9801

"between the two stores", perhaps.

[David.Xu]

Line 9800 can't be moved up to maximize horizontal space and reduce vertical space, because clang-formate does not allow this. When it is moved up, it will be moved down after executing "git clang-formate" command.

[jmolloy]

Hi,

As David has now finished his internship at ARM, I took the liberty of committing this after updating it for review feedback.

• Vertical space maximised as requested.
• reachesChainWithoutSideEffects() is not required, as it is guaranteed that the two chains are both stores with no intervening chains.

Committed in r218569.

Cheers,

James

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

       return Chain;
     }
   }
+  // If this is a store followed by a store with the same value to the same
+  // location, then the store is dead/noop.
+  if (StoreSDNode *ST1 = dyn_cast<StoreSDNode>(Chain)) {
+    if (ST1->getBasePtr() == Ptr && ST->getMemoryVT() == ST1->getMemoryVT() &&
+        ST1->getValue() == Value && ST->isUnindexed() && !ST->isVolatile() &&
+        ST1->isUnindexed() &&
+        !ST1->isVolatile() &&
+        // There can't be any side effects between the store and store, such as
+        // a call or store.
+        Chain.reachesChainWithoutSideEffects(SDValue(ST1, 0))) {
+      // The store is dead, remove it.
+      return Chain;
+    }
+  }
 
   // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
   // truncating store.  We can do this even if this is already a truncstore.

test/CodeGen/AArch64/Redundantstore.ll

+;llc -O3 -march=aarch64 < %s | FileCheck %s 
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+@end_of_array = common global i8* null, align 8
+; CHECK-LABLE: @test
+; CHECK: stur
+; CHECK-NOT: stur
+define i8* @test(i32 %size) {
+entry:
+  %0 = load i8** @end_of_array, align 8
+  %conv = sext i32 %size to i64
+  %and = and i64 %conv, -8
+  %conv2 = trunc i64 %and to i32
+  %add.ptr.sum = add nsw i64 %and, -4
+  %add.ptr3 = getelementptr inbounds i8* %0, i64 %add.ptr.sum
+  %size4 = bitcast i8* %add.ptr3 to i32*
+  store i32 %conv2, i32* %size4, align 4
+  %add.ptr.sum9 = add nsw i64 %and, -4
+  %add.ptr5 = getelementptr inbounds i8* %0, i64 %add.ptr.sum9
+  %size6 = bitcast i8* %add.ptr5 to i32*
+  store i32 %conv2, i32* %size6, align 4
+  ret i8* %0
+}
+