[PATCH] D31782: [CodeGenPrepare]Skip sext promotion if operands for multiple users is detected

[Evgeny Astigeevich via Phabricator via llvm-commits]

eastig added inline comments.


================
Comment at: lib/CodeGen/CodeGenPrepare.cpp:4572-4573
     // Get the action to perform the promotion.
-    TypePromotionHelper::Action TPH =
-        TypePromotionHelper::getAction(I, InsertedInsts, *TLI, PromotedInsts);
+    TypePromotionHelper::Action TPH = TypePromotionHelper::getAction(
+        I, InsertedInsts, *TLI, PromotedInsts, hasMultiUserOperand);
     // Check if we can promote.
----------------
junbuml wrote:
> eastig wrote:
> > Do you assume that the Exts vector always has only one instruction? Otherwise the value of hasMultiUserOperand can be changed from true to false.
> I do not assume only one instruction in Exts. After hasMultiUserOperand is initialized as false in  optimizeExt(), hasMultiUserOperand can turn into true only when encountering any operand used by multiple users. I cannot see the case where hasMultiUserOperand turn from true to false ? Please let me know if you see such case. 
Yes, you are right. It can be changed only from false to true and never back.


================
Comment at: test/CodeGen/AArch64/arm64-addr-type-promotion.ll:99-100
+  %add = add nsw i32 %j, 1
+  %sext1 = sext i32 %add to i64
+  %arrayidx = getelementptr inbounds %struct.GlobalData, %struct.GlobalData* @global_data, i64 0, i32 0, i64 %sext1
+  store float %s, float* %arrayidx, align 4
----------------
junbuml wrote:
> eastig wrote:
> > Does this guarantee lowering to multi usage? Is it possible to write a test in MIR?
> In this test, I intended to make %add have two users:  1) in %sext1 and 2)  in %j.l.  In such case, we do not allow moving %sext up through %add. 
I see. Thank you for explanation.


https://reviews.llvm.org/D31782