[GVN] Execute performLoopLoadPRE ahead of PerformLoadPRE

[jaykang10]

Summary

There are cfgs which both of loop load pre and load pre can handle.

In this case, it could be better to perform loop load pre first because it can reduce the number of execution of load or it can help sccp pass propagate constant.

It fixes https://bugs.llvm.org/show_bug.cgi?id=51141

[nikic]

Could you please pre-commit the test, so the difference becomes visible?

[jaykang10]

Updated test output with pre-committed test.

[jaykang10]

In D108204#2955600, @nikic wrote:

Could you please pre-commit the test, so the difference becomes visible?

Yep, done. Thanks for comment!

[xbolva00]

@reames ?

[nikic]

LGTM

[jaykang10]

Rebased

[jaykang10]

@nikic Thanks for updating code. Let me push this change.

[fhahn]

It looks like this may cause a regression where we fail to remove some dead code https://github.com/llvm/llvm-project/issues/53322

@jaykang10 it would be great if you could take a look

llvm/lib/Transforms/Scalar/GVN.cpp

bool GVN::processNonLocalLoad(LoadInst *Load) {
   }
 
   // Step 4: Eliminate partial redundancy.
   if (!isPREEnabled() || !isLoadPREEnabled())
     return Changed;
   if (!isLoadInLoopPREEnabled() && LI && LI->getLoopFor(Load->getParent()))
     return Changed;
 
-  if (PerformLoadPRE(Load, ValuesPerBlock, UnavailableBlocks) ||
-      performLoopLoadPRE(Load, ValuesPerBlock, UnavailableBlocks))
+  if (performLoopLoadPRE(Load, ValuesPerBlock, UnavailableBlocks) ||
+      PerformLoadPRE(Load, ValuesPerBlock, UnavailableBlocks))
     return true;
 
   return Changed;
 }
 
 static bool impliesEquivalanceIfTrue(CmpInst* Cmp) {
   if (Cmp->getPredicate() == CmpInst::Predicate::ICMP_EQ)
     return true;

llvm/test/Transforms/GVN/gvn-loop-load-pre-order.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt < %s -loops -gvn -S | FileCheck -check-prefix=CHECK-PRE %s
 ; RUN: opt < %s -loops -gvn -sccp -S | FileCheck -check-prefix=CHECK-SCCP %s
 
 @b = external global i32, align 4
 
 define void @main(i1 %cond) {
 ; CHECK-PRE-LABEL: @main(
 ; CHECK-PRE-NEXT:  entry:
 ; CHECK-PRE-NEXT:    store i32 0, i32* @b, align 4
 ; CHECK-PRE-NEXT:    br label [[WHILE_BODY3:%.*]]
 ; CHECK-PRE:       while.cond.loopexit.loopexit:
 ; CHECK-PRE-NEXT:    ret void
 ; CHECK-PRE:       while.body3:
-; CHECK-PRE-NEXT:    [[TMP0:%.*]] = phi i32 [ [[DOTPRE:%.*]], [[IF_END_WHILE_BODY3_CRIT_EDGE:%.*]] ], [ 0, [[ENTRY:%.*]] ]
+; CHECK-PRE-NEXT:    [[TMP0:%.*]] = phi i32 [ [[TMP1:%.*]], [[IF_END:%.*]] ], [ 0, [[ENTRY:%.*]] ]
 ; CHECK-PRE-NEXT:    [[TOBOOL4_NOT:%.*]] = icmp eq i32 [[TMP0]], 0
-; CHECK-PRE-NEXT:    br i1 [[TOBOOL4_NOT]], label [[IF_END:%.*]], label [[IF_THEN:%.*]]
+; CHECK-PRE-NEXT:    br i1 [[TOBOOL4_NOT]], label [[IF_END]], label [[IF_THEN:%.*]]
 ; CHECK-PRE:       if.then:
 ; CHECK-PRE-NEXT:    tail call void @foo()
+; CHECK-PRE-NEXT:    [[DOTPRE:%.*]] = load i32, i32* @b, align 4
 ; CHECK-PRE-NEXT:    br label [[IF_END]]
 ; CHECK-PRE:       if.end:
-; CHECK-PRE-NEXT:    br i1 [[COND:%.*]], label [[WHILE_COND_LOOPEXIT_LOOPEXIT:%.*]], label [[IF_END_WHILE_BODY3_CRIT_EDGE]]
-; CHECK-PRE:       if.end.while.body3_crit_edge:
-; CHECK-PRE-NEXT:    [[DOTPRE]] = load i32, i32* @b, align 4
-; CHECK-PRE-NEXT:    br label [[WHILE_BODY3]]
+; CHECK-PRE-NEXT:    [[TMP1]] = phi i32 [ [[DOTPRE]], [[IF_THEN]] ], [ 0, [[WHILE_BODY3]] ]
+; CHECK-PRE-NEXT:    br i1 [[COND:%.*]], label [[WHILE_COND_LOOPEXIT_LOOPEXIT:%.*]], label [[WHILE_BODY3]]
 ;
 ; CHECK-SCCP-LABEL: @main(
 ; CHECK-SCCP-NEXT:  entry:
 ; CHECK-SCCP-NEXT:    store i32 0, i32* @b, align 4
 ; CHECK-SCCP-NEXT:    br label [[WHILE_BODY3:%.*]]
 ; CHECK-SCCP:       while.cond.loopexit.loopexit:
 ; CHECK-SCCP-NEXT:    ret void
 ; CHECK-SCCP:       while.body3:
-; CHECK-SCCP-NEXT:    [[TMP0:%.*]] = phi i32 [ [[DOTPRE:%.*]], [[IF_END_WHILE_BODY3_CRIT_EDGE:%.*]] ], [ 0, [[ENTRY:%.*]] ]
-; CHECK-SCCP-NEXT:    [[TOBOOL4_NOT:%.*]] = icmp eq i32 [[TMP0]], 0
-; CHECK-SCCP-NEXT:    br i1 [[TOBOOL4_NOT]], label [[IF_END:%.*]], label [[IF_THEN:%.*]]
+; CHECK-SCCP-NEXT:    br i1 true, label [[IF_END:%.*]], label [[IF_THEN:%.*]]
 ; CHECK-SCCP:       if.then:
-; CHECK-SCCP-NEXT:    tail call void @foo()
 ; CHECK-SCCP-NEXT:    br label [[IF_END]]
 ; CHECK-SCCP:       if.end:
-; CHECK-SCCP-NEXT:    br i1 [[COND:%.*]], label [[WHILE_COND_LOOPEXIT_LOOPEXIT:%.*]], label [[IF_END_WHILE_BODY3_CRIT_EDGE]]
-; CHECK-SCCP:       if.end.while.body3_crit_edge:
-; CHECK-SCCP-NEXT:    [[DOTPRE]] = load i32, i32* @b, align 4
-; CHECK-SCCP-NEXT:    br label [[WHILE_BODY3]]
+; CHECK-SCCP-NEXT:    br i1 [[COND:%.*]], label [[WHILE_COND_LOOPEXIT_LOOPEXIT:%.*]], label [[WHILE_BODY3]]
 ;
 entry:
   store i32 0, i32* @b
   br label %while.body3.lr.ph
 
 while.cond.loopexit.loopexit:                     ; preds = %if.end
   ret void