[llvm] ddbc728 - [PGO][PGSO] Update BFI in CodeGenPrepare::optimizeSelectInst.

[Hiroshi Yamauchi via llvm-commits]

Author: Hiroshi Yamauchi
Date: 2020-01-22T08:36:54-08:00
New Revision: ddbc728828c70728473b47c9f7427aa9514f3d17

URL: https://github.com/llvm/llvm-project/commit/ddbc728828c70728473b47c9f7427aa9514f3d17
DIFF: https://github.com/llvm/llvm-project/commit/ddbc728828c70728473b47c9f7427aa9514f3d17.diff

LOG: [PGO][PGSO] Update BFI in CodeGenPrepare::optimizeSelectInst.

Summary:
Without the BFI update, some hot blocks are incorrectly treated as cold code.

This fixes a FDO perf regression in the TSVC benchmark from D71288.

Reviewers: davidxl

Subscribers: hiraditya, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D73146

Added: 
    

Modified: 
    llvm/lib/CodeGen/CodeGenPrepare.cpp
    llvm/test/CodeGen/X86/cmov-into-branch.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/CodeGen/CodeGenPrepare.cpp b/llvm/lib/CodeGen/CodeGenPrepare.cpp
index 9639336da726..a44d965029dc 100644
--- a/llvm/lib/CodeGen/CodeGenPrepare.cpp
+++ b/llvm/lib/CodeGen/CodeGenPrepare.cpp
@@ -6138,6 +6138,7 @@ bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) {
   BasicBlock *StartBlock = SI->getParent();
   BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(LastSI));
   BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
+  BFI->setBlockFreq(EndBlock, BFI->getBlockFreq(StartBlock).getFrequency());
 
   // Delete the unconditional branch that was just created by the split.
   StartBlock->getTerminator()->eraseFromParent();

diff  --git a/llvm/test/CodeGen/X86/cmov-into-branch.ll b/llvm/test/CodeGen/X86/cmov-into-branch.ll
index 6cc81ac0833b..d9ab2550f791 100644
--- a/llvm/test/CodeGen/X86/cmov-into-branch.ll
+++ b/llvm/test/CodeGen/X86/cmov-into-branch.ll
@@ -165,6 +165,30 @@ define i32 @weighted_select_pgso(i32 %a, i32 %b) !prof !14 {
   ret i32 %sel
 }
 
+; If two selects in a row are predictable, turn them into branches.
+define i32 @weighted_selects(i32 %a, i32 %b) !prof !19 {
+; CHECK-LABEL: weighted_selects:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    movl %esi, %eax
+; CHECK-NEXT:    testl %edi, %edi
+; CHECK-NEXT:    movl %edi, %ecx
+; CHECK-NEXT:    jne .LBB11_2
+; CHECK-NEXT:  # %bb.1: # %select.false
+; CHECK-NEXT:    movl %eax, %ecx
+; CHECK-NEXT:  .LBB11_2: # %select.end
+; CHECK-NEXT:    testl %ecx, %ecx
+; CHECK-NEXT:    jne .LBB11_4
+; CHECK-NEXT:  # %bb.3: # %select.false2
+; CHECK-NEXT:    movl %edi, %eax
+; CHECK-NEXT:  .LBB11_4: # %select.end1
+; CHECK-NEXT:    retq
+  %cmp = icmp ne i32 %a, 0
+  %sel = select i1 %cmp, i32 %a, i32 %b, !prof !16
+  %cmp1 = icmp ne i32 %sel, 0
+  %sel1 = select i1 %cmp1, i32 %b, i32 %a, !prof !16
+  ret i32 %sel1
+}
+
 !llvm.module.flags = !{!0}
 !0 = !{i32 1, !"ProfileSummary", !1}
 !1 = !{!2, !3, !4, !5, !6, !7, !8, !9}
@@ -185,3 +209,4 @@ define i32 @weighted_select_pgso(i32 %a, i32 %b) !prof !14 {
 !16 = !{!"branch_weights", i32 1, i32 100}
 !17 = !{!"branch_weights", i32 100, i32 1}
 !18 = !{!"branch_weights", i32 0, i32 0}
+!19 = !{!"function_entry_count", i64 100}