[PATCH] D119777: [X86] Introduce x86-cmov-converter-force-all

[Amir Ayupov via Phabricator via llvm-commits]

Amir created this revision.
Herald added subscribers: pengfei, hiraditya.
Amir requested review of this revision.
Herald added a project: LLVM.
Herald added a subscriber: llvm-commits.

Introduce an option to expand all CMOV groups into hammocks, matching GCC's
`-fno-if-conversion2` flag. The motivation is to leave CMOV conversion
opportunities to a binary optimizer that can make the decision based on branch
misprediction rate (available e.g. in Intel's LBR).


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D119777

Files:
  llvm/lib/Target/X86/X86CmovConversion.cpp


Index: llvm/lib/Target/X86/X86CmovConversion.cpp
===================================================================
--- llvm/lib/Target/X86/X86CmovConversion.cpp
+++ llvm/lib/Target/X86/X86CmovConversion.cpp
@@ -97,6 +97,11 @@
     cl::desc("Convert cmovs to branches whenever they have memory operands."),
     cl::init(true), cl::Hidden);
 
+static cl::opt<bool> ForceAll(
+    "x86-cmov-converter-force-all",
+    cl::desc("Convert all cmovs to branches."),
+    cl::init(false), cl::Hidden);
+
 namespace {
 
 /// Converts X86 cmov instructions into branches when profitable.
@@ -174,11 +179,11 @@
   TSchedModel.init(&STI);
 
   // Before we handle the more subtle cases of register-register CMOVs inside
-  // of potentially hot loops, we want to quickly remove all CMOVs with
-  // a memory operand. The CMOV will risk a stall waiting for the load to
-  // complete that speculative execution behind a branch is better suited to
-  // handle on modern x86 chips.
-  if (ForceMemOperand) {
+  // of potentially hot loops, we want to quickly remove all CMOVs (ForceAll) or
+  // the ones with a memory operand (ForceMemOperand option). The latter CMOV
+  // will risk a stall waiting for the load to complete that speculative
+  // execution behind a branch is better suited to handle on modern x86 chips.
+  if (ForceMemOperand || ForceAll) {
     CmovGroups AllCmovGroups;
     SmallVector<MachineBasicBlock *, 4> Blocks;
     for (auto &MBB : MF)
@@ -186,7 +191,8 @@
     if (collectCmovCandidates(Blocks, AllCmovGroups, /*IncludeLoads*/ true)) {
       for (auto &Group : AllCmovGroups) {
         // Skip any group that doesn't do at least one memory operand cmov.
-        if (llvm::none_of(Group, [&](MachineInstr *I) { return I->mayLoad(); }))
+        if (ForceMemOperand &&
+            llvm::none_of(Group, [&](MachineInstr *I) { return I->mayLoad(); }))
           continue;
 
         // For CMOV groups which we can rewrite and which contain a memory load,


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D119777.408606.patch
Type: text/x-patch
Size: 1977 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20220214/d791e8d0/attachment.bin>