SpeculateAroundPHIs: Respect convergent

[arsenm]

Summary

This would break for speculatable and convergent intrinsic calls.

[jdoerfert]

I don't think we need the AMDGPU test case specialization here. A generic test with a convergent and one with a non-duplicatable call would make more sense to me.

[arsenm]

In D62531#1519636, @jdoerfert wrote:

I don't think we need the AMDGPU test case specialization here. A generic test with a convergent and one with a non-duplicatable call would make more sense to me.

It concludes an arbitrary call is too expensive, so the problem doesn't manifest. I didn't see any control to override the cost

[arsenm]

A fake intrinsic with an llvm. prefix works for the test

[jdoerfert]

Last request from my side, a positive test that looks exactly the same but without the convergent/noduplicate to make sure it is actually transformed.

Otherwise, this looks good to me.

[arsenm]

r361957

lib/Transforms/Scalar/SpeculateAroundPHIs.cpp

for (Use &U : PN.uses()) {
     // FIXME: We use a blunt hammer of requiring them to be in the same basic
     // block. We should consider using actual post-dominance here in the
     // future.
     if (UI->getParent() != PhiBB) {
       LLVM_DEBUG(dbgs() << "  Unsafe: use in a different BB: " << *UI << "\n");
       return false;
     }
 
+    if (auto CS = ImmutableCallSite(UI)) {
+      if (CS.isConvergent() || CS.cannotDuplicate()) {
+        LLVM_DEBUG(dbgs() << "  Unsafe: convergent "
+                   "callsite cannot de duplicated: " << *UI << '\n');
+        return false;
+      }
+    }
+
     // FIXME: This check is much too conservative. We're not going to move these
     // instructions onto new dynamic paths through the program unless there is
     // a call instruction between the use and the PHI node. And memory isn't
     // changing unless there is a store in that same sequence. We should
     // probably change this to do at least a limited scan of the intervening
     // instructions and allow handling stores in easily proven safe cases.
     if (mayBeMemoryDependent(*UI)) {
       LLVM_DEBUG(dbgs() << "  Unsafe: can't speculate use: " << *UI << "\n");

test/Transforms/SpeculateAroundPHIs/convergent.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -S -passes=spec-phis < %s | FileCheck %s
+; Make sure convergent and noduplicate calls aren't duplicated.
+
+declare i32 @llvm.convergent(i32) #0
+declare i32 @llvm.noduplicate(i32) #1
+
+define i32 @test_convergent(i1 %flag, i32 %arg) #0 {
+; CHECK-LABEL: @test_convergent(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    br i1 [[FLAG:%.*]], label [[A:%.*]], label [[B:%.*]]
+; CHECK:       a:
+; CHECK-NEXT:    br label [[EXIT:%.*]]
+; CHECK:       b:
+; CHECK-NEXT:    br label [[EXIT]]
+; CHECK:       exit:
+; CHECK-NEXT:    [[P:%.*]] = phi i32 [ 7, [[A]] ], [ 11, [[B]] ]
+; CHECK-NEXT:    [[SUM:%.*]] = call i32 @llvm.convergent(i32 [[P]])
+; CHECK-NEXT:    ret i32 [[SUM]]
+;
+entry:
+  br i1 %flag, label %a, label %b
+
+a:
+  br label %exit
+
+b:
+  br label %exit
+
+exit:
+  %p = phi i32 [ 7, %a ], [ 11, %b ]
+  %sum = call i32 @llvm.convergent(i32 %p)
+  ret i32 %sum
+}
+
+define i32 @test_noduplicate(i1 %flag, i32 %arg) #1 {
+; CHECK-LABEL: @test_noduplicate(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    br i1 [[FLAG:%.*]], label [[A:%.*]], label [[B:%.*]]
+; CHECK:       a:
+; CHECK-NEXT:    br label [[EXIT:%.*]]
+; CHECK:       b:
+; CHECK-NEXT:    br label [[EXIT]]
+; CHECK:       exit:
+; CHECK-NEXT:    [[P:%.*]] = phi i32 [ 7, [[A]] ], [ 11, [[B]] ]
+; CHECK-NEXT:    [[SUM:%.*]] = call i32 @llvm.noduplicate(i32 [[P]])
+; CHECK-NEXT:    ret i32 [[SUM]]
+;
+entry:
+  br i1 %flag, label %a, label %b
+
+a:
+  br label %exit
+
+b:
+  br label %exit
+
+exit:
+  %p = phi i32 [ 7, %a ], [ 11, %b ]
+  %sum = call i32 @llvm.noduplicate(i32 %p)
+  ret i32 %sum
+}
+
+attributes #0 = { nounwind readnone convergent speculatable }
+attributes #1 = { nounwind readnone noduplicate speculatable }