[EarlyCSE] Simplify guard intrinsics

[sanjoy]

Summary

This change teaches EarlyCSE some basic properties of guard intrinsics:

• Guard intrinsics read all memory, but don't write to any memory
• After a guard has executed, the condition it was guarding on can be assumed to be true
• Guard intrinsics on a constant true are no-ops

[majnemer]

lib/Transforms/Scalar/EarlyCSE.cpp

560 ↗

(On Diff #55158)

Any reason against using m_One here?

[sanjoy]

• Address review

[reames]

LGTM w/minor comments.

lib/Transforms/Scalar/EarlyCSE.cpp

561 ↗

(On Diff #55160)

You should consider (as a future change) moving this into isInstructionTriviallyDead.

test/Transforms/EarlyCSE/guards.ll

26 ↗

(On Diff #55160)

If you used a sub here, you could check for ret i32 0

45 ↗

(On Diff #55160)

I'd suggest splitting this into two test cases.

llvm/trunk/lib/Transforms/Scalar/EarlyCSE.cpp

for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
     // they're marked as such to ensure preservation of control dependencies),
     // and this pass will not disturb any of the assumption's control
     // dependencies.
     if (match(Inst, m_Intrinsic<Intrinsic::assume>())) {
       DEBUG(dbgs() << "EarlyCSE skipping assumption: " << *Inst << '\n');
       continue;
     }
 
+    if (match(Inst, m_Intrinsic<Intrinsic::experimental_guard>())) {
+      Value *Cond = cast<CallInst>(Inst)->getArgOperand(0);
+
+      if (match(Cond, m_One())) {
+        // Elide guards on true, since operationally they're no-ops.  In the
+        // future we can consider more sophisticated tradeoffs here with
+        // consideration to potential for check widening, but for now we keep
+        // things simple.
+        Inst->eraseFromParent();
+      } else if (auto *CondI = dyn_cast<Instruction>(Cond)) {
+        // The condition we're on guarding here is true for all dominated
+        // locations.
+        if (SimpleValue::canHandle(CondI))
+          AvailableValues.insert(CondI, ConstantInt::getTrue(BB->getContext()));
+      }
+
+      // Guard intrinsics read all memory, but don't write any memory.
+      // Accordingly, don't update the generation but consume the last store (to
+      // avoid an incorrect DSE).
+      LastStore = nullptr;
+      continue;
+    }
+
     // If the instruction can be simplified (e.g. X+0 = X) then replace it with
     // its simpler value.
     if (Value *V = SimplifyInstruction(Inst, DL, &TLI, &DT, &AC)) {
       DEBUG(dbgs() << "EarlyCSE Simplify: " << *Inst << "  to: " << *V << '\n');
       Inst->replaceAllUsesWith(V);
       Inst->eraseFromParent();
       Changed = true;
       ++NumSimplify;

llvm/trunk/test/Transforms/EarlyCSE/guards.ll

+; RUN: opt -S -early-cse < %s | FileCheck %s
+
+declare void @llvm.experimental.guard(i1,...)
+
+define i32 @test0(i32* %ptr, i1 %cond) {
+; We can do store to load forwarding over a guard, since it does not
+; clobber memory
+
+; CHECK-LABEL: @test0(
+; CHECK-NEXT:  store i32 40, i32* %ptr
+; CHECK-NEXT:  call void (i1, ...) @llvm.experimental.guard(i1 %cond) [ "deopt"() ]
+; CHECK-NEXT:  ret i32 40
+
+  store i32 40, i32* %ptr
+  call void(i1,...) @llvm.experimental.guard(i1 %cond) [ "deopt"() ]
+  %rval = load i32, i32* %ptr
+  ret i32 %rval
+}
+
+define i32 @test1(i32* %val, i1 %cond) {
+; We can CSE loads over a guard, since it does not clobber memory
+
+; CHECK-LABEL: @test1(
+; CHECK-NEXT:  %val0 = load i32, i32* %val
+; CHECK-NEXT:  call void (i1, ...) @llvm.experimental.guard(i1 %cond) [ "deopt"() ]
+; CHECK-NEXT:  ret i32 0
+
+  %val0 = load i32, i32* %val
+  call void(i1,...) @llvm.experimental.guard(i1 %cond) [ "deopt"() ]
+  %val1 = load i32, i32* %val
+  %rval = sub i32 %val0, %val1
+  ret i32 %rval
+}
+
+define i32 @test2() {
+; Guards on "true" get removed
+
+; CHECK-LABEL: @test2(
+; CHECK-NEXT: ret i32 0
+  call void(i1, ...) @llvm.experimental.guard(i1 true) [ "deopt"() ]
+  ret i32 0
+}
+
+define i32 @test3(i32 %val) {
+; After a guard has executed the condition it was guarding is known to
+; be true.
+
+; CHECK-LABEL: @test3(
+; CHECK-NEXT:  %cond0 = icmp slt i32 %val, 40
+; CHECK-NEXT:  call void (i1, ...) @llvm.experimental.guard(i1 %cond0) [ "deopt"() ]
+; CHECK-NEXT:  ret i32 -1
+
+  %cond0 = icmp slt i32 %val, 40
+  call void(i1,...) @llvm.experimental.guard(i1 %cond0) [ "deopt"() ]
+  %cond1 = icmp slt i32 %val, 40
+  call void(i1,...) @llvm.experimental.guard(i1 %cond1) [ "deopt"() ]
+
+  %cond2 = icmp slt i32 %val, 40
+  %rval = sext i1 %cond2 to i32
+  ret i32 %rval
+}
+
+define i32 @test3.unhandled(i32 %val) {
+; After a guard has executed the condition it was guarding is known to
+; be true.
+
+; CHECK-LABEL: @test3.unhandled(
+; CHECK-NEXT:  %cond0 = icmp slt i32 %val, 40
+; CHECK-NEXT:  call void (i1, ...) @llvm.experimental.guard(i1 %cond0) [ "deopt"() ]
+; CHECK-NEXT:  %cond1 = icmp sge i32 %val, 40
+; CHECK-NEXT:  call void (i1, ...) @llvm.experimental.guard(i1 %cond1) [ "deopt"() ]
+; CHECK-NEXT:  ret i32 0
+
+; Demonstrates a case we do not yet handle (it is legal to fold %cond2
+; to false)
+  %cond0 = icmp slt i32 %val, 40
+  call void(i1,...) @llvm.experimental.guard(i1 %cond0) [ "deopt"() ]
+  %cond1 = icmp sge i32 %val, 40
+  call void(i1,...) @llvm.experimental.guard(i1 %cond1) [ "deopt"() ]
+  ret i32 0
+}
+
+define i32 @test4(i32 %val, i1 %c) {
+; Same as test3, but with some control flow involved.
+
+; CHECK-LABEL: @test4(
+; CHECK: entry:
+; CHECK-NEXT:  %cond0 = icmp slt i32 %val, 40
+; CHECK-NEXT:  call void (i1, ...) @llvm.experimental.guard(i1 %cond0
+; CHECK-NEXT:  br label %bb0
+
+; CHECK:     bb0:
+; CHECK-NEXT:  %cond2 = icmp ult i32 %val, 200
+; CHECK-NEXT:  call void (i1, ...) @llvm.experimental.guard(i1 %cond2
+; CHECK-NEXT:  br i1 %c, label %left, label %right
+
+; CHECK:     left:
+; CHECK-NEXT:  ret i32 0
+
+; CHECK:     right:
+; CHECK-NEXT:  ret i32 20
+
+entry:
+  %cond0 = icmp slt i32 %val, 40
+  call void(i1,...) @llvm.experimental.guard(i1 %cond0) [ "deopt"() ]
+  %cond1 = icmp slt i32 %val, 40
+  call void(i1,...) @llvm.experimental.guard(i1 %cond1) [ "deopt"() ]
+  br label %bb0
+
+bb0:
+  %cond2 = icmp ult i32 %val, 200
+  call void(i1,...) @llvm.experimental.guard(i1 %cond2) [ "deopt"() ]
+  br i1 %c, label %left, label %right
+
+left:
+  %cond3 = icmp ult i32 %val, 200
+  call void(i1,...) @llvm.experimental.guard(i1 %cond3) [ "deopt"() ]
+  ret i32 0
+
+right:
+ ret i32 20
+}
+
+define i32 @test5(i32 %val, i1 %c) {
+; Same as test4, but the %left block has mutliple predecessors.
+
+; CHECK-LABEL: @test5(
+
+; CHECK: entry:
+; CHECK-NEXT:  %cond0 = icmp slt i32 %val, 40
+; CHECK-NEXT:  call void (i1, ...) @llvm.experimental.guard(i1 %cond0
+; CHECK-NEXT:  br label %bb0
+
+; CHECK: bb0:
+; CHECK-NEXT:  %cond2 = icmp ult i32 %val, 200
+; CHECK-NEXT:  call void (i1, ...) @llvm.experimental.guard(i1 %cond2
+; CHECK-NEXT:  br i1 %c, label %left, label %right
+
+; CHECK: left:
+; CHECK-NEXT:  br label %right
+
+; CHECK: right:
+; CHECK-NEXT:  br label %left
+
+entry:
+  %cond0 = icmp slt i32 %val, 40
+  call void(i1,...) @llvm.experimental.guard(i1 %cond0) [ "deopt"() ]
+  %cond1 = icmp slt i32 %val, 40
+  call void(i1,...) @llvm.experimental.guard(i1 %cond1) [ "deopt"() ]
+  br label %bb0
+
+bb0:
+  %cond2 = icmp ult i32 %val, 200
+  call void(i1,...) @llvm.experimental.guard(i1 %cond2) [ "deopt"() ]
+  br i1 %c, label %left, label %right
+
+left:
+  %cond3 = icmp ult i32 %val, 200
+  call void(i1,...) @llvm.experimental.guard(i1 %cond3) [ "deopt"() ]
+  br label %right
+
+right:
+  br label %left
+}
+
+define void @test6(i1 %c, i32* %ptr) {
+; Check that we do not DSE over calls to @llvm.experimental.guard.
+; Guard intrinsics do _read_ memory, so th call to guard below needs
+; to see the store of 500 to %ptr
+
+; CHECK-LABEL: @test6(
+; CHECK-NEXT:  store i32 500, i32* %ptr
+; CHECK-NEXT:  call void (i1, ...) @llvm.experimental.guard(i1 %c) [ "deopt"() ]
+; CHECK-NEXT:  store i32 600, i32* %ptr
+
+
+  store i32 500, i32* %ptr
+  call void(i1,...) @llvm.experimental.guard(i1 %c) [ "deopt"() ]
+  store i32 600, i32* %ptr
+  ret void
+}