InstCombine/AMDGPU: Fix constant folding of llvm.amdgcn.{icmp,fcmp}

[nhaehnle]

Summary

The return value of these intrinsics should always have 0 bits for
inactive threads. This means that when all arguments are constant
and the comparison evaluates to true, the intrinsic should return
the current exec mask.

Fixes some GL_ARB_shader_ballot tests.

[arsenm]

lib/Transforms/InstCombine/InstCombineCalls.cpp

3438

No else

3442–3445

Should have a comment about why this is done

test/Transforms/InstCombine/amdgcn-intrinsics.ll

1538–1540

Check that attributes 4 contains convergent?

[nhaehnle]

Address review comments and apply some more formatting fixes.

[arsenm]

LGTM

lib/Transforms/InstCombine/InstCombineCalls.cpp

3435–3437

Multiline, so braces

test/Transforms/InstCombine/amdgcn-intrinsics.ll

1538–1540

This is really missing nounwind?

[nhaehnle]

lib/Transforms/InstCombine/InstCombineCalls.cpp

3435–3437

Fixing this before I commit.

test/Transforms/InstCombine/amdgcn-intrinsics.ll

1538–1540

Yes. It's an attribute of the call-site, not of the function. read_register itself is nounwind readonly, but doesn't have the convergent attribute.

lib/Transforms/InstCombine/InstCombineCalls.cpp

if ((IsInteger && (CCVal < CmpInst::FIRST_ICMP_PREDICATE ||
       break;
 
     Value *Src0 = II->getArgOperand(0);
     Value *Src1 = II->getArgOperand(1);
 
     if (auto *CSrc0 = dyn_cast<Constant>(Src0)) {
       if (auto *CSrc1 = dyn_cast<Constant>(Src1)) {
         Constant *CCmp = ConstantExpr::getCompare(CCVal, CSrc0, CSrc1);
-        return replaceInstUsesWith(*II,
-                                   ConstantExpr::getSExt(CCmp, II->getType()));
+        if (CCmp->isNullValue())
+          return replaceInstUsesWith(
+              *II, ConstantExpr::getSExt(CCmp, II->getType()));
+
+        // The result of V_ICMP/V_FCMP assembly instructions (which this
+        // intrinsic exposes) is one bit per thread, masked with the EXEC
+        // register (which contains the bitmask of live threads). So a
+        // comparison that always returns true is the same as a read of the
+        // EXEC register.
+        Value *NewF = Intrinsic::getDeclaration(
+            II->getModule(), Intrinsic::read_register, II->getType());
+        Metadata *MDArgs[] = {MDString::get(II->getContext(), "exec")};
+        MDNode *MD = MDNode::get(II->getContext(), MDArgs);
+        Value *Args[] = {MetadataAsValue::get(II->getContext(), MD)};
+        CallInst *NewCall = Builder->CreateCall(NewF, Args);
+        NewCall->addAttribute(AttributeList::FunctionIndex,
+                              Attribute::Convergent);
+        NewCall->takeName(II);
+        return replaceInstUsesWith(*II, NewCall);
       }
 
       // Canonicalize constants to RHS.
       CmpInst::Predicate SwapPred
         = CmpInst::getSwappedPredicate(static_cast<CmpInst::Predicate>(CCVal));
       II->setArgOperand(0, Src1);
       II->setArgOperand(1, Src0);
       II->setArgOperand(2, ConstantInt::get(CC->getType(),

test/Transforms/InstCombine/amdgcn-intrinsics.ll

 ; CHECK-LABEL: @icmp_constant_inputs_false(
 ; CHECK: ret i64 0
 define i64 @icmp_constant_inputs_false() {
   %result = call i64 @llvm.amdgcn.icmp.i32(i32 9, i32 8, i32 32)
   ret i64 %result
 }
 
 ; CHECK-LABEL: @icmp_constant_inputs_true(
-; CHECK: ret i64 -1
+; CHECK: %result = call i64 @llvm.read_register.i64(metadata !0) #4
 define i64 @icmp_constant_inputs_true() {
   %result = call i64 @llvm.amdgcn.icmp.i32(i32 9, i32 8, i32 34)
   ret i64 %result
 }
 
 ; CHECK-LABEL: @icmp_constant_to_rhs_slt(
 ; CHECK: %result = call i64 @llvm.amdgcn.icmp.i32(i32 %x, i32 9, i32 38)
 define i64 @icmp_constant_to_rhs_slt(i32 %x) {

 ; CHECK-LABEL: @fcmp_constant_inputs_false(
 ; CHECK: ret i64 0
 define i64 @fcmp_constant_inputs_false() {
   %result = call i64 @llvm.amdgcn.fcmp.f32(float 2.0, float 4.0, i32 1)
   ret i64 %result
 }
 
 ; CHECK-LABEL: @fcmp_constant_inputs_true(
-; CHECK: ret i64 -1
+; CHECK: %result = call i64 @llvm.read_register.i64(metadata !0) #4
 define i64 @fcmp_constant_inputs_true() {
   %result = call i64 @llvm.amdgcn.fcmp.f32(float 2.0, float 4.0, i32 4)
   ret i64 %result
 }
 
 ; CHECK-LABEL: @fcmp_constant_to_rhs_olt(
 ; CHECK: %result = call i64 @llvm.amdgcn.fcmp.f32(float %x, float 4.000000e+00, i32 2)
 define i64 @fcmp_constant_to_rhs_olt(float %x) {
   %result = call i64 @llvm.amdgcn.fcmp.f32(float 4.0, float %x, i32 4)
   ret i64 %result
 }
+
+; CHECK: attributes #4 = { convergent }