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

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.

Reviewers: arsenm

Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye

Differential Revision: https://reviews.llvm.org/D32344

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@301195 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/InstCombine/InstCombineCalls.cpp b/lib/Transforms/InstCombine/InstCombineCalls.cpp
index e7aa1a45737141f2ceca58825ff999129246c12c..0ea1c7be42214dfd9662510b7be05fbe57e09309 100644 (file)
--- a/lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ b/lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -3432,8 +3432,26 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) {
     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.

diff --git a/test/Transforms/InstCombine/amdgcn-intrinsics.ll b/test/Transforms/InstCombine/amdgcn-intrinsics.ll
index deae5502bcdb8b65b4fec489cae70ace1ce4e249..357085fd31fee9fee838f34eca2472115e427c08 100644 (file)
--- a/test/Transforms/InstCombine/amdgcn-intrinsics.ll
+++ b/test/Transforms/InstCombine/amdgcn-intrinsics.ll
@@ -1259,7 +1259,7 @@ define i64 @icmp_constant_inputs_false() {
 }
 
 ; 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
@@ -1524,7 +1524,7 @@ define i64 @fcmp_constant_inputs_false() {
 }
 
 ; 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
@@ -1536,3 +1536,5 @@ 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 }