[PATCH] D32344: InstCombine/AMDGPU: Fix constant folding of llvm.amdgcn.{icmp, fcmp}

[Nicolai Hähnle via Phabricator via llvm-commits]

nhaehnle updated this revision to Diff 96413.
nhaehnle added a comment.

Address review comments and apply some more formatting fixes.


https://reviews.llvm.org/D32344

Files:
  lib/Transforms/InstCombine/InstCombineCalls.cpp
  test/Transforms/InstCombine/amdgcn-intrinsics.ll


Index: test/Transforms/InstCombine/amdgcn-intrinsics.ll
===================================================================
--- test/Transforms/InstCombine/amdgcn-intrinsics.ll
+++ test/Transforms/InstCombine/amdgcn-intrinsics.ll
@@ -1259,7 +1259,7 @@
 }
 
 ; 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 @@
 }
 
 ; 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 @@
   %result = call i64 @llvm.amdgcn.fcmp.f32(float 4.0, float %x, i32 4)
   ret i64 %result
 }
+
+; CHECK: attributes #4 = { convergent }
Index: lib/Transforms/InstCombine/InstCombineCalls.cpp
===================================================================
--- lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -3432,8 +3432,25 @@
     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.


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D32344.96413.patch
Type: text/x-patch
Size: 2628 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170424/0c6782e9/attachment.bin>