[rs4gc] don't force a conflict for a canonical broadcast

[reames]

Summary

Minor follow up to D98122 noticed while writing tests for that.

A broadcast is a shufflevector where only one input is used. Because of the way we handle constants (undef is a constant), the canonical shuffle sees a meet of (some value) and (nullptr). Given this, every broadcast gets treated as a conflict and a new base pointer computation is added.

The other way to tackle this would be to change constant handling specifically for undefs, but this seems easier.

llvm/lib/Transforms/Scalar/RewriteStatepointsForGC.cpp

if (PHINode *PN = dyn_cast<PHINode>(BDV)) {
       F(SI->getTrueValue());
       F(SI->getFalseValue());
     } else if (auto *EE = dyn_cast<ExtractElementInst>(BDV)) {
       F(EE->getVectorOperand());
     } else if (auto *IE = dyn_cast<InsertElementInst>(BDV)) {
       F(IE->getOperand(0));
       F(IE->getOperand(1));
     } else if (auto *SV = dyn_cast<ShuffleVectorInst>(BDV)) {
+      // For a canonical broadcast, ignore the undef argument
+      // (without this, we insert a parallel base shuffle for every broadcast)
       F(SV->getOperand(0));
+      if (!SV->isZeroEltSplat())
         F(SV->getOperand(1));
     } else {
       llvm_unreachable("unexpected BDV type");
     }
   };
 
 
   // Recursively fill in all base defining values reachable from the initial
   // one for which we don't already know a definite base value for
#endif
       auto *BaseSV = cast<ShuffleVectorInst>(State.getBaseValue());
       auto *BdvSV = cast<ShuffleVectorInst>(BDV);
       auto UpdateOperand = [&](int OperandIdx) {
         Value *InVal = BdvSV->getOperand(OperandIdx);
         Value *Base = getBaseForInput(InVal, BaseSV);
         BaseSV->setOperand(OperandIdx, Base);
       };
       UpdateOperand(0); // vector operand
+      if (!BdvSV->isZeroEltSplat())
         UpdateOperand(1); // vector operand
     }
   }
 
 #ifndef NDEBUG
   VerifyStates();
 #endif
 
   // Cache all of our results so we can cheaply reuse them

llvm/test/Transforms/RewriteStatepointsForGC/base-inference.ll

 }
 
 ; TODO: Special case worth handling - we interpret the shuffle as if we need
 ; to select the base pointers from either input when the mask is known.
 define <2 x i8 addrspace(1)*> @test_shuffle_broadcast(i8 addrspace(1)* %a) gc "statepoint-example" {
 ; CHECK-LABEL: @test_shuffle_broadcast(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[IE:%.*]] = insertelement <2 x i8 addrspace(1)*> zeroinitializer, i8 addrspace(1)* [[A:%.*]], i64 0
-; CHECK-NEXT:    [[BROADCAST_BASE:%.*]] = shufflevector <2 x i8 addrspace(1)*> [[IE]], <2 x i8 addrspace(1)*> zeroinitializer, <2 x i32> zeroinitializer, !is_base_value !0
 ; CHECK-NEXT:    [[BROADCAST:%.*]] = shufflevector <2 x i8 addrspace(1)*> [[IE]], <2 x i8 addrspace(1)*> undef, <2 x i32> zeroinitializer
-; CHECK-NEXT:    [[STATEPOINT_TOKEN:%.*]] = call token (i64, i32, void ()*, i32, i32, ...) @llvm.experimental.gc.statepoint.p0f_isVoidf(i64 2882400000, i32 0, void ()* @foo, i32 0, i32 0, i32 0, i32 0) [ "gc-live"(<2 x i8 addrspace(1)*> [[BROADCAST]], <2 x i8 addrspace(1)*> [[BROADCAST_BASE]]) ]
-; CHECK-NEXT:    [[BROADCAST_RELOCATED:%.*]] = call coldcc <2 x i8 addrspace(1)*> @llvm.experimental.gc.relocate.v2p1i8(token [[STATEPOINT_TOKEN]], i32 1, i32 0)
-; CHECK-NEXT:    [[BROADCAST_BASE_RELOCATED:%.*]] = call coldcc <2 x i8 addrspace(1)*> @llvm.experimental.gc.relocate.v2p1i8(token [[STATEPOINT_TOKEN]], i32 1, i32 1)
+; CHECK-NEXT:    [[STATEPOINT_TOKEN:%.*]] = call token (i64, i32, void ()*, i32, i32, ...) @llvm.experimental.gc.statepoint.p0f_isVoidf(i64 2882400000, i32 0, void ()* @foo, i32 0, i32 0, i32 0, i32 0) [ "gc-live"(<2 x i8 addrspace(1)*> [[BROADCAST]]) ]
+; CHECK-NEXT:    [[BROADCAST_RELOCATED:%.*]] = call coldcc <2 x i8 addrspace(1)*> @llvm.experimental.gc.relocate.v2p1i8(token [[STATEPOINT_TOKEN]], i32 0, i32 0)
 ; CHECK-NEXT:    ret <2 x i8 addrspace(1)*> [[BROADCAST_RELOCATED]]
 ;
 entry:
   %ie = insertelement <2 x i8 addrspace(1)*> zeroinitializer, i8 addrspace(1)* %a, i64 0
   %broadcast = shufflevector <2 x i8 addrspace(1)*> %ie, <2 x i8 addrspace(1)*> undef, <2 x i32> zeroinitializer
   call void @foo()
   ret <2 x i8 addrspace(1)*> %broadcast
 }

llvm/test/Transforms/RewriteStatepointsForGC/scalar-base-vector.ll

 ; CHECK-NEXT:    [[BASE21:%.*]] = call i8 addrspace(1)* @llvm.experimental.gc.result.p1i8(token [[STATEPOINT_TOKEN]])
 ; CHECK-NEXT:    br label [[SECOND]]
 ; CHECK:       second:
 ; CHECK-NEXT:    [[PHI:%.*]] = phi i8 addrspace(1)* [ [[BASE1:%.*]], [[ENTRY:%.*]] ], [ [[BASE21]], [[FIRST]] ]
 ; CHECK-NEXT:    [[BASE_I32:%.*]] = bitcast i8 addrspace(1)* [[PHI]] to i32 addrspace(1)*
 ; CHECK-NEXT:    [[CAST:%.*]] = bitcast i8 addrspace(1)* [[PHI]] to i32 addrspace(1)*
 ; CHECK-NEXT:    [[DOTSPLATINSERT_BASE:%.*]] = insertelement <2 x i32 addrspace(1)*> zeroinitializer, i32 addrspace(1)* [[CAST]], i32 0, !is_base_value !0
 ; CHECK-NEXT:    [[DOTSPLATINSERT:%.*]] = insertelement <2 x i32 addrspace(1)*> poison, i32 addrspace(1)* [[BASE_I32]], i32 0
-; CHECK-NEXT:    [[DOTSPLAT_BASE:%.*]] = shufflevector <2 x i32 addrspace(1)*> [[DOTSPLATINSERT_BASE]], <2 x i32 addrspace(1)*> zeroinitializer, <2 x i32> zeroinitializer, !is_base_value !0
+; CHECK-NEXT:    [[DOTSPLAT_BASE:%.*]] = shufflevector <2 x i32 addrspace(1)*> [[DOTSPLATINSERT_BASE]], <2 x i32 addrspace(1)*> undef, <2 x i32> zeroinitializer, !is_base_value !0
 ; CHECK-NEXT:    [[DOTSPLAT:%.*]] = shufflevector <2 x i32 addrspace(1)*> [[DOTSPLATINSERT]], <2 x i32 addrspace(1)*> poison, <2 x i32> zeroinitializer
 ; CHECK-NEXT:    [[VEC:%.*]] = getelementptr i32, <2 x i32 addrspace(1)*> [[DOTSPLAT]], <2 x i64> [[OFFSETS:%.*]]
 ; CHECK-NEXT:    [[PTR_BASE:%.*]] = extractelement <2 x i32 addrspace(1)*> [[DOTSPLAT_BASE]], i32 1, !is_base_value !0
 ; CHECK-NEXT:    [[PTR:%.*]] = extractelement <2 x i32 addrspace(1)*> [[VEC]], i32 1
 ; CHECK-NEXT:    [[STATEPOINT_TOKEN2:%.*]] = call token (i64, i32, void ()*, i32, i32, ...) @llvm.experimental.gc.statepoint.p0f_isVoidf(i64 2882400000, i32 0, void ()* @do_safepoint, i32 0, i32 0, i32 0, i32 0) [ "deopt"(i32 0, i32 -1, i32 0, i32 0, i32 0), "gc-live"(i32 addrspace(1)* [[PTR]], i32 addrspace(1)* [[PTR_BASE]]) ]
 ; CHECK-NEXT:    [[PTR_RELOCATED:%.*]] = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(token [[STATEPOINT_TOKEN2]], i32 1, i32 0)
 ; CHECK-NEXT:    [[PTR_RELOCATED_CASTED:%.*]] = bitcast i8 addrspace(1)* [[PTR_RELOCATED]] to i32 addrspace(1)*
 ; CHECK-NEXT:    [[PTR_BASE_RELOCATED:%.*]] = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(token [[STATEPOINT_TOKEN2]], i32 1, i32 1)

 
 define i32 addrspace(1)* @test2(i8 addrspace(1)* %base, <2 x i64> %offsets) gc "statepoint-example" {
 ; CHECK-LABEL: @test2(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[BASE_I32:%.*]] = bitcast i8 addrspace(1)* [[BASE:%.*]] to i32 addrspace(1)*
 ; CHECK-NEXT:    [[CAST:%.*]] = bitcast i8 addrspace(1)* [[BASE]] to i32 addrspace(1)*
 ; CHECK-NEXT:    [[DOTSPLATINSERT_BASE:%.*]] = insertelement <2 x i32 addrspace(1)*> zeroinitializer, i32 addrspace(1)* [[CAST]], i32 0, !is_base_value !0
 ; CHECK-NEXT:    [[DOTSPLATINSERT:%.*]] = insertelement <2 x i32 addrspace(1)*> poison, i32 addrspace(1)* [[BASE_I32]], i32 0
-; CHECK-NEXT:    [[DOTSPLAT_BASE:%.*]] = shufflevector <2 x i32 addrspace(1)*> [[DOTSPLATINSERT_BASE]], <2 x i32 addrspace(1)*> zeroinitializer, <2 x i32> zeroinitializer, !is_base_value !0
+; CHECK-NEXT:    [[DOTSPLAT_BASE:%.*]] = shufflevector <2 x i32 addrspace(1)*> [[DOTSPLATINSERT_BASE]], <2 x i32 addrspace(1)*> undef, <2 x i32> zeroinitializer, !is_base_value !0
 ; CHECK-NEXT:    [[DOTSPLAT:%.*]] = shufflevector <2 x i32 addrspace(1)*> [[DOTSPLATINSERT]], <2 x i32 addrspace(1)*> poison, <2 x i32> zeroinitializer
 ; CHECK-NEXT:    [[VEC:%.*]] = getelementptr i32, <2 x i32 addrspace(1)*> [[DOTSPLAT]], <2 x i64> [[OFFSETS:%.*]]
 ; CHECK-NEXT:    [[PTR_BASE:%.*]] = extractelement <2 x i32 addrspace(1)*> [[DOTSPLAT_BASE]], i32 1, !is_base_value !0
 ; CHECK-NEXT:    [[PTR:%.*]] = extractelement <2 x i32 addrspace(1)*> [[VEC]], i32 1
 ; CHECK-NEXT:    [[STATEPOINT_TOKEN:%.*]] = call token (i64, i32, void ()*, i32, i32, ...) @llvm.experimental.gc.statepoint.p0f_isVoidf(i64 2882400000, i32 0, void ()* @do_safepoint, i32 0, i32 0, i32 0, i32 0) [ "gc-live"(i32 addrspace(1)* [[PTR]], i32 addrspace(1)* [[PTR_BASE]]) ]
 ; CHECK-NEXT:    [[PTR_RELOCATED:%.*]] = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(token [[STATEPOINT_TOKEN]], i32 1, i32 0)
 ; CHECK-NEXT:    [[PTR_RELOCATED_CASTED:%.*]] = bitcast i8 addrspace(1)* [[PTR_RELOCATED]] to i32 addrspace(1)*
 ; CHECK-NEXT:    [[PTR_BASE_RELOCATED:%.*]] = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(token [[STATEPOINT_TOKEN]], i32 1, i32 1)

 
 define i32 addrspace(1)* @test4(i8 addrspace(1)* %base, <2 x i64> %offsets) gc "statepoint-example" {
 ; CHECK-LABEL: @test4(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[BASE_I32:%.*]] = bitcast i8 addrspace(1)* [[BASE:%.*]] to i32 addrspace(1)*
 ; CHECK-NEXT:    [[CAST:%.*]] = bitcast i8 addrspace(1)* [[BASE]] to i32 addrspace(1)*
 ; CHECK-NEXT:    [[DOTSPLATINSERT_BASE:%.*]] = insertelement <2 x i32 addrspace(1)*> zeroinitializer, i32 addrspace(1)* [[CAST]], i32 0, !is_base_value !0
 ; CHECK-NEXT:    [[DOTSPLATINSERT:%.*]] = insertelement <2 x i32 addrspace(1)*> poison, i32 addrspace(1)* [[BASE_I32]], i32 0
-; CHECK-NEXT:    [[DOTSPLAT_BASE:%.*]] = shufflevector <2 x i32 addrspace(1)*> [[DOTSPLATINSERT_BASE]], <2 x i32 addrspace(1)*> zeroinitializer, <2 x i32> zeroinitializer, !is_base_value !0
+; CHECK-NEXT:    [[DOTSPLAT_BASE:%.*]] = shufflevector <2 x i32 addrspace(1)*> [[DOTSPLATINSERT_BASE]], <2 x i32 addrspace(1)*> undef, <2 x i32> zeroinitializer, !is_base_value !0
 ; CHECK-NEXT:    [[DOTSPLAT:%.*]] = shufflevector <2 x i32 addrspace(1)*> [[DOTSPLATINSERT]], <2 x i32 addrspace(1)*> poison, <2 x i32> zeroinitializer
 ; CHECK-NEXT:    [[VEC:%.*]] = getelementptr i32, <2 x i32 addrspace(1)*> [[DOTSPLAT]], <2 x i64> [[OFFSETS:%.*]]
 ; CHECK-NEXT:    [[STATEPOINT_TOKEN:%.*]] = call token (i64, i32, void ()*, i32, i32, ...) @llvm.experimental.gc.statepoint.p0f_isVoidf(i64 2882400000, i32 0, void ()* @do_safepoint, i32 0, i32 0, i32 0, i32 0) [ "gc-live"(<2 x i32 addrspace(1)*> [[VEC]], <2 x i32 addrspace(1)*> [[DOTSPLAT_BASE]]) ]
 ; CHECK-NEXT:    [[VEC_RELOCATED:%.*]] = call coldcc <2 x i8 addrspace(1)*> @llvm.experimental.gc.relocate.v2p1i8(token [[STATEPOINT_TOKEN]], i32 1, i32 0)
 ; CHECK-NEXT:    [[VEC_RELOCATED_CASTED:%.*]] = bitcast <2 x i8 addrspace(1)*> [[VEC_RELOCATED]] to <2 x i32 addrspace(1)*>
 ; CHECK-NEXT:    [[DOTSPLAT_BASE_RELOCATED:%.*]] = call coldcc <2 x i8 addrspace(1)*> @llvm.experimental.gc.relocate.v2p1i8(token [[STATEPOINT_TOKEN]], i32 1, i32 1)
 ; CHECK-NEXT:    [[DOTSPLAT_BASE_RELOCATED_CASTED:%.*]] = bitcast <2 x i8 addrspace(1)*> [[DOTSPLAT_BASE_RELOCATED]] to <2 x i32 addrspace(1)*>
 ; CHECK-NEXT:    [[PTR:%.*]] = extractelement <2 x i32 addrspace(1)*> [[VEC_RELOCATED_CASTED]], i32 1