[PATCH] D98315: [rs4gc] don't force a conflict for a canonical broadcast

Philip Reames via Phabricator via llvm-commits llvm-commits at lists.llvm.org
Tue Mar 16 12:59:28 PDT 2021 

This revision was landed with ongoing or failed builds.
This revision was automatically updated to reflect the committed changes.
Closed by commit rGef884e155da7: [rs4gc] don't force a conflict for a canonical broadcast (authored by reames).

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D98315/new/

https://reviews.llvm.org/D98315

Files:
  llvm/lib/Transforms/Scalar/RewriteStatepointsForGC.cpp
  llvm/test/Transforms/RewriteStatepointsForGC/base-inference.ll
  llvm/test/Transforms/RewriteStatepointsForGC/scalar-base-vector.ll


Index: llvm/test/Transforms/RewriteStatepointsForGC/scalar-base-vector.ll
===================================================================
--- llvm/test/Transforms/RewriteStatepointsForGC/scalar-base-vector.ll
+++ llvm/test/Transforms/RewriteStatepointsForGC/scalar-base-vector.ll
@@ -19,7 +19,7 @@
 ; 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
@@ -54,7 +54,7 @@
 ; 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
@@ -102,7 +102,7 @@
 ; 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]]) ]
Index: llvm/test/Transforms/RewriteStatepointsForGC/base-inference.ll
===================================================================
--- llvm/test/Transforms/RewriteStatepointsForGC/base-inference.ll
+++ llvm/test/Transforms/RewriteStatepointsForGC/base-inference.ll
@@ -227,11 +227,9 @@
 ; 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:
Index: llvm/lib/Transforms/Scalar/RewriteStatepointsForGC.cpp
===================================================================
--- llvm/lib/Transforms/Scalar/RewriteStatepointsForGC.cpp
+++ llvm/lib/Transforms/Scalar/RewriteStatepointsForGC.cpp
@@ -855,8 +855,11 @@
       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));
-      F(SV->getOperand(1));
+      if (!SV->isZeroEltSplat())
+        F(SV->getOperand(1));
     } else {
       llvm_unreachable("unexpected BDV type");
     }
@@ -1214,7 +1217,8 @@
         BaseSV->setOperand(OperandIdx, Base);
       };
       UpdateOperand(0); // vector operand
-      UpdateOperand(1); // vector operand
+      if (!BdvSV->isZeroEltSplat())
+        UpdateOperand(1); // vector operand
     }
   }
 


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D98315.331073.patch
Type: text/x-patch
Size: 6815 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20210316/90ac1a46/attachment.bin>

More information about the llvm-commits mailing list