[llvm] ef884e1 - [rs4gc] don't force a conflict for a canonical broadcast
Philip Reames via llvm-commits
llvm-commits at lists.llvm.org
Tue Mar 16 12:59:12 PDT 2021
Author: Philip Reames
Date: 2021-03-16T12:59:06-07:00
New Revision: ef884e155da7a46225f3441faa45d6d9e4249f8d
URL: https://github.com/llvm/llvm-project/commit/ef884e155da7a46225f3441faa45d6d9e4249f8d
DIFF: https://github.com/llvm/llvm-project/commit/ef884e155da7a46225f3441faa45d6d9e4249f8d.diff
LOG: [rs4gc] don't force a conflict for a canonical broadcast
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.
Differential Revision: https://reviews.llvm.org/D98315
Added:
Modified:
llvm/lib/Transforms/Scalar/RewriteStatepointsForGC.cpp
llvm/test/Transforms/RewriteStatepointsForGC/base-inference.ll
llvm/test/Transforms/RewriteStatepointsForGC/scalar-base-vector.ll
Removed:
################################################################################
diff --git a/llvm/lib/Transforms/Scalar/RewriteStatepointsForGC.cpp b/llvm/lib/Transforms/Scalar/RewriteStatepointsForGC.cpp
index 1ab4946476cc..755ebb881622 100644
--- a/llvm/lib/Transforms/Scalar/RewriteStatepointsForGC.cpp
+++ b/llvm/lib/Transforms/Scalar/RewriteStatepointsForGC.cpp
@@ -855,8 +855,11 @@ static Value *findBasePointer(Value *I, DefiningValueMapTy &Cache) {
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 @@ static Value *findBasePointer(Value *I, DefiningValueMapTy &Cache) {
BaseSV->setOperand(OperandIdx, Base);
};
UpdateOperand(0); // vector operand
- UpdateOperand(1); // vector operand
+ if (!BdvSV->isZeroEltSplat())
+ UpdateOperand(1); // vector operand
}
}
diff --git a/llvm/test/Transforms/RewriteStatepointsForGC/base-inference.ll b/llvm/test/Transforms/RewriteStatepointsForGC/base-inference.ll
index ddb229234b4e..0df3043fd4ca 100644
--- a/llvm/test/Transforms/RewriteStatepointsForGC/base-inference.ll
+++ b/llvm/test/Transforms/RewriteStatepointsForGC/base-inference.ll
@@ -227,11 +227,9 @@ define <2 x i8 addrspace(1)*> @test_shuffle_broadcast(i8 addrspace(1)* %a) gc "s
; 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:
diff --git a/llvm/test/Transforms/RewriteStatepointsForGC/scalar-base-vector.ll b/llvm/test/Transforms/RewriteStatepointsForGC/scalar-base-vector.ll
index c45a580119a7..2fb066608b4e 100644
--- a/llvm/test/Transforms/RewriteStatepointsForGC/scalar-base-vector.ll
+++ b/llvm/test/Transforms/RewriteStatepointsForGC/scalar-base-vector.ll
@@ -19,7 +19,7 @@ define i32 addrspace(1)* @test1(i8 addrspace(1)* %base1, <2 x i64> %offsets) gc
; 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 @@ define i32 addrspace(1)* @test2(i8 addrspace(1)* %base, <2 x i64> %offsets) gc "
; 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 @@ define i32 addrspace(1)* @test4(i8 addrspace(1)* %base, <2 x i64> %offsets) gc "
; 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]]) ]
More information about the llvm-commits
mailing list