[llvm] c490658 - [VPlan] Use skipCostComputation when pre-computing induction costs.
Florian Hahn via llvm-commits
llvm-commits at lists.llvm.org
Thu Aug 29 13:21:16 PDT 2024
Author: Florian Hahn
Date: 2024-08-29T21:20:00+01:00
New Revision: c4906588ce47de33d59bcd95f3e82ce2c3e61c23
URL: https://github.com/llvm/llvm-project/commit/c4906588ce47de33d59bcd95f3e82ce2c3e61c23
DIFF: https://github.com/llvm/llvm-project/commit/c4906588ce47de33d59bcd95f3e82ce2c3e61c23.diff
LOG: [VPlan] Use skipCostComputation when pre-computing induction costs.
This ensures we skip any instructions identified to be ignored by the
legacy cost model as well. Fixes a divergence between legacy and
VPlan-based cost model.
Fixes https://github.com/llvm/llvm-project/issues/106417.
Added:
llvm/test/Transforms/LoopVectorize/RISCV/induction-costs.ll
Modified:
llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
Removed:
################################################################################
diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index 4cc75e2e754603..6babfd1eee9108 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -7113,7 +7113,7 @@ LoopVectorizationPlanner::precomputeCosts(VPlan &Plan, ElementCount VF,
IVInsts.push_back(CI);
}
for (Instruction *IVInst : IVInsts) {
- if (!CostCtx.SkipCostComputation.insert(IVInst).second)
+ if (CostCtx.skipCostComputation(IVInst, VF.isVector()))
continue;
InstructionCost InductionCost = CostCtx.getLegacyCost(IVInst, VF);
LLVM_DEBUG({
@@ -7121,6 +7121,7 @@ LoopVectorizationPlanner::precomputeCosts(VPlan &Plan, ElementCount VF,
<< ": induction instruction " << *IVInst << "\n";
});
Cost += InductionCost;
+ CostCtx.SkipCostComputation.insert(IVInst);
}
}
diff --git a/llvm/test/Transforms/LoopVectorize/RISCV/induction-costs.ll b/llvm/test/Transforms/LoopVectorize/RISCV/induction-costs.ll
new file mode 100644
index 00000000000000..bee7bb7bd61622
--- /dev/null
+++ b/llvm/test/Transforms/LoopVectorize/RISCV/induction-costs.ll
@@ -0,0 +1,192 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -p loop-vectorize -S %s | FileCheck %s
+
+target datalayout = "e-m:e-p:64:64-i64:64-i128:128-n32:64-S128"
+target triple = "riscv64-unknown-linux-gnu"
+
+; Test case for https://github.com/llvm/llvm-project/issues/106417.
+define void @skip_free_iv_truncate(i16 %x, ptr %A) #0 {
+; CHECK-LABEL: define void @skip_free_iv_truncate(
+; CHECK-SAME: i16 [[X:%.*]], ptr [[A:%.*]]) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT: [[ENTRY:.*]]:
+; CHECK-NEXT: [[X_I32:%.*]] = sext i16 [[X]] to i32
+; CHECK-NEXT: [[X_I64:%.*]] = sext i16 [[X]] to i64
+; CHECK-NEXT: [[INVARIANT_GEP:%.*]] = getelementptr i8, ptr [[A]], i64 -8
+; CHECK-NEXT: [[SMAX20:%.*]] = call i64 @llvm.smax.i64(i64 [[X_I64]], i64 99)
+; CHECK-NEXT: [[TMP0:%.*]] = sub i64 [[SMAX20]], [[X_I64]]
+; CHECK-NEXT: [[UMIN21:%.*]] = call i64 @llvm.umin.i64(i64 [[TMP0]], i64 1)
+; CHECK-NEXT: [[TMP1:%.*]] = sub i64 [[SMAX20]], [[UMIN21]]
+; CHECK-NEXT: [[TMP2:%.*]] = sub i64 [[TMP1]], [[X_I64]]
+; CHECK-NEXT: [[TMP3:%.*]] = udiv i64 [[TMP2]], 3
+; CHECK-NEXT: [[TMP4:%.*]] = add i64 [[UMIN21]], [[TMP3]]
+; CHECK-NEXT: [[TMP5:%.*]] = add i64 [[TMP4]], 1
+; CHECK-NEXT: [[TMP6:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT: [[TMP7:%.*]] = mul i64 [[TMP6]], 8
+; CHECK-NEXT: [[TMP8:%.*]] = call i64 @llvm.umax.i64(i64 288, i64 [[TMP7]])
+; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ule i64 [[TMP5]], [[TMP8]]
+; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_SCEVCHECK:.*]]
+; CHECK: [[VECTOR_SCEVCHECK]]:
+; CHECK-NEXT: [[SMAX:%.*]] = call i64 @llvm.smax.i64(i64 [[X_I64]], i64 99)
+; CHECK-NEXT: [[TMP9:%.*]] = sub i64 [[SMAX]], [[X_I64]]
+; CHECK-NEXT: [[UMIN:%.*]] = call i64 @llvm.umin.i64(i64 [[TMP9]], i64 1)
+; CHECK-NEXT: [[TMP10:%.*]] = sub i64 [[SMAX]], [[UMIN]]
+; CHECK-NEXT: [[TMP11:%.*]] = sub i64 [[TMP10]], [[X_I64]]
+; CHECK-NEXT: [[TMP12:%.*]] = udiv i64 [[TMP11]], 3
+; CHECK-NEXT: [[TMP13:%.*]] = add i64 [[UMIN]], [[TMP12]]
+; CHECK-NEXT: [[TMP14:%.*]] = shl nsw i64 [[X_I64]], 1
+; CHECK-NEXT: [[SCEVGEP:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP14]]
+; CHECK-NEXT: [[MUL:%.*]] = call { i64, i1 } @llvm.umul.with.overflow.i64(i64 6, i64 [[TMP13]])
+; CHECK-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i64, i1 } [[MUL]], 0
+; CHECK-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i64, i1 } [[MUL]], 1
+; CHECK-NEXT: [[TMP15:%.*]] = sub i64 0, [[MUL_RESULT]]
+; CHECK-NEXT: [[TMP16:%.*]] = getelementptr i8, ptr [[SCEVGEP]], i64 [[MUL_RESULT]]
+; CHECK-NEXT: [[TMP17:%.*]] = icmp ult ptr [[TMP16]], [[SCEVGEP]]
+; CHECK-NEXT: [[TMP18:%.*]] = or i1 [[TMP17]], [[MUL_OVERFLOW]]
+; CHECK-NEXT: [[TMP19:%.*]] = shl nsw i64 [[X_I64]], 3
+; CHECK-NEXT: [[SCEVGEP1:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP19]]
+; CHECK-NEXT: [[MUL2:%.*]] = call { i64, i1 } @llvm.umul.with.overflow.i64(i64 24, i64 [[TMP13]])
+; CHECK-NEXT: [[MUL_RESULT3:%.*]] = extractvalue { i64, i1 } [[MUL2]], 0
+; CHECK-NEXT: [[MUL_OVERFLOW4:%.*]] = extractvalue { i64, i1 } [[MUL2]], 1
+; CHECK-NEXT: [[TMP20:%.*]] = sub i64 0, [[MUL_RESULT3]]
+; CHECK-NEXT: [[TMP21:%.*]] = getelementptr i8, ptr [[SCEVGEP1]], i64 [[MUL_RESULT3]]
+; CHECK-NEXT: [[TMP22:%.*]] = icmp ult ptr [[TMP21]], [[SCEVGEP1]]
+; CHECK-NEXT: [[TMP23:%.*]] = or i1 [[TMP22]], [[MUL_OVERFLOW4]]
+; CHECK-NEXT: [[TMP24:%.*]] = add nsw i64 [[TMP19]], -8
+; CHECK-NEXT: [[SCEVGEP5:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP24]]
+; CHECK-NEXT: [[MUL6:%.*]] = call { i64, i1 } @llvm.umul.with.overflow.i64(i64 24, i64 [[TMP13]])
+; CHECK-NEXT: [[MUL_RESULT7:%.*]] = extractvalue { i64, i1 } [[MUL6]], 0
+; CHECK-NEXT: [[MUL_OVERFLOW8:%.*]] = extractvalue { i64, i1 } [[MUL6]], 1
+; CHECK-NEXT: [[TMP25:%.*]] = sub i64 0, [[MUL_RESULT7]]
+; CHECK-NEXT: [[TMP26:%.*]] = getelementptr i8, ptr [[SCEVGEP5]], i64 [[MUL_RESULT7]]
+; CHECK-NEXT: [[TMP27:%.*]] = icmp ult ptr [[TMP26]], [[SCEVGEP5]]
+; CHECK-NEXT: [[TMP28:%.*]] = or i1 [[TMP27]], [[MUL_OVERFLOW8]]
+; CHECK-NEXT: [[TMP29:%.*]] = or i1 [[TMP18]], [[TMP23]]
+; CHECK-NEXT: [[TMP30:%.*]] = or i1 [[TMP29]], [[TMP28]]
+; CHECK-NEXT: br i1 [[TMP30]], label %[[SCALAR_PH]], label %[[VECTOR_MEMCHECK:.*]]
+; CHECK: [[VECTOR_MEMCHECK]]:
+; CHECK-NEXT: [[TMP31:%.*]] = shl nsw i64 [[X_I64]], 1
+; CHECK-NEXT: [[SCEVGEP9:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP31]]
+; CHECK-NEXT: [[SMAX10:%.*]] = call i64 @llvm.smax.i64(i64 [[X_I64]], i64 99)
+; CHECK-NEXT: [[TMP32:%.*]] = sub i64 [[SMAX10]], [[X_I64]]
+; CHECK-NEXT: [[UMIN11:%.*]] = call i64 @llvm.umin.i64(i64 [[TMP32]], i64 1)
+; CHECK-NEXT: [[TMP33:%.*]] = sub i64 [[SMAX10]], [[UMIN11]]
+; CHECK-NEXT: [[TMP34:%.*]] = sub i64 [[TMP33]], [[X_I64]]
+; CHECK-NEXT: [[TMP35:%.*]] = udiv i64 [[TMP34]], 3
+; CHECK-NEXT: [[TMP36:%.*]] = add i64 [[UMIN11]], [[TMP35]]
+; CHECK-NEXT: [[TMP37:%.*]] = mul i64 [[TMP36]], 6
+; CHECK-NEXT: [[TMP38:%.*]] = add i64 [[TMP37]], [[TMP31]]
+; CHECK-NEXT: [[TMP39:%.*]] = add i64 [[TMP38]], 2
+; CHECK-NEXT: [[SCEVGEP12:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP39]]
+; CHECK-NEXT: [[TMP40:%.*]] = shl nsw i64 [[X_I64]], 3
+; CHECK-NEXT: [[SCEVGEP13:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP40]]
+; CHECK-NEXT: [[TMP41:%.*]] = mul i64 [[TMP36]], 24
+; CHECK-NEXT: [[TMP42:%.*]] = add i64 [[TMP41]], [[TMP40]]
+; CHECK-NEXT: [[TMP43:%.*]] = add i64 [[TMP42]], 8
+; CHECK-NEXT: [[SCEVGEP14:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP43]]
+; CHECK-NEXT: [[TMP44:%.*]] = add nsw i64 [[TMP40]], -8
+; CHECK-NEXT: [[SCEVGEP15:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP44]]
+; CHECK-NEXT: [[SCEVGEP16:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP42]]
+; CHECK-NEXT: [[BOUND0:%.*]] = icmp ult ptr [[SCEVGEP9]], [[SCEVGEP14]]
+; CHECK-NEXT: [[BOUND1:%.*]] = icmp ult ptr [[SCEVGEP13]], [[SCEVGEP12]]
+; CHECK-NEXT: [[FOUND_CONFLICT:%.*]] = and i1 [[BOUND0]], [[BOUND1]]
+; CHECK-NEXT: [[BOUND017:%.*]] = icmp ult ptr [[SCEVGEP9]], [[SCEVGEP16]]
+; CHECK-NEXT: [[BOUND118:%.*]] = icmp ult ptr [[SCEVGEP15]], [[SCEVGEP12]]
+; CHECK-NEXT: [[FOUND_CONFLICT19:%.*]] = and i1 [[BOUND017]], [[BOUND118]]
+; CHECK-NEXT: [[CONFLICT_RDX:%.*]] = or i1 [[FOUND_CONFLICT]], [[FOUND_CONFLICT19]]
+; CHECK-NEXT: br i1 [[CONFLICT_RDX]], label %[[SCALAR_PH]], label %[[VECTOR_PH:.*]]
+; CHECK: [[VECTOR_PH]]:
+; CHECK-NEXT: [[TMP45:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT: [[TMP46:%.*]] = mul i64 [[TMP45]], 8
+; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[TMP5]], [[TMP46]]
+; CHECK-NEXT: [[TMP47:%.*]] = icmp eq i64 [[N_MOD_VF]], 0
+; CHECK-NEXT: [[TMP48:%.*]] = select i1 [[TMP47]], i64 [[TMP46]], i64 [[N_MOD_VF]]
+; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[TMP5]], [[TMP48]]
+; CHECK-NEXT: [[TMP49:%.*]] = mul i64 [[N_VEC]], 3
+; CHECK-NEXT: [[IND_END:%.*]] = add i64 [[X_I64]], [[TMP49]]
+; CHECK-NEXT: [[DOTCAST:%.*]] = trunc i64 [[N_VEC]] to i32
+; CHECK-NEXT: [[TMP50:%.*]] = mul i32 [[DOTCAST]], 3
+; CHECK-NEXT: [[IND_END22:%.*]] = add i32 [[X_I32]], [[TMP50]]
+; CHECK-NEXT: [[TMP51:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT: [[TMP52:%.*]] = mul i64 [[TMP51]], 8
+; CHECK-NEXT: [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 8 x i64> poison, i64 [[X_I64]], i64 0
+; CHECK-NEXT: [[DOTSPLAT:%.*]] = shufflevector <vscale x 8 x i64> [[DOTSPLATINSERT]], <vscale x 8 x i64> poison, <vscale x 8 x i32> zeroinitializer
+; CHECK-NEXT: [[TMP53:%.*]] = call <vscale x 8 x i64> @llvm.stepvector.nxv8i64()
+; CHECK-NEXT: [[TMP54:%.*]] = add <vscale x 8 x i64> [[TMP53]], zeroinitializer
+; CHECK-NEXT: [[TMP55:%.*]] = mul <vscale x 8 x i64> [[TMP54]], shufflevector (<vscale x 8 x i64> insertelement (<vscale x 8 x i64> poison, i64 3, i64 0), <vscale x 8 x i64> poison, <vscale x 8 x i32> zeroinitializer)
+; CHECK-NEXT: [[INDUCTION:%.*]] = add <vscale x 8 x i64> [[DOTSPLAT]], [[TMP55]]
+; CHECK-NEXT: [[TMP56:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT: [[TMP57:%.*]] = mul i64 [[TMP56]], 8
+; CHECK-NEXT: [[TMP58:%.*]] = mul i64 3, [[TMP57]]
+; CHECK-NEXT: [[DOTSPLATINSERT24:%.*]] = insertelement <vscale x 8 x i64> poison, i64 [[TMP58]], i64 0
+; CHECK-NEXT: [[DOTSPLAT25:%.*]] = shufflevector <vscale x 8 x i64> [[DOTSPLATINSERT24]], <vscale x 8 x i64> poison, <vscale x 8 x i32> zeroinitializer
+; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
+; CHECK: [[VECTOR_BODY]]:
+; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT: [[VEC_IND:%.*]] = phi <vscale x 8 x i64> [ [[INDUCTION]], %[[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT: [[TMP59:%.*]] = getelementptr i16, ptr [[A]], <vscale x 8 x i64> [[VEC_IND]]
+; CHECK-NEXT: call void @llvm.masked.scatter.nxv8i16.nxv8p0(<vscale x 8 x i16> zeroinitializer, <vscale x 8 x ptr> [[TMP59]], i32 2, <vscale x 8 x i1> shufflevector (<vscale x 8 x i1> insertelement (<vscale x 8 x i1> poison, i1 true, i64 0), <vscale x 8 x i1> poison, <vscale x 8 x i32> zeroinitializer)), !alias.scope [[META0:![0-9]+]], !noalias [[META3:![0-9]+]]
+; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP52]]
+; CHECK-NEXT: [[VEC_IND_NEXT]] = add <vscale x 8 x i64> [[VEC_IND]], [[DOTSPLAT25]]
+; CHECK-NEXT: [[TMP60:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-NEXT: br i1 [[TMP60]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP6:![0-9]+]]
+; CHECK: [[MIDDLE_BLOCK]]:
+; CHECK-NEXT: br label %[[SCALAR_PH]]
+; CHECK: [[SCALAR_PH]]:
+; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[IND_END]], %[[MIDDLE_BLOCK]] ], [ [[X_I64]], %[[ENTRY]] ], [ [[X_I64]], %[[VECTOR_SCEVCHECK]] ], [ [[X_I64]], %[[VECTOR_MEMCHECK]] ]
+; CHECK-NEXT: [[BC_RESUME_VAL23:%.*]] = phi i32 [ [[IND_END22]], %[[MIDDLE_BLOCK]] ], [ [[X_I32]], %[[ENTRY]] ], [ [[X_I32]], %[[VECTOR_SCEVCHECK]] ], [ [[X_I32]], %[[VECTOR_MEMCHECK]] ]
+; CHECK-NEXT: br label %[[LOOP:.*]]
+; CHECK: [[LOOP]]:
+; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT: [[IV_CONV:%.*]] = phi i32 [ [[BC_RESUME_VAL23]], %[[SCALAR_PH]] ], [ [[TMP64:%.*]], %[[LOOP]] ]
+; CHECK-NEXT: [[GEP_I64:%.*]] = getelementptr i64, ptr [[A]], i64 [[IV]]
+; CHECK-NEXT: [[TMP61:%.*]] = load i64, ptr [[GEP_I64]], align 8
+; CHECK-NEXT: [[TMP62:%.*]] = sext i32 [[IV_CONV]] to i64
+; CHECK-NEXT: [[GEP_CONV:%.*]] = getelementptr i64, ptr [[INVARIANT_GEP]], i64 [[TMP62]]
+; CHECK-NEXT: [[TMP63:%.*]] = load i64, ptr [[GEP_CONV]], align 8
+; CHECK-NEXT: [[GEP_I16:%.*]] = getelementptr i16, ptr [[A]], i64 [[IV]]
+; CHECK-NEXT: store i16 0, ptr [[GEP_I16]], align 2
+; CHECK-NEXT: [[IV_NEXT]] = add i64 [[IV]], 3
+; CHECK-NEXT: [[TMP64]] = trunc i64 [[IV_NEXT]] to i32
+; CHECK-NEXT: [[C:%.*]] = icmp slt i64 [[IV]], 99
+; CHECK-NEXT: br i1 [[C]], label %[[LOOP]], label %[[EXIT:.*]], !llvm.loop [[LOOP9:![0-9]+]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: ret void
+;
+entry:
+ %x.i32 = sext i16 %x to i32
+ %x.i64 = sext i16 %x to i64
+ %invariant.gep = getelementptr i8, ptr %A, i64 -8
+ br label %loop
+
+loop:
+ %iv = phi i64 [ %x.i64, %entry ], [ %iv.next, %loop ]
+ %iv.conv = phi i32 [ %x.i32, %entry ], [ %5, %loop ]
+ %gep.i64 = getelementptr i64, ptr %A, i64 %iv
+ %2 = load i64, ptr %gep.i64, align 8
+ %3 = sext i32 %iv.conv to i64
+ %gep.conv = getelementptr i64, ptr %invariant.gep, i64 %3
+ %4 = load i64, ptr %gep.conv, align 8
+ %gep.i16 = getelementptr i16, ptr %A, i64 %iv
+ store i16 0, ptr %gep.i16, align 2
+ %iv.next = add i64 %iv, 3
+ %5 = trunc i64 %iv.next to i32
+ %c = icmp slt i64 %iv, 99
+ br i1 %c, label %loop, label %exit
+
+exit:
+ ret void
+}
+
+attributes #0 = { "target-features"="+64bit,+v,+zvl256b" }
+;.
+; CHECK: [[META0]] = !{[[META1:![0-9]+]]}
+; CHECK: [[META1]] = distinct !{[[META1]], [[META2:![0-9]+]]}
+; CHECK: [[META2]] = distinct !{[[META2]], !"LVerDomain"}
+; CHECK: [[META3]] = !{[[META4:![0-9]+]], [[META5:![0-9]+]]}
+; CHECK: [[META4]] = distinct !{[[META4]], [[META2]]}
+; CHECK: [[META5]] = distinct !{[[META5]], [[META2]]}
+; CHECK: [[LOOP6]] = distinct !{[[LOOP6]], [[META7:![0-9]+]], [[META8:![0-9]+]]}
+; CHECK: [[META7]] = !{!"llvm.loop.isvectorized", i32 1}
+; CHECK: [[META8]] = !{!"llvm.loop.unroll.runtime.disable"}
+; CHECK: [[LOOP9]] = distinct !{[[LOOP9]], [[META7]]}
+;.
More information about the llvm-commits
mailing list