Skip to content

Commit 4a7f60d

Browse files
lukel97lukel97
authored
[VPlan] Handle VPWidenCastRecipe without underlying value in EVL transform (#120194)
This fixes a crash that shows up when building SPEC CPU 2017 with EVL tail folding on RISC-V. A VPWidenCastRecipe doesn't always have an underlying value, and in the case of this crash this happens whenever a widened cast is created via truncateToMinimalBitwidths. Fix this by just using the opcode stored in the recipe itself. I think a similar issue exists with VPWidenIntrinsicRecipe and how it's widened, but I haven't run into any crashes with it just yet.
1 parent b24caf3 commit 4a7f60d

File tree

2 files changed

+91
-8
lines changed

2 files changed

+91
-8
lines changed

‎llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp

Lines changed: 7 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -1505,24 +1505,23 @@ static void transformRecipestoEVLRecipes(VPlan &Plan, VPValue &EVL) {
15051505
CInst->getDebugLoc());
15061506
})
15071507
.Case<VPWidenCastRecipe>(
1508-
[&](VPWidenCastRecipe *CInst) -> VPRecipeBase * {
1509-
auto *CI = dyn_cast<CastInst>(CInst->getUnderlyingInstr());
1508+
[&](VPWidenCastRecipe *CastR) -> VPRecipeBase * {
15101509
Intrinsic::ID VPID =
1511-
VPIntrinsic::getForOpcode(CI->getOpcode());
1510+
VPIntrinsic::getForOpcode(CastR->getOpcode());
15121511
assert(VPID != Intrinsic::not_intrinsic &&
15131512
"Expected vp.casts Instrinsic");
15141513

1515-
SmallVector<VPValue *> Ops(CInst->operands());
1514+
SmallVector<VPValue *> Ops(CastR->operands());
15161515
assert(VPIntrinsic::getMaskParamPos(VPID) &&
15171516
VPIntrinsic::getVectorLengthParamPos(VPID) &&
15181517
"Expected VP intrinsic");
1519-
VPValue *Mask = Plan.getOrAddLiveIn(ConstantInt::getTrue(
1520-
IntegerType::getInt1Ty(CI->getContext())));
1518+
VPValue *Mask =
1519+
Plan.getOrAddLiveIn(ConstantInt::getTrue(Ctx));
15211520
Ops.push_back(Mask);
15221521
Ops.push_back(&EVL);
15231522
return new VPWidenIntrinsicRecipe(
1524-
VPID, Ops, TypeInfo.inferScalarType(CInst),
1525-
CInst->getDebugLoc());
1523+
VPID, Ops, TypeInfo.inferScalarType(CastR),
1524+
CastR->getDebugLoc());
15261525
})
15271526
.Case<VPWidenSelectRecipe>([&](VPWidenSelectRecipe *Sel) {
15281527
SmallVector<VPValue *> Ops(Sel->operands());

‎llvm/test/Transforms/LoopVectorize/RISCV/truncate-to-minimal-bitwidth-evl-crash.ll

Lines changed: 84 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,84 @@
1+
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
2+
; RUN: opt -passes=loop-vectorize -force-tail-folding-style=data-with-evl -prefer-predicate-over-epilogue=predicate-dont-vectorize -mtriple=riscv64 -mattr=+v -S %s | FileCheck %s
3+
4+
; Make sure we don't crash when transforming a VPWidenCastRecipe created without
5+
; an underlying value to an EVL recipe. This occurs in this test via
6+
; VPlanTransforms::truncateToMinimalBitwidths
7+
8+
define void @truncate_to_minimal_bitwidths_widen_cast_recipe(ptr %src) {
9+
; CHECK-LABEL: define void @truncate_to_minimal_bitwidths_widen_cast_recipe(
10+
; CHECK-SAME: ptr [[SRC:%.*]]) #[[ATTR0:[0-9]+]] {
11+
; CHECK-NEXT: [[ENTRY:.*]]:
12+
; CHECK-NEXT: br i1 false, label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
13+
; CHECK: [[VECTOR_PH]]:
14+
; CHECK-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
15+
; CHECK-NEXT: [[TMP1:%.*]] = sub i64 [[TMP0]], 1
16+
; CHECK-NEXT: [[N_RND_UP:%.*]] = add i64 2, [[TMP1]]
17+
; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N_RND_UP]], [[TMP0]]
18+
; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[N_RND_UP]], [[N_MOD_VF]]
19+
; CHECK-NEXT: [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
20+
; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
21+
; CHECK: [[VECTOR_BODY]]:
22+
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
23+
; CHECK-NEXT: [[EVL_BASED_IV:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_EVL_NEXT:%.*]], %[[VECTOR_BODY]] ]
24+
; CHECK-NEXT: [[AVL:%.*]] = sub i64 2, [[EVL_BASED_IV]]
25+
; CHECK-NEXT: [[TMP3:%.*]] = call i32 @llvm.experimental.get.vector.length.i64(i64 [[AVL]], i32 1, i1 true)
26+
; CHECK-NEXT: [[TMP4:%.*]] = add i64 [[EVL_BASED_IV]], 0
27+
; CHECK-NEXT: [[TMP5:%.*]] = getelementptr i8, ptr [[SRC]], i64 [[TMP4]]
28+
; CHECK-NEXT: [[TMP6:%.*]] = getelementptr i8, ptr [[TMP5]], i32 0
29+
; CHECK-NEXT: [[VP_OP_LOAD:%.*]] = call <vscale x 1 x i8> @llvm.vp.load.nxv1i8.p0(ptr align 1 [[TMP6]], <vscale x 1 x i1> splat (i1 true), i32 [[TMP3]])
30+
; CHECK-NEXT: [[TMP7:%.*]] = call <vscale x 1 x i16> @llvm.vp.zext.nxv1i16.nxv1i8(<vscale x 1 x i8> [[VP_OP_LOAD]], <vscale x 1 x i1> splat (i1 true), i32 [[TMP3]])
31+
; CHECK-NEXT: [[VP_OP:%.*]] = call <vscale x 1 x i16> @llvm.vp.mul.nxv1i16(<vscale x 1 x i16> zeroinitializer, <vscale x 1 x i16> [[TMP7]], <vscale x 1 x i1> splat (i1 true), i32 [[TMP3]])
32+
; CHECK-NEXT: [[VP_OP1:%.*]] = call <vscale x 1 x i16> @llvm.vp.lshr.nxv1i16(<vscale x 1 x i16> [[VP_OP]], <vscale x 1 x i16> trunc (<vscale x 1 x i32> splat (i32 1) to <vscale x 1 x i16>), <vscale x 1 x i1> splat (i1 true), i32 [[TMP3]])
33+
; CHECK-NEXT: [[TMP8:%.*]] = call <vscale x 1 x i8> @llvm.vp.trunc.nxv1i8.nxv1i16(<vscale x 1 x i16> [[VP_OP1]], <vscale x 1 x i1> splat (i1 true), i32 [[TMP3]])
34+
; CHECK-NEXT: call void @llvm.vp.scatter.nxv1i8.nxv1p0(<vscale x 1 x i8> [[TMP8]], <vscale x 1 x ptr> align 1 zeroinitializer, <vscale x 1 x i1> splat (i1 true), i32 [[TMP3]])
35+
; CHECK-NEXT: [[TMP9:%.*]] = zext i32 [[TMP3]] to i64
36+
; CHECK-NEXT: [[INDEX_EVL_NEXT]] = add nuw i64 [[TMP9]], [[EVL_BASED_IV]]
37+
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP2]]
38+
; CHECK-NEXT: [[TMP10:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
39+
; CHECK-NEXT: br i1 [[TMP10]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
40+
; CHECK: [[MIDDLE_BLOCK]]:
41+
; CHECK-NEXT: br i1 true, label %[[EXIT:.*]], label %[[SCALAR_PH]]
42+
; CHECK: [[SCALAR_PH]]:
43+
; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
44+
; CHECK-NEXT: br label %[[LOOP:.*]]
45+
; CHECK: [[LOOP]]:
46+
; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
47+
; CHECK-NEXT: [[GEP_SRC:%.*]] = getelementptr i8, ptr [[SRC]], i64 [[IV]]
48+
; CHECK-NEXT: [[TMP11:%.*]] = load i8, ptr [[GEP_SRC]], align 1
49+
; CHECK-NEXT: [[CONV:%.*]] = zext i8 [[TMP11]] to i32
50+
; CHECK-NEXT: [[MUL16:%.*]] = mul i32 0, [[CONV]]
51+
; CHECK-NEXT: [[SHR35:%.*]] = lshr i32 [[MUL16]], 1
52+
; CHECK-NEXT: [[CONV36:%.*]] = trunc i32 [[SHR35]] to i8
53+
; CHECK-NEXT: store i8 [[CONV36]], ptr null, align 1
54+
; CHECK-NEXT: [[IV_NEXT]] = add i64 [[IV]], 1
55+
; CHECK-NEXT: [[EC:%.*]] = icmp eq i64 [[IV]], 1
56+
; CHECK-NEXT: br i1 [[EC]], label %[[EXIT]], label %[[LOOP]], !llvm.loop [[LOOP3:![0-9]+]]
57+
; CHECK: [[EXIT]]:
58+
; CHECK-NEXT: ret void
59+
;
60+
entry:
61+
br label %loop
62+
63+
loop: ; preds = %loop, %entry
64+
%iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
65+
%gep.src = getelementptr i8, ptr %src, i64 %iv
66+
%0 = load i8, ptr %gep.src, align 1
67+
%conv = zext i8 %0 to i32
68+
%mul16 = mul i32 0, %conv
69+
%shr35 = lshr i32 %mul16, 1
70+
%conv36 = trunc i32 %shr35 to i8
71+
store i8 %conv36, ptr null, align 1
72+
%iv.next = add i64 %iv, 1
73+
%ec = icmp eq i64 %iv, 1
74+
br i1 %ec, label %exit, label %loop
75+
76+
exit: ; preds = %loop
77+
ret void
78+
}
79+
;.
80+
; CHECK: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}
81+
; CHECK: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}
82+
; CHECK: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}
83+
; CHECK: [[LOOP3]] = distinct !{[[LOOP3]], [[META2]], [[META1]]}
84+
;.

0 commit comments

Comments
 (0)