[VPlan] Preserve underlying value when EVL transforming intrinsics/casts #121552
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This fixes a crash that shows up with EVL tail folding due to a mismatch between the legacy and VPlan cost models. In the legacy model a cast instruction may have its cost ignored if it's part of a reduction/induction variable. The VPlan cost model will check if the underying value of a recipe is skipped in Ctx.skipCostComputation, so this works for normally widened casts. However when transformed to a VP intrinsic we drop the underlying value and we now end up costing the previously skipped values, which leads to the mismatch. This fixes the crash by preserving the underlying cast instruction. It means that a VPWidenIntrinsicRecipe's underlying value might not always be a CallInst anymore, so there was one place where I had to relax a cast<> to a dyn_cast<>. I wasn't able to provide a test case for the crash itself since we can't yet do EVL tail folding with widened IVs, but it will become a problem later down the line. , but only after induction variable widening support is enabled so I'm not really able to provide
lukel97
requested review from
Mel-Chen,
alexey-bataev,
arcbbb,
fhahn and
LiqinWeng
|
@llvm/pr-subscribers-vectorizers @llvm/pr-subscribers-llvm-transforms Author: Luke Lau (lukel97) ChangesThis fixes a crash that shows up with EVL tail folding due to a mismatch between the legacy and VPlan cost models. In the legacy model a cast instruction may have its cost ignored if it's part of a reduction/induction variable. The VPlan cost model will check if the underying value of a recipe is skipped in Ctx.skipCostComputation, so this works for normally widened casts. However when transformed to a VP intrinsic we drop the underlying value and we now end up costing the previously skipped values, which leads to the mismatch. This fixes the crash by preserving the underlying cast instruction. It means that a VPWidenIntrinsicRecipe's underlying value might not always be a CallInst anymore, so there was one place where I had to relax a cast<> to a dyn_cast<>. I wasn't able to provide a test case for the crash itself since we can't yet do EVL tail folding with widened IVs, but it will become a problem later down the line. Patch is 23.35 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/121552.diff 4 Files Affected:
diff --git a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
index 77c08839dbfa95..c14146839e5e11 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
@@ -985,7 +985,7 @@ void VPWidenIntrinsicRecipe::execute(VPTransformState &State) {
assert(VectorF &&
"Can't retrieve vector intrinsic or vector-predication intrinsics.");
- auto *CI = cast_or_null<CallInst>(getUnderlyingValue());
+ auto *CI = dyn_cast_if_present<CallInst>(getUnderlyingValue());
SmallVector<OperandBundleDef, 1> OpBundles;
if (CI)
CI->getOperandBundlesAsDefs(OpBundles);
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index 8ac2bd5160c267..8751d626f86a12 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -1506,8 +1506,12 @@ static VPRecipeBase *createEVLRecipe(VPValue *HeaderMask,
SmallVector<VPValue *> Ops(CR->operands());
Ops.push_back(&AllOneMask);
Ops.push_back(&EVL);
- return new VPWidenIntrinsicRecipe(
+ auto *VPR = new VPWidenIntrinsicRecipe(
VPID, Ops, TypeInfo.inferScalarType(CR), CR->getDebugLoc());
+ // Preserve the underlying value as some cast instructions may have
+ // their cost skipped in collectValuesToIgnore
+ VPR->setUnderlyingValue(CR->getUnderlyingValue());
+ return VPR;
})
.Case<VPWidenSelectRecipe>([&](VPWidenSelectRecipe *Sel) {
SmallVector<VPValue *> Ops(Sel->operands());
diff --git a/llvm/test/Transforms/LoopVectorize/RISCV/vplan-vp-call-intrinsics.ll b/llvm/test/Transforms/LoopVectorize/RISCV/vplan-vp-call-intrinsics.ll
index 05c687a2053533..a318eb4406d319 100644
--- a/llvm/test/Transforms/LoopVectorize/RISCV/vplan-vp-call-intrinsics.ll
+++ b/llvm/test/Transforms/LoopVectorize/RISCV/vplan-vp-call-intrinsics.ll
@@ -27,10 +27,10 @@ define void @vp_smax(ptr %a, ptr %b, ptr %c, i64 %N) {
; IF-EVL-NEXT: CLONE ir<[[GEP2:%.+]]> = getelementptr inbounds ir<%c>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR2:%[0-9]+]]> = vector-pointer ir<[[GEP2]]>
; IF-EVL-NEXT: WIDEN ir<[[LD2:%.+]]> = vp.load vp<[[PTR2]]>, vp<[[EVL]]>
-; IF-EVL-NEXT: WIDEN-INTRINSIC vp<[[SMAX:%.+]]> = call llvm.vp.smax(ir<[[LD1]]>, ir<[[LD2]]>, ir<true>, vp<[[EVL]]>)
+; IF-EVL-NEXT: WIDEN-INTRINSIC ir<[[SMAX:%.+]]> = call llvm.vp.smax(ir<[[LD1]]>, ir<[[LD2]]>, ir<true>, vp<[[EVL]]>)
; IF-EVL-NEXT: CLONE ir<[[GEP3:%.+]]> = getelementptr inbounds ir<%a>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR3:%[0-9]+]]> = vector-pointer ir<[[GEP3]]>
-; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR3]]>, vp<[[SMAX]]>, vp<[[EVL]]>
+; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR3]]>, ir<[[SMAX]]>, vp<[[EVL]]>
; IF-EVL-NEXT: SCALAR-CAST vp<[[CAST:%[0-9]+]]> = zext vp<[[EVL]]> to i64
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT]]> = add vp<[[CAST]]>, vp<[[EVL_PHI]]>
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT_EXIT:%.+]]> = add vp<[[IV]]>, vp<[[VFUF]]>
@@ -80,10 +80,10 @@ define void @vp_smin(ptr %a, ptr %b, ptr %c, i64 %N) {
; IF-EVL-NEXT: CLONE ir<[[GEP2:%.+]]> = getelementptr inbounds ir<%c>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR2:%[0-9]+]]> = vector-pointer ir<[[GEP2]]>
; IF-EVL-NEXT: WIDEN ir<[[LD2:%.+]]> = vp.load vp<[[PTR2]]>, vp<[[EVL]]>
-; IF-EVL-NEXT: WIDEN-INTRINSIC vp<[[SMIN:%.+]]> = call llvm.vp.smin(ir<[[LD1]]>, ir<[[LD2]]>, ir<true>, vp<[[EVL]]>)
+; IF-EVL-NEXT: WIDEN-INTRINSIC ir<[[SMIN:%.+]]> = call llvm.vp.smin(ir<[[LD1]]>, ir<[[LD2]]>, ir<true>, vp<[[EVL]]>)
; IF-EVL-NEXT: CLONE ir<[[GEP3:%.+]]> = getelementptr inbounds ir<%a>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR3:%[0-9]+]]> = vector-pointer ir<[[GEP3]]>
-; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR3]]>, vp<[[SMIN]]>, vp<[[EVL]]>
+; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR3]]>, ir<[[SMIN]]>, vp<[[EVL]]>
; IF-EVL-NEXT: SCALAR-CAST vp<[[CAST:%[0-9]+]]> = zext vp<[[EVL]]> to i64
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT]]> = add vp<[[CAST]]>, vp<[[EVL_PHI]]>
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT_EXIT:%.+]]> = add vp<[[IV]]>, vp<[[VFUF]]>
@@ -133,10 +133,10 @@ define void @vp_umax(ptr %a, ptr %b, ptr %c, i64 %N) {
; IF-EVL-NEXT: CLONE ir<[[GEP2:%.+]]> = getelementptr inbounds ir<%c>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR2:%[0-9]+]]> = vector-pointer ir<[[GEP2]]>
; IF-EVL-NEXT: WIDEN ir<[[LD2:%.+]]> = vp.load vp<[[PTR2]]>, vp<[[EVL]]>
-; IF-EVL-NEXT: WIDEN-INTRINSIC vp<[[UMAX:%.+]]> = call llvm.vp.umax(ir<[[LD1]]>, ir<[[LD2]]>, ir<true>, vp<[[EVL]]>)
+; IF-EVL-NEXT: WIDEN-INTRINSIC ir<[[UMAX:%.+]]> = call llvm.vp.umax(ir<[[LD1]]>, ir<[[LD2]]>, ir<true>, vp<[[EVL]]>)
; IF-EVL-NEXT: CLONE ir<[[GEP3:%.+]]> = getelementptr inbounds ir<%a>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR3:%[0-9]+]]> = vector-pointer ir<[[GEP3]]>
-; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR3]]>, vp<[[UMAX]]>, vp<[[EVL]]>
+; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR3]]>, ir<[[UMAX]]>, vp<[[EVL]]>
; IF-EVL-NEXT: SCALAR-CAST vp<[[CAST:%[0-9]+]]> = zext vp<[[EVL]]> to i64
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT]]> = add vp<[[CAST]]>, vp<[[EVL_PHI]]>
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT_EXIT:%.+]]> = add vp<[[IV]]>, vp<[[VFUF]]>
@@ -186,10 +186,10 @@ define void @vp_umin(ptr %a, ptr %b, ptr %c, i64 %N) {
; IF-EVL-NEXT: CLONE ir<[[GEP2:%.+]]> = getelementptr inbounds ir<%c>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR2:%[0-9]+]]> = vector-pointer ir<[[GEP2]]>
; IF-EVL-NEXT: WIDEN ir<[[LD2:%.+]]> = vp.load vp<[[PTR2]]>, vp<[[EVL]]>
-; IF-EVL-NEXT: WIDEN-INTRINSIC vp<[[UMIN:%.+]]> = call llvm.vp.umin(ir<[[LD1]]>, ir<[[LD2]]>, ir<true>, vp<[[EVL]]>)
+; IF-EVL-NEXT: WIDEN-INTRINSIC ir<[[UMIN:%.+]]> = call llvm.vp.umin(ir<[[LD1]]>, ir<[[LD2]]>, ir<true>, vp<[[EVL]]>)
; IF-EVL-NEXT: CLONE ir<[[GEP3:%.+]]> = getelementptr inbounds ir<%a>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR3:%[0-9]+]]> = vector-pointer ir<[[GEP3]]>
-; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR3]]>, vp<[[UMIN]]>, vp<[[EVL]]>
+; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR3]]>, ir<[[UMIN]]>, vp<[[EVL]]>
; IF-EVL-NEXT: SCALAR-CAST vp<[[CAST:%[0-9]+]]> = zext vp<[[EVL]]> to i64
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT]]> = add vp<[[CAST]]>, vp<[[EVL_PHI]]>
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT_EXIT:%.+]]> = add vp<[[IV]]>, vp<[[VFUF]]>
@@ -236,10 +236,10 @@ define void @vp_ctlz(ptr %a, ptr %b, i64 %N) {
; IF-EVL-NEXT: CLONE ir<[[GEP1:%.+]]> = getelementptr inbounds ir<%b>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR1:%[0-9]+]]> = vector-pointer ir<[[GEP1]]>
; IF-EVL-NEXT: WIDEN ir<[[LD1:%.+]]> = vp.load vp<[[PTR1]]>, vp<[[EVL]]>
-; IF-EVL-NEXT: WIDEN-INTRINSIC vp<[[CTLZ:%.+]]> = call llvm.vp.ctlz(ir<[[LD1]]>, ir<true>, ir<true>, vp<[[EVL]]>)
+; IF-EVL-NEXT: WIDEN-INTRINSIC ir<[[CTLZ:%.+]]> = call llvm.vp.ctlz(ir<[[LD1]]>, ir<true>, ir<true>, vp<[[EVL]]>)
; IF-EVL-NEXT: CLONE ir<[[GEP2:%.+]]> = getelementptr inbounds ir<%a>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR2:%[0-9]+]]> = vector-pointer ir<[[GEP2]]>
-; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR2]]>, vp<[[CTLZ]]>, vp<[[EVL]]>
+; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR2]]>, ir<[[CTLZ]]>, vp<[[EVL]]>
; IF-EVL-NEXT: SCALAR-CAST vp<[[CAST:%[0-9]+]]> = zext vp<[[EVL]]> to i64
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT]]> = add vp<[[CAST]]>, vp<[[EVL_PHI]]>
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT_EXIT:%.+]]> = add vp<[[IV]]>, vp<[[VFUF]]>
@@ -284,10 +284,10 @@ define void @vp_cttz(ptr %a, ptr %b, i64 %N) {
; IF-EVL-NEXT: CLONE ir<[[GEP1:%.+]]> = getelementptr inbounds ir<%b>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR1:%[0-9]+]]> = vector-pointer ir<[[GEP1]]>
; IF-EVL-NEXT: WIDEN ir<[[LD1:%.+]]> = vp.load vp<[[PTR1]]>, vp<[[EVL]]>
-; IF-EVL-NEXT: WIDEN-INTRINSIC vp<[[CTTZ:%.+]]> = call llvm.vp.cttz(ir<[[LD1]]>, ir<true>, ir<true>, vp<[[EVL]]>)
+; IF-EVL-NEXT: WIDEN-INTRINSIC ir<[[CTTZ:%.+]]> = call llvm.vp.cttz(ir<[[LD1]]>, ir<true>, ir<true>, vp<[[EVL]]>)
; IF-EVL-NEXT: CLONE ir<[[GEP2:%.+]]> = getelementptr inbounds ir<%a>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR2:%[0-9]+]]> = vector-pointer ir<[[GEP2]]>
-; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR2]]>, vp<[[CTTZ]]>, vp<[[EVL]]>
+; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR2]]>, ir<[[CTTZ]]>, vp<[[EVL]]>
; IF-EVL-NEXT: SCALAR-CAST vp<[[CAST:%[0-9]+]]> = zext vp<[[EVL]]> to i64
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT]]> = add vp<[[CAST]]>, vp<[[EVL_PHI]]>
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT_EXIT:%.+]]> = add vp<[[IV]]>, vp<[[VFUF]]>
@@ -332,12 +332,12 @@ define void @vp_lrint(ptr %a, ptr %b, i64 %N) {
; IF-EVL-NEXT: CLONE ir<[[GEP1:%.+]]> = getelementptr inbounds ir<%b>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR1:%[0-9]+]]> = vector-pointer ir<[[GEP1]]>
; IF-EVL-NEXT: WIDEN ir<[[LD1:%.+]]> = vp.load vp<[[PTR1]]>, vp<[[EVL]]>
-; IF-EVL-NEXT: WIDEN-INTRINSIC vp<[[FPEXT:%.+]]> = call llvm.vp.fpext(ir<[[LD1]]>, ir<true>, vp<[[EVL]]>)
-; IF-EVL-NEXT: WIDEN-INTRINSIC vp<[[LRINT:%.+]]> = call llvm.vp.lrint(vp<[[FPEXT]]>, ir<true>, vp<[[EVL]]>)
-; IF-EVL-NEXT: WIDEN-INTRINSIC vp<[[TRUNC:%.+]]> = call llvm.vp.trunc(vp<[[LRINT]]>, ir<true>, vp<[[EVL]]>)
+; IF-EVL-NEXT: WIDEN-INTRINSIC ir<[[FPEXT:%.+]]> = call llvm.vp.fpext(ir<[[LD1]]>, ir<true>, vp<[[EVL]]>)
+; IF-EVL-NEXT: WIDEN-INTRINSIC ir<[[LRINT:%.+]]> = call llvm.vp.lrint(ir<[[FPEXT]]>, ir<true>, vp<[[EVL]]>)
+; IF-EVL-NEXT: WIDEN-INTRINSIC ir<[[TRUNC:%.+]]> = call llvm.vp.trunc(ir<[[LRINT]]>, ir<true>, vp<[[EVL]]>)
; IF-EVL-NEXT: CLONE ir<[[GEP2:%.+]]> = getelementptr inbounds ir<%a>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR2:%[0-9]+]]> = vector-pointer ir<[[GEP2]]>
-; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR2]]>, vp<[[TRUNC]]>, vp<[[EVL]]>
+; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR2]]>, ir<[[TRUNC]]>, vp<[[EVL]]>
; IF-EVL-NEXT: SCALAR-CAST vp<[[CAST:%[0-9]+]]> = zext vp<[[EVL]]> to i64
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT]]> = add vp<[[CAST]]>, vp<[[EVL_PHI]]>
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT_EXIT:%.+]]> = add vp<[[IV]]>, vp<[[VFUF]]>
@@ -384,12 +384,12 @@ define void @vp_llrint(ptr %a, ptr %b, i64 %N) {
; IF-EVL-NEXT: CLONE ir<[[GEP1:%.+]]> = getelementptr inbounds ir<%b>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR1:%[0-9]+]]> = vector-pointer ir<[[GEP1]]>
; IF-EVL-NEXT: WIDEN ir<[[LD1:%.+]]> = vp.load vp<[[PTR1]]>, vp<[[EVL]]>
-; IF-EVL-NEXT: WIDEN-INTRINSIC vp<[[FPEXT:%.+]]> = call llvm.vp.fpext(ir<[[LD1]]>, ir<true>, vp<[[EVL]]>)
-; IF-EVL-NEXT: WIDEN-INTRINSIC vp<[[LLRINT:%.+]]> = call llvm.vp.llrint(vp<[[FPEXT]]>, ir<true>, vp<[[EVL]]>)
-; IF-EVL-NEXT: WIDEN-INTRINSIC vp<[[TRUNC:%.+]]> = call llvm.vp.trunc(vp<[[LLRINT]]>, ir<true>, vp<[[EVL]]>)
+; IF-EVL-NEXT: WIDEN-INTRINSIC ir<[[FPEXT:%.+]]> = call llvm.vp.fpext(ir<[[LD1]]>, ir<true>, vp<[[EVL]]>)
+; IF-EVL-NEXT: WIDEN-INTRINSIC ir<[[LLRINT:%.+]]> = call llvm.vp.llrint(ir<[[FPEXT]]>, ir<true>, vp<[[EVL]]>)
+; IF-EVL-NEXT: WIDEN-INTRINSIC ir<[[TRUNC:%.+]]> = call llvm.vp.trunc(ir<[[LLRINT]]>, ir<true>, vp<[[EVL]]>)
; IF-EVL-NEXT: CLONE ir<[[GEP2:%.+]]> = getelementptr inbounds ir<%a>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR2:%[0-9]+]]> = vector-pointer ir<[[GEP2]]>
-; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR2]]>, vp<[[TRUNC]]>, vp<[[EVL]]>
+; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR2]]>, ir<[[TRUNC]]>, vp<[[EVL]]>
; IF-EVL-NEXT: SCALAR-CAST vp<[[CAST:%[0-9]+]]> = zext vp<[[EVL]]> to i64
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT]]> = add vp<[[CAST]]>, vp<[[EVL_PHI]]>
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT_EXIT:%.+]]> = add vp<[[IV]]>, vp<[[VFUF]]>
@@ -436,10 +436,10 @@ define void @vp_abs(ptr %a, ptr %b, i64 %N) {
; IF-EVL-NEXT: CLONE ir<[[GEP1:%.+]]> = getelementptr inbounds ir<%b>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR1:%[0-9]+]]> = vector-pointer ir<[[GEP1]]>
; IF-EVL-NEXT: WIDEN ir<[[LD1:%.+]]> = vp.load vp<[[PTR1]]>, vp<[[EVL]]>
-; IF-EVL-NEXT: WIDEN-INTRINSIC vp<[[ABS:%.+]]> = call llvm.vp.abs(ir<[[LD1]]>, ir<true>, ir<true>, vp<[[EVL]]>)
+; IF-EVL-NEXT: WIDEN-INTRINSIC ir<[[ABS:%.+]]> = call llvm.vp.abs(ir<[[LD1]]>, ir<true>, ir<true>, vp<[[EVL]]>)
; IF-EVL-NEXT: CLONE ir<[[GEP2:%.+]]> = getelementptr inbounds ir<%a>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR2:%[0-9]+]]> = vector-pointer ir<[[GEP2]]>
-; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR2]]>, vp<[[ABS]]>, vp<[[EVL]]>
+; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR2]]>, ir<[[ABS]]>, vp<[[EVL]]>
; IF-EVL-NEXT: SCALAR-CAST vp<[[CAST:%[0-9]+]]> = zext vp<[[EVL]]> to i64
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT]]> = add vp<[[CAST]]>, vp<[[EVL_PHI]]>
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT_EXIT:%.+]]> = add vp<[[IV]]>, vp<[[VFUF]]>
diff --git a/llvm/test/Transforms/LoopVectorize/RISCV/vplan-vp-cast-intrinsics.ll b/llvm/test/Transforms/LoopVectorize/RISCV/vplan-vp-cast-intrinsics.ll
index f4acfd6185644d..cd2639aafe52f9 100644
--- a/llvm/test/Transforms/LoopVectorize/RISCV/vplan-vp-cast-intrinsics.ll
+++ b/llvm/test/Transforms/LoopVectorize/RISCV/vplan-vp-cast-intrinsics.ll
@@ -23,10 +23,10 @@ define void @vp_sext(ptr %a, ptr %b, i64 %N) {
; IF-EVL-NEXT: CLONE ir<[[GEP1:%.+]]> = getelementptr inbounds ir<%b>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR1:%[0-9]+]]> = vector-pointer ir<[[GEP1]]>
; IF-EVL-NEXT: WIDEN ir<[[LD1:%.+]]> = vp.load vp<[[PTR1]]>, vp<[[EVL]]>
-; IF-EVL-NEXT: WIDEN-INTRINSIC vp<[[SEXT:%.+]]> = call llvm.vp.sext(ir<[[LD1]]>, ir<true>, vp<[[EVL]]>)
+; IF-EVL-NEXT: WIDEN-INTRINSIC ir<[[SEXT:%.+]]> = call llvm.vp.sext(ir<[[LD1]]>, ir<true>, vp<[[EVL]]>)
; IF-EVL-NEXT: CLONE ir<[[GEP2:%.+]]> = getelementptr inbounds ir<%a>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR2:%[0-9]+]]> = vector-pointer ir<[[GEP2]]>
-; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR2]]>, vp<[[SEXT]]>, vp<[[EVL]]>
+; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR2]]>, ir<[[SEXT]]>, vp<[[EVL]]>
; IF-EVL-NEXT: SCALAR-CAST vp<[[CAST:%[0-9]+]]> = zext vp<[[EVL]]> to i64
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT]]> = add vp<[[CAST]]>, vp<[[EVL_PHI]]>
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT_EXIT:%.+]]> = add vp<[[IV]]>, vp<[[VFUF]]>
@@ -73,10 +73,10 @@ define void @vp_zext(ptr %a, ptr %b, i64 %N) {
; IF-EVL-NEXT: CLONE ir<[[GEP1:%.+]]> = getelementptr inbounds ir<%b>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR1:%[0-9]+]]> = vector-pointer ir<[[GEP1]]>
; IF-EVL-NEXT: WIDEN ir<[[LD1:%.+]]> = vp.load vp<[[PTR1]]>, vp<[[EVL]]>
-; IF-EVL-NEXT: WIDEN-INTRINSIC vp<[[ZEXT:%.+]]> = call llvm.vp.zext(ir<[[LD1]]>, ir<true>, vp<[[EVL]]>)
+; IF-EVL-NEXT: WIDEN-INTRINSIC ir<[[ZEXT:%.+]]> = call llvm.vp.zext(ir<[[LD1]]>, ir<true>, vp<[[EVL]]>)
; IF-EVL-NEXT: CLONE ir<[[GEP2:%.+]]> = getelementptr inbounds ir<%a>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR2:%[0-9]+]]> = vector-pointer ir<[[GEP2]]>
-; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR2]]>, vp<[[ZEXT]]>, vp<[[EVL]]>
+; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR2]]>, ir<[[ZEXT]]>, vp<[[EVL]]>
; IF-EVL-NEXT: SCALAR-CAST vp<[[CAST:%[0-9]+]]> = zext vp<[[EVL]]> to i64
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT]]> = add vp<[[CAST]]>, vp<[[EVL_PHI]]>
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT_EXIT:%.+]]> = add vp<[[IV]]>, vp<[[VFUF]]>
@@ -121,10 +121,10 @@ define void @vp_trunc(ptr %a, ptr %b, i64 %N) {
; IF-EVL-NEXT: CLONE ir<[[GEP1:%.+]]> = getelementptr inbounds ir<%b>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR1:%[0-9]+]]> = vector-pointer ir<[[GEP1]]>
; IF-EVL-NEXT: WIDEN ir<[[LD1:%.+]]> = vp.load vp<[[PTR1]]>, vp<[[EVL]]>
-; IF-EVL-NEXT: WIDEN-INTRINSIC vp<[[TRUNC:%.+]]> = call llvm.vp.trunc(ir<[[LD1]]>, ir<true>, vp<[[EVL]]>)
+; IF-EVL-NEXT: WIDEN-INTRINSIC ir<[[TRUNC:%.+]]> = call llvm.vp.trunc(ir<[[LD1]]>, ir<true>, vp<[[EVL]]>)
; IF-EVL-NEXT: CLONE ir<[[GEP2:%.+]]> = getelementptr inbounds ir<%a>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR2:%[0-9]+]]> = vector-pointer ir<[[GEP2]]>
-; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR2]]>, vp<[[TRUNC]]>, vp<[[EVL]]>
+; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR2]]>, ir<[[TRUNC]]>, vp<[[EVL]]>
; IF-EVL-NEXT: SCALAR-CAST vp<[[CAST:%[0-9]+]]> = zext vp<[[EVL]]> to i64
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT]]> = add vp<[[CAST]]>, vp<[[EVL_PHI]]>
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT_EXIT:%.+]]> = add vp<[[IV]]>, vp<[[VFUF]]>
@@ -169,10 +169,10 @@ define void @vp_fpext(ptr %a, ptr %b, i64 %N) {
; IF-EVL-NEXT: CLONE ir<[[GEP1:%.+]]> = getelementptr inbounds ir<%b>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR1:%[0-9]+]]> = vector-pointer ir<[[GEP1]]>
; IF-EVL-NEXT: WIDEN ir<[[LD1:%.+]]> = vp.load vp<[[PTR1]]>, vp<[[EVL]]>
-; IF-EVL-NEXT: WIDEN-INTRINSIC vp<[[FPEXT:%.+]]> = call llvm.vp.fpext(ir<[[LD1]]>, ir<true>, vp<[[EVL]]>)
+; IF-EVL-NEXT: WIDEN-INTRINSIC ir<[[FPEXT:%.+]]> = call llvm.vp.fpext(ir<[[LD1]]>, ir<true>, vp<[[EVL]]>)
; IF-EVL-NEXT: CLONE ir<[[GEP2:%.+]]> = getelementptr inbounds ir<%a>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR2:%[0-9]+]]> = vector-pointer ir<[[GEP2]]>
-; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR2]]>, vp<[[FPEXT]]>, vp<[[EVL]]>
+; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR2]]>, ir<[[FPEXT]]>, vp<[[EVL]]>
; IF-EVL-NEXT: SCALAR-CAST vp<[[CAST:%[0-9]+]]> = zext vp<[[EVL]]> to i64
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT]]> = add vp<[[CAST]]>, vp<[[EVL_PHI]]>
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT_EXIT:%.+]]> = add vp<[[IV]]>, vp<[[VFUF]]>
@@ -217,10 +217,10 @@ define void @vp_fptrunc(ptr %a, ptr %b, i64 %N) {
; IF-EVL-NEXT: CLONE ir<[[GEP1:%.+]]> = getelementptr inbounds ir<%b>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR1:%[0-9]+]]> = vector-pointer ir<[[GEP1]]>
; IF-EVL-NEXT: WIDEN ir<[[LD1:%.+]]> = vp.load vp<[[PTR1]]>, vp<[[EVL]]>
-; IF-EVL-NEXT: WIDEN-INTRINSIC vp<[[FPTRUNC:%.+]]> = call llvm.vp.fptrunc(ir<[[LD1]]>, ir<true>, vp<[[EVL]]>)
+; IF-EVL-NEXT: WIDEN-INTRINSIC ir<[[FPTRUNC:%.+]]> = call llvm.vp.fptrunc(ir<[[LD1]]>, ir<true>, vp<[[EVL]]>)
; IF-EVL-NEXT: CLONE ir<[[GEP2:%.+]]> = getelementptr inbounds ir<%a>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR2:%[0-9]+]]> = vector-pointer ir<[[GEP2]]>
-; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR2]]>, vp<[[FPTRUNC]]>, vp<[[EVL]]>
+; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR2]]>, ir<[[FPTRUNC]]>, vp<[[EVL]]>
; IF-EVL-NEXT: SCALAR-CAST vp<[[CAST:%[0-9]+]]> = zext vp<[[EVL]]> to i64
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT]]> = add vp<[[CAST]]>, vp<[[EVL_PHI]]>
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT_EXIT:%.+]]> = add vp<[[IV]]>, vp<[[VFUF]]>
@@ -265,10 +265,10 @@ define void @vp_sitofp(ptr %a, ptr %b, i64 %N) {
; IF-EVL-NEXT: CLONE ir<[[GEP1:%.+]]> = getelementptr inbounds ir<%b>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR1:%[0-9]+]]> = vector-pointer ir<[[GEP1]]>
; IF-EVL-NEXT: WIDEN ir<[[LD1:%.+]]> = vp.load vp<[[PTR1]]>, vp<[[EVL]]>
-; IF-EVL-NEXT: WIDEN-INTRINSIC vp<[[SITOFP:%.+]]> = call llvm.vp.sitofp(ir<[[LD1]]>, ir<true>, vp<[[EVL]]>)
+; IF-EVL-NEXT: WIDEN-INTRINSIC ir<[[SITOFP:%.+]]> = call llvm.vp.sitofp(ir<[[LD1]]>, ir<true>, vp<[[EVL]]>)
; IF-EVL-NEXT: CLONE ir<[[GEP2:%.+]]> = getelementptr inbounds ir<%a>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR2:%[0-9]+]]> = vector-pointer ir<[[GEP2]]>
-; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR2]]>, vp<[[SITOFP]]>, vp<[[EVL]]>
+; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR2]]>, ir<[[SITOFP]]>, vp<[[EVL]]>
; IF-EVL-NEXT: SCALAR-CAST vp<[[CAST:%[0-9]+]]> = zext vp<[[EVL]]> to i64
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT]]> = add vp<[[CAST]]>, vp<[[EVL_PHI]]>
; IF-EVL-NEXT: EMIT vp<[[IV_NEXT_EXIT:%.+]]> = add vp<[[IV]]>, vp<[[VFUF]]>
@@ -313,10 +313,10 @@ define void @vp_uitofp(ptr %a, ptr %b, i64 %N) {
; IF-EVL-NEXT: CLONE ir<[[GEP1:%.+]]> = getelementptr inbounds ir<%b>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR1:%[0-9]+]]> = vector-pointer ir<[[GEP1]]>
; IF-EVL-NEXT: WIDEN ir<[[LD1:%.+]]> = vp.load vp<[[PTR1]]>, vp<[[EVL]]>
-; IF-EVL-NEXT: WIDEN-INTRINSIC vp<[[UITOFP:%.+]]> = call llvm.vp.uitofp(ir<[[LD1]]>, ir<true>, vp<[[EVL]]>)
+; IF-EVL-NEXT: WIDEN-INTRINSIC ir<[[UITOFP:%.+]]> = call llvm.vp.uitofp(ir<[[LD1]]>, ir<true>, vp<[[EVL]]>)
; IF-EVL-NEXT: CLONE ir<[[GEP2:%.+]]> = getelementptr inbounds ir<%a>, vp<[[ST]]>
; IF-EVL-NEXT: vp<[[PTR2:%[0-9]+]]> = vector-pointer ir<[[GEP2]]>
-; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR2]]>, vp<[[UITOFP]]>, vp<[[EVL]]>
+; IF-EVL-NEXT: WIDEN vp.store vp<[[PTR2]]>, ir<[[UITOFP]]>, vp<[[EVL]]>
; IF-EVL-NEXT: SCALAR-CAST vp<[[CAST:%[0-9]+]]> = zext vp<...
[truncated]
|
|
Does #113903 fix the issue? cc @ElvisWang123 |
Currently, not. |
LiqinWeng
reviewed
Jan 6, 2025
| VPID, Ops, TypeInfo.inferScalarType(CR), CR->getDebugLoc()); | ||
| // Preserve the underlying value as some cast instructions may have | ||
| // their cost skipped in collectValuesToIgnore | ||
| VPR->setUnderlyingValue(CR->getUnderlyingValue()); |
There was a problem hiding this comment.
Add a condition CR->getUnderlingValue != nullptr
There was a problem hiding this comment.
As in an assertion? I think it's possible for a VPWidenIntrinsicRecipe to have no underlying value if constructed with the VPWidenIntrinsicRecipe(Intrinsic::ID, ArrayRef<VPValue *>, Type *, DebugLoc) constructor
There was a problem hiding this comment.
I think he suggested something like this:
if (CR->getUnderlingValue())
VPR->setUnderlyingValue(CR->getUnderlyingValue());
There was a problem hiding this comment.
I think that just does the same thing if CR->getUnderlyingValue() is null. Happy to add it if people prefer it stylistically though
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This fixes a crash that shows up with EVL tail folding due to a mismatch between the legacy and VPlan cost models.
In the legacy model a cast instruction may have its cost ignored if it's part of a reduction/induction variable.
The VPlan cost model will check if the underying value of a recipe is skipped in Ctx.skipCostComputation, so this works for normally widened casts.
However when transformed to a VP intrinsic we drop the underlying value and we now end up costing the previously skipped values, which leads to the mismatch.
This fixes the crash by preserving the underlying cast instruction. It means that a VPWidenIntrinsicRecipe's underlying value might not always be a CallInst anymore, so there was one place where I had to relax a cast<> to a dyn_cast<>.
I wasn't able to provide a test case for the crash itself since we can't yet do EVL tail folding with widened IVs, but it will become a problem later down the line.