[llvm] [SDAG] [X86] Extend SplitVecOp_VSETCC for STRICT_FSETCCS (PR #116768)
via llvm-commits
llvm-commits at lists.llvm.org
Wed Nov 20 22:01:55 PST 2024
https://github.com/abhishek-kaushik22 updated https://github.com/llvm/llvm-project/pull/116768
>From 90b27e75897ee756e35dd23bd157a454cb42abdc Mon Sep 17 00:00:00 2001
From: abhishek-kaushik22 <abhishek.kaushik at intel.com>
Date: Tue, 19 Nov 2024 13:58:41 +0530
Subject: [PATCH 1/6] [SDAG] [X86] Extend SplitVecOp_VSETCC for STRICT_FSETCCS
---
.../SelectionDAG/LegalizeVectorTypes.cpp | 11 ++++---
.../X86/vec-strict-cmp-512-skx-fp64.ll | 31 +++++++++++++++++++
2 files changed, 38 insertions(+), 4 deletions(-)
create mode 100644 llvm/test/CodeGen/X86/vec-strict-cmp-512-skx-fp64.ll
diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
index eccda73548e874..3dda807ed65259 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
@@ -3235,6 +3235,7 @@ bool DAGTypeLegalizer::SplitVectorOperand(SDNode *N, unsigned OpNo) {
case ISD::VP_SETCC:
case ISD::STRICT_FSETCC:
+ case ISD::STRICT_FSETCCS:
case ISD::SETCC: Res = SplitVecOp_VSETCC(N); break;
case ISD::BITCAST: Res = SplitVecOp_BITCAST(N); break;
case ISD::EXTRACT_SUBVECTOR: Res = SplitVecOp_EXTRACT_SUBVECTOR(N); break;
@@ -4236,7 +4237,8 @@ SDValue DAGTypeLegalizer::SplitVecOp_TruncateHelper(SDNode *N) {
}
SDValue DAGTypeLegalizer::SplitVecOp_VSETCC(SDNode *N) {
- bool isStrict = N->getOpcode() == ISD::STRICT_FSETCC;
+ bool isStrict = N->getOpcode() == ISD::STRICT_FSETCC ||
+ N->getOpcode() == ISD::STRICT_FSETCCS;
assert(N->getValueType(0).isVector() &&
N->getOperand(isStrict ? 1 : 0).getValueType().isVector() &&
"Operand types must be vectors");
@@ -4255,11 +4257,12 @@ SDValue DAGTypeLegalizer::SplitVecOp_VSETCC(SDNode *N) {
if (N->getOpcode() == ISD::SETCC) {
LoRes = DAG.getNode(ISD::SETCC, DL, PartResVT, Lo0, Lo1, N->getOperand(2));
HiRes = DAG.getNode(ISD::SETCC, DL, PartResVT, Hi0, Hi1, N->getOperand(2));
- } else if (N->getOpcode() == ISD::STRICT_FSETCC) {
- LoRes = DAG.getNode(ISD::STRICT_FSETCC, DL,
+ } else if (N->getOpcode() == ISD::STRICT_FSETCC ||
+ N->getOpcode() == ISD::STRICT_FSETCCS) {
+ LoRes = DAG.getNode(N->getOpcode(), DL,
DAG.getVTList(PartResVT, N->getValueType(1)),
N->getOperand(0), Lo0, Lo1, N->getOperand(3));
- HiRes = DAG.getNode(ISD::STRICT_FSETCC, DL,
+ HiRes = DAG.getNode(N->getOpcode(), DL,
DAG.getVTList(PartResVT, N->getValueType(1)),
N->getOperand(0), Hi0, Hi1, N->getOperand(3));
SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
diff --git a/llvm/test/CodeGen/X86/vec-strict-cmp-512-skx-fp64.ll b/llvm/test/CodeGen/X86/vec-strict-cmp-512-skx-fp64.ll
new file mode 100644
index 00000000000000..181d0cd0ef1d7a
--- /dev/null
+++ b/llvm/test/CodeGen/X86/vec-strict-cmp-512-skx-fp64.ll
@@ -0,0 +1,31 @@
+; RUN: llc < %s -mtriple=x86_64 -mcpu=skx | FileCheck %s --check-prefixes=SKX
+
+define <16 x i32> @test_v16f64_ogt(<16 x i32> %a, <16 x i32> %b, <16 x double> %f1, <16 x double> %f2) #0 {
+; SKX-LABEL: test_v16f64_ogt
+; SKX: # %bb.0:
+; SKX-NEXT: pushq %rbp
+; SKX-NEXT: movq %rsp, %rbp
+; SKX-NEXT: andq $-32, %rsp
+; SKX-NEXT: subq $32, %rsp
+; SKX-NEXT: vcmpgtpd 80(%rbp), %ymm6, %k0
+; SKX-NEXT: vcmpgtpd 112(%rbp), %ymm7, %k1
+; SKX-NEXT: kshiftlb $4, %k1, %k1
+; SKX-NEXT: korb %k1, %k0, %k1
+; SKX-NEXT: vcmpgtpd 16(%rbp), %ymm4, %k0
+; SKX-NEXT: vcmpgtpd 48(%rbp), %ymm5, %k2
+; SKX-NEXT: kshiftlb $4, %k2, %k2
+; SKX-NEXT: korb %k2, %k0, %k2
+; SKX-NEXT: vpblendmd %ymm0, %ymm2, %ymm0 {%k2}
+; SKX-NEXT: vpblendmd %ymm1, %ymm3, %ymm1 {%k1}
+; SKX-NEXT: movq %rbp, %rsp
+; SKX-NEXT: popq %rbp
+; SKX-NEXT: retq
+ %cond = tail call <16 x i1> @llvm.experimental.constrained.fcmps.v16f64(
+ <16 x double> %f1, <16 x double> %f2, metadata !"ogt", metadata !"fpexcept.maytrap")
+ %res = select <16 x i1> %cond, <16 x i32> %a, <16 x i32> %b
+ ret <16 x i32> %res
+}
+
+declare <16 x i1> @llvm.experimental.constrained.fcmps.v16f64(<16 x double>, <16 x double>, metadata, metadata)
+
+attributes #0 = { nounwind strictfp "min-legal-vector-width"="0" }
\ No newline at end of file
>From 401449798893724789acb4077c38834b8d3422a2 Mon Sep 17 00:00:00 2001
From: abhishek-kaushik22 <abhishek.kaushik at intel.com>
Date: Tue, 19 Nov 2024 14:00:31 +0530
Subject: [PATCH 2/6] Update vec-strict-cmp-512-skx-fp64.ll
---
llvm/test/CodeGen/X86/vec-strict-cmp-512-skx-fp64.ll | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/llvm/test/CodeGen/X86/vec-strict-cmp-512-skx-fp64.ll b/llvm/test/CodeGen/X86/vec-strict-cmp-512-skx-fp64.ll
index 181d0cd0ef1d7a..c3d4052c553f59 100644
--- a/llvm/test/CodeGen/X86/vec-strict-cmp-512-skx-fp64.ll
+++ b/llvm/test/CodeGen/X86/vec-strict-cmp-512-skx-fp64.ll
@@ -28,4 +28,4 @@ define <16 x i32> @test_v16f64_ogt(<16 x i32> %a, <16 x i32> %b, <16 x double> %
declare <16 x i1> @llvm.experimental.constrained.fcmps.v16f64(<16 x double>, <16 x double>, metadata, metadata)
-attributes #0 = { nounwind strictfp "min-legal-vector-width"="0" }
\ No newline at end of file
+attributes #0 = { nounwind strictfp "min-legal-vector-width"="0" }
>From 79d4592e02738b614b1a72254ee7ce31be80ec4e Mon Sep 17 00:00:00 2001
From: abhishek-kaushik22 <abhishek.kaushik at intel.com>
Date: Tue, 19 Nov 2024 15:48:36 +0530
Subject: [PATCH 3/6] Use a variable for N->getOpcode() calls
---
.../SelectionDAG/LegalizeVectorTypes.cpp | 17 +++++++++--------
1 file changed, 9 insertions(+), 8 deletions(-)
diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
index 3dda807ed65259..8ac0835d719e16 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
@@ -4237,8 +4237,9 @@ SDValue DAGTypeLegalizer::SplitVecOp_TruncateHelper(SDNode *N) {
}
SDValue DAGTypeLegalizer::SplitVecOp_VSETCC(SDNode *N) {
- bool isStrict = N->getOpcode() == ISD::STRICT_FSETCC ||
- N->getOpcode() == ISD::STRICT_FSETCCS;
+ auto Opc = N->getOpcode();
+ bool isStrict = Opc == ISD::STRICT_FSETCC ||
+ Opc == ISD::STRICT_FSETCCS;
assert(N->getValueType(0).isVector() &&
N->getOperand(isStrict ? 1 : 0).getValueType().isVector() &&
"Operand types must be vectors");
@@ -4254,22 +4255,22 @@ SDValue DAGTypeLegalizer::SplitVecOp_VSETCC(SDNode *N) {
EVT PartResVT = EVT::getVectorVT(Context, MVT::i1, PartEltCnt);
EVT WideResVT = EVT::getVectorVT(Context, MVT::i1, PartEltCnt*2);
- if (N->getOpcode() == ISD::SETCC) {
+ if (Opc == ISD::SETCC) {
LoRes = DAG.getNode(ISD::SETCC, DL, PartResVT, Lo0, Lo1, N->getOperand(2));
HiRes = DAG.getNode(ISD::SETCC, DL, PartResVT, Hi0, Hi1, N->getOperand(2));
- } else if (N->getOpcode() == ISD::STRICT_FSETCC ||
- N->getOpcode() == ISD::STRICT_FSETCCS) {
- LoRes = DAG.getNode(N->getOpcode(), DL,
+ } else if (Opc == ISD::STRICT_FSETCC ||
+ Opc == ISD::STRICT_FSETCCS) {
+ LoRes = DAG.getNode(Opc, DL,
DAG.getVTList(PartResVT, N->getValueType(1)),
N->getOperand(0), Lo0, Lo1, N->getOperand(3));
- HiRes = DAG.getNode(N->getOpcode(), DL,
+ HiRes = DAG.getNode(Opc, DL,
DAG.getVTList(PartResVT, N->getValueType(1)),
N->getOperand(0), Hi0, Hi1, N->getOperand(3));
SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
LoRes.getValue(1), HiRes.getValue(1));
ReplaceValueWith(SDValue(N, 1), NewChain);
} else {
- assert(N->getOpcode() == ISD::VP_SETCC && "Expected VP_SETCC opcode");
+ assert(Opc == ISD::VP_SETCC && "Expected VP_SETCC opcode");
SDValue MaskLo, MaskHi, EVLLo, EVLHi;
std::tie(MaskLo, MaskHi) = SplitMask(N->getOperand(3));
std::tie(EVLLo, EVLHi) =
>From 6aa250588b370c49081ad0dc3c2f983e3a9c91a0 Mon Sep 17 00:00:00 2001
From: abhishek-kaushik22 <abhishek.kaushik at intel.com>
Date: Tue, 19 Nov 2024 15:54:30 +0530
Subject: [PATCH 4/6] Update LegalizeVectorTypes.cpp
---
.../lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp | 12 ++++--------
1 file changed, 4 insertions(+), 8 deletions(-)
diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
index 8ac0835d719e16..d891e818ca56db 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
@@ -4238,8 +4238,7 @@ SDValue DAGTypeLegalizer::SplitVecOp_TruncateHelper(SDNode *N) {
SDValue DAGTypeLegalizer::SplitVecOp_VSETCC(SDNode *N) {
auto Opc = N->getOpcode();
- bool isStrict = Opc == ISD::STRICT_FSETCC ||
- Opc == ISD::STRICT_FSETCCS;
+ bool isStrict = Opc == ISD::STRICT_FSETCC || Opc == ISD::STRICT_FSETCCS;
assert(N->getValueType(0).isVector() &&
N->getOperand(isStrict ? 1 : 0).getValueType().isVector() &&
"Operand types must be vectors");
@@ -4258,13 +4257,10 @@ SDValue DAGTypeLegalizer::SplitVecOp_VSETCC(SDNode *N) {
if (Opc == ISD::SETCC) {
LoRes = DAG.getNode(ISD::SETCC, DL, PartResVT, Lo0, Lo1, N->getOperand(2));
HiRes = DAG.getNode(ISD::SETCC, DL, PartResVT, Hi0, Hi1, N->getOperand(2));
- } else if (Opc == ISD::STRICT_FSETCC ||
- Opc == ISD::STRICT_FSETCCS) {
- LoRes = DAG.getNode(Opc, DL,
- DAG.getVTList(PartResVT, N->getValueType(1)),
+ } else if (Opc == ISD::STRICT_FSETCC || Opc == ISD::STRICT_FSETCCS) {
+ LoRes = DAG.getNode(Opc, DL, DAG.getVTList(PartResVT, N->getValueType(1)),
N->getOperand(0), Lo0, Lo1, N->getOperand(3));
- HiRes = DAG.getNode(Opc, DL,
- DAG.getVTList(PartResVT, N->getValueType(1)),
+ HiRes = DAG.getNode(Opc, DL, DAG.getVTList(PartResVT, N->getValueType(1)),
N->getOperand(0), Hi0, Hi1, N->getOperand(3));
SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
LoRes.getValue(1), HiRes.getValue(1));
>From 6f69ac5167c1515e78278317b27e39ffe39506e0 Mon Sep 17 00:00:00 2001
From: abhishek-kaushik22 <abhishek.kaushik at intel.com>
Date: Tue, 19 Nov 2024 16:09:24 +0530
Subject: [PATCH 5/6] Update LegalizeVectorTypes.cpp
---
llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
index d891e818ca56db..3d24b039aeafc0 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
@@ -4257,7 +4257,7 @@ SDValue DAGTypeLegalizer::SplitVecOp_VSETCC(SDNode *N) {
if (Opc == ISD::SETCC) {
LoRes = DAG.getNode(ISD::SETCC, DL, PartResVT, Lo0, Lo1, N->getOperand(2));
HiRes = DAG.getNode(ISD::SETCC, DL, PartResVT, Hi0, Hi1, N->getOperand(2));
- } else if (Opc == ISD::STRICT_FSETCC || Opc == ISD::STRICT_FSETCCS) {
+ } else if (isStrict) {
LoRes = DAG.getNode(Opc, DL, DAG.getVTList(PartResVT, N->getValueType(1)),
N->getOperand(0), Lo0, Lo1, N->getOperand(3));
HiRes = DAG.getNode(Opc, DL, DAG.getVTList(PartResVT, N->getValueType(1)),
>From be062db6513df450229b105039c7ce05fb96d523 Mon Sep 17 00:00:00 2001
From: abhishek-kaushik22 <abhishek.kaushik at intel.com>
Date: Thu, 21 Nov 2024 11:31:38 +0530
Subject: [PATCH 6/6] Add test to existing file, replace auto with unsigned
---
.../SelectionDAG/LegalizeVectorTypes.cpp | 2 +-
.../X86/vec-strict-cmp-512-skx-fp64.ll | 31 -------------------
.../CodeGen/X86/vec-strict-cmp-512-skx.ll | 27 ++++++++++++++++
3 files changed, 28 insertions(+), 32 deletions(-)
delete mode 100644 llvm/test/CodeGen/X86/vec-strict-cmp-512-skx-fp64.ll
diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
index 3d24b039aeafc0..465128099f4447 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
@@ -4237,7 +4237,7 @@ SDValue DAGTypeLegalizer::SplitVecOp_TruncateHelper(SDNode *N) {
}
SDValue DAGTypeLegalizer::SplitVecOp_VSETCC(SDNode *N) {
- auto Opc = N->getOpcode();
+ unsigned Opc = N->getOpcode();
bool isStrict = Opc == ISD::STRICT_FSETCC || Opc == ISD::STRICT_FSETCCS;
assert(N->getValueType(0).isVector() &&
N->getOperand(isStrict ? 1 : 0).getValueType().isVector() &&
diff --git a/llvm/test/CodeGen/X86/vec-strict-cmp-512-skx-fp64.ll b/llvm/test/CodeGen/X86/vec-strict-cmp-512-skx-fp64.ll
deleted file mode 100644
index c3d4052c553f59..00000000000000
--- a/llvm/test/CodeGen/X86/vec-strict-cmp-512-skx-fp64.ll
+++ /dev/null
@@ -1,31 +0,0 @@
-; RUN: llc < %s -mtriple=x86_64 -mcpu=skx | FileCheck %s --check-prefixes=SKX
-
-define <16 x i32> @test_v16f64_ogt(<16 x i32> %a, <16 x i32> %b, <16 x double> %f1, <16 x double> %f2) #0 {
-; SKX-LABEL: test_v16f64_ogt
-; SKX: # %bb.0:
-; SKX-NEXT: pushq %rbp
-; SKX-NEXT: movq %rsp, %rbp
-; SKX-NEXT: andq $-32, %rsp
-; SKX-NEXT: subq $32, %rsp
-; SKX-NEXT: vcmpgtpd 80(%rbp), %ymm6, %k0
-; SKX-NEXT: vcmpgtpd 112(%rbp), %ymm7, %k1
-; SKX-NEXT: kshiftlb $4, %k1, %k1
-; SKX-NEXT: korb %k1, %k0, %k1
-; SKX-NEXT: vcmpgtpd 16(%rbp), %ymm4, %k0
-; SKX-NEXT: vcmpgtpd 48(%rbp), %ymm5, %k2
-; SKX-NEXT: kshiftlb $4, %k2, %k2
-; SKX-NEXT: korb %k2, %k0, %k2
-; SKX-NEXT: vpblendmd %ymm0, %ymm2, %ymm0 {%k2}
-; SKX-NEXT: vpblendmd %ymm1, %ymm3, %ymm1 {%k1}
-; SKX-NEXT: movq %rbp, %rsp
-; SKX-NEXT: popq %rbp
-; SKX-NEXT: retq
- %cond = tail call <16 x i1> @llvm.experimental.constrained.fcmps.v16f64(
- <16 x double> %f1, <16 x double> %f2, metadata !"ogt", metadata !"fpexcept.maytrap")
- %res = select <16 x i1> %cond, <16 x i32> %a, <16 x i32> %b
- ret <16 x i32> %res
-}
-
-declare <16 x i1> @llvm.experimental.constrained.fcmps.v16f64(<16 x double>, <16 x double>, metadata, metadata)
-
-attributes #0 = { nounwind strictfp "min-legal-vector-width"="0" }
diff --git a/llvm/test/CodeGen/X86/vec-strict-cmp-512-skx.ll b/llvm/test/CodeGen/X86/vec-strict-cmp-512-skx.ll
index 3028b749673784..0ec33ad800912f 100644
--- a/llvm/test/CodeGen/X86/vec-strict-cmp-512-skx.ll
+++ b/llvm/test/CodeGen/X86/vec-strict-cmp-512-skx.ll
@@ -34,7 +34,34 @@ define <8 x i32> @test_v8f64_oeq_q(<8 x i32> %a, <8 x i32> %b, <8 x double> %f1,
ret <8 x i32> %res
}
+define <16 x i32> @test_v16f64_ogt(<16 x i32> %a, <16 x i32> %b, <16 x double> %f1, <16 x double> %f2) #0 {
+; SKX-LABEL: test_v16f64_ogt
+; SKX: # %bb.0:
+; SKX-NEXT: pushq %rbp
+; SKX-NEXT: movq %rsp, %rbp
+; SKX-NEXT: andq $-32, %rsp
+; SKX-NEXT: subq $32, %rsp
+; SKX-NEXT: vcmpgtpd 80(%rbp), %ymm6, %k0
+; SKX-NEXT: vcmpgtpd 112(%rbp), %ymm7, %k1
+; SKX-NEXT: kshiftlb $4, %k1, %k1
+; SKX-NEXT: korb %k1, %k0, %k1
+; SKX-NEXT: vcmpgtpd 16(%rbp), %ymm4, %k0
+; SKX-NEXT: vcmpgtpd 48(%rbp), %ymm5, %k2
+; SKX-NEXT: kshiftlb $4, %k2, %k2
+; SKX-NEXT: korb %k2, %k0, %k2
+; SKX-NEXT: vpblendmd %ymm0, %ymm2, %ymm0 {%k2}
+; SKX-NEXT: vpblendmd %ymm1, %ymm3, %ymm1 {%k1}
+; SKX-NEXT: movq %rbp, %rsp
+; SKX-NEXT: popq %rbp
+; SKX-NEXT: retq
+ %cond = tail call <16 x i1> @llvm.experimental.constrained.fcmps.v16f64(
+ <16 x double> %f1, <16 x double> %f2, metadata !"ogt", metadata !"fpexcept.maytrap")
+ %res = select <16 x i1> %cond, <16 x i32> %a, <16 x i32> %b
+ ret <16 x i32> %res
+}
+
declare <16 x i1> @llvm.experimental.constrained.fcmp.v16f32(<16 x float>, <16 x float>, metadata, metadata)
declare <8 x i1> @llvm.experimental.constrained.fcmp.v8f64(<8 x double>, <8 x double>, metadata, metadata)
+declare <16 x i1> @llvm.experimental.constrained.fcmps.v16f64(<16 x double>, <16 x double>, metadata, metadata)
attributes #0 = { nounwind strictfp "min-legal-vector-width"="0" }
More information about the llvm-commits
mailing list