[clang-tools-extra] [llvm] [clang] [PowerPC] Implement llvm.set.rounding intrinsic (PR #67302)
Qiu Chaofan via cfe-commits
cfe-commits at lists.llvm.org
Wed Jan 10 01:28:23 PST 2024
https://github.com/ecnelises updated https://github.com/llvm/llvm-project/pull/67302
>From a1567f579531c3abbd1f4e9b7c7edd2f95ead42c Mon Sep 17 00:00:00 2001
From: Qiu Chaofan <qiucofan at cn.ibm.com>
Date: Mon, 25 Sep 2023 17:10:51 +0800
Subject: [PATCH 1/6] [PowerPC] Implement llvm.set.rounding intrinsic
According to LangRef, llvm.set.rounding sets rounding mode by integer argument:
0 - toward zero
1 - to nearest, ties to even
2 - toward positive infinity
3 - toward negative infinity
4 - to nearest, ties away from zero
While PowerPC ISA says:
0 - to nearest
1 - toward zero
2 - toward positive infinity
3 - toward negative infinity
This patch maps the argument and write into last two bits of FPSCR (rounding mode).
---
llvm/lib/Target/PowerPC/PPCISelLowering.cpp | 80 ++++++++
llvm/lib/Target/PowerPC/PPCISelLowering.h | 1 +
llvm/test/CodeGen/PowerPC/frounds.ll | 194 +++++++++++++++++++-
3 files changed, 274 insertions(+), 1 deletion(-)
diff --git a/llvm/lib/Target/PowerPC/PPCISelLowering.cpp b/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
index f4e3531980d165..4e5ff0cb716966 100644
--- a/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
+++ b/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
@@ -426,6 +426,7 @@ PPCTargetLowering::PPCTargetLowering(const PPCTargetMachine &TM,
setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
setOperationAction(ISD::GET_ROUNDING, MVT::i32, Custom);
+ setOperationAction(ISD::SET_ROUNDING, MVT::Other, Custom);
// If we're enabling GP optimizations, use hardware square root
if (!Subtarget.hasFSQRT() &&
@@ -8898,6 +8899,83 @@ SDValue PPCTargetLowering::LowerINT_TO_FP(SDValue Op,
return FP;
}
+SDValue PPCTargetLowering::LowerSET_ROUNDING(SDValue Op,
+ SelectionDAG &DAG) const {
+ SDLoc Dl(Op);
+ MachineFunction &MF = DAG.getMachineFunction();
+ EVT PtrVT = getPointerTy(MF.getDataLayout());
+ SDValue Chain = Op.getOperand(0);
+
+ // If requested mode is constant, just use simpler mtfsb.
+ if (auto *CVal = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
+ uint64_t Mode = CVal->getZExtValue();
+ if (Mode >= 4)
+ llvm_unreachable("Unsupported rounding mode!");
+ unsigned InternalRnd = Mode ^ (~(Mode >> 1) & 1);
+ SDNode *SetHi = DAG.getMachineNode(
+ (InternalRnd & 2) ? PPC::MTFSB1 : PPC::MTFSB0, Dl, MVT::Other,
+ {DAG.getConstant(30, Dl, MVT::i32, true), Chain});
+ SDNode *SetLo = DAG.getMachineNode(
+ (InternalRnd & 1) ? PPC::MTFSB1 : PPC::MTFSB0, Dl, MVT::Other,
+ {DAG.getConstant(31, Dl, MVT::i32, true), SDValue(SetHi, 0)});
+ return SDValue(SetLo, 0);
+ }
+
+ // Use x ^ (~(x >> 1) & 1) to transform LLVM rounding mode to Power format.
+ SDValue One = DAG.getConstant(1, Dl, MVT::i32);
+ SDValue SrcFlag = DAG.getNode(ISD::AND, Dl, MVT::i32, Op.getOperand(1),
+ DAG.getConstant(3, Dl, MVT::i32));
+ SDValue DstFlag = DAG.getNode(
+ ISD::XOR, Dl, MVT::i32, SrcFlag,
+ DAG.getNode(ISD::AND, Dl, MVT::i32,
+ DAG.getNOT(Dl,
+ DAG.getNode(ISD::SRL, Dl, MVT::i32, SrcFlag, One),
+ MVT::i32),
+ One));
+ SDValue MFFS = DAG.getNode(PPCISD::MFFS, Dl, {MVT::f64, MVT::Other}, Chain);
+ Chain = MFFS.getValue(1);
+ SDValue NewFPSCR;
+ if (isTypeLegal(MVT::i64)) {
+ // Set the last two bits (rounding mode) of bitcasted FPSCR.
+ NewFPSCR = DAG.getNode(
+ ISD::OR, Dl, MVT::i64,
+ DAG.getNode(ISD::AND, Dl, MVT::i64,
+ DAG.getNode(ISD::BITCAST, Dl, MVT::i64, MFFS),
+ DAG.getNOT(Dl, DAG.getConstant(3, Dl, MVT::i64), MVT::i64)),
+ DAG.getNode(ISD::ZERO_EXTEND, Dl, MVT::i64, DstFlag));
+ NewFPSCR = DAG.getNode(ISD::BITCAST, Dl, MVT::f64, NewFPSCR);
+ } else {
+ // In 32-bit mode, store f64, load and update the lower half.
+ int SSFI = MF.getFrameInfo().CreateStackObject(8, Align(8), false);
+ SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
+ Chain = DAG.getStore(Chain, Dl, MFFS, StackSlot, MachinePointerInfo());
+ SDValue Addr;
+ if (Subtarget.isLittleEndian())
+ Addr = StackSlot;
+ else
+ Addr = DAG.getNode(ISD::ADD, Dl, PtrVT, StackSlot,
+ DAG.getConstant(4, Dl, PtrVT));
+ SDValue Tmp = DAG.getLoad(MVT::i32, Dl, Chain, Addr, MachinePointerInfo());
+ Chain = Tmp.getValue(1);
+
+ Tmp = DAG.getNode(
+ ISD::OR, Dl, MVT::i32,
+ DAG.getNode(ISD::AND, Dl, MVT::i32, Tmp,
+ DAG.getNOT(Dl, DAG.getConstant(3, Dl, MVT::i32), MVT::i32)),
+ DstFlag);
+
+ Chain = DAG.getStore(Chain, Dl, Tmp, Addr, MachinePointerInfo());
+ NewFPSCR =
+ DAG.getLoad(MVT::f64, Dl, Chain, StackSlot, MachinePointerInfo());
+ Chain = NewFPSCR.getValue(1);
+ }
+ SDValue Zero = DAG.getConstant(0, Dl, MVT::i32, true);
+ SDNode *MTFSF = DAG.getMachineNode(
+ PPC::MTFSF, Dl, MVT::Other,
+ {DAG.getConstant(255, Dl, MVT::i32, true), NewFPSCR, Zero, Zero, Chain});
+ return SDValue(MTFSF, 0);
+}
+
SDValue PPCTargetLowering::LowerGET_ROUNDING(SDValue Op,
SelectionDAG &DAG) const {
SDLoc dl(Op);
@@ -11647,6 +11725,8 @@ SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
case ISD::UINT_TO_FP:
case ISD::SINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
case ISD::GET_ROUNDING: return LowerGET_ROUNDING(Op, DAG);
+ case ISD::SET_ROUNDING:
+ return LowerSET_ROUNDING(Op, DAG);
// Lower 64-bit shifts.
case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
diff --git a/llvm/lib/Target/PowerPC/PPCISelLowering.h b/llvm/lib/Target/PowerPC/PPCISelLowering.h
index 7c62e370f1536a..6c197327593f0f 100644
--- a/llvm/lib/Target/PowerPC/PPCISelLowering.h
+++ b/llvm/lib/Target/PowerPC/PPCISelLowering.h
@@ -1286,6 +1286,7 @@ namespace llvm {
const SDLoc &dl) const;
SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
SDValue LowerGET_ROUNDING(SDValue Op, SelectionDAG &DAG) const;
+ SDValue LowerSET_ROUNDING(SDValue Op, SelectionDAG &DAG) const;
SDValue LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const;
SDValue LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const;
SDValue LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const;
diff --git a/llvm/test/CodeGen/PowerPC/frounds.ll b/llvm/test/CodeGen/PowerPC/frounds.ll
index c1f7181b30f3f6..1944af1687d0ef 100644
--- a/llvm/test/CodeGen/PowerPC/frounds.ll
+++ b/llvm/test/CodeGen/PowerPC/frounds.ll
@@ -77,4 +77,196 @@ return: ; preds = %entry
ret i32 %retval3
}
-declare i32 @llvm.get.rounding() nounwind
+define void @setrnd_tozero() {
+; PPC32-LABEL: setrnd_tozero:
+; PPC32: # %bb.0: # %entry
+; PPC32-NEXT: mtfsb0 30
+; PPC32-NEXT: mtfsb1 31
+; PPC32-NEXT: blr
+;
+; PPC64-LABEL: setrnd_tozero:
+; PPC64: # %bb.0: # %entry
+; PPC64-NEXT: mtfsb0 30
+; PPC64-NEXT: mtfsb1 31
+; PPC64-NEXT: blr
+;
+; PPC64LE-LABEL: setrnd_tozero:
+; PPC64LE: # %bb.0: # %entry
+; PPC64LE-NEXT: mtfsb0 30
+; PPC64LE-NEXT: mtfsb1 31
+; PPC64LE-NEXT: blr
+;
+; DM-LABEL: setrnd_tozero:
+; DM: # %bb.0: # %entry
+; DM-NEXT: mtfsb0 30
+; DM-NEXT: mtfsb1 31
+; DM-NEXT: blr
+entry:
+ call void @llvm.set.rounding(i32 0)
+ ret void
+}
+
+define void @setrnd_tonearest_tieeven() {
+; PPC32-LABEL: setrnd_tonearest_tieeven:
+; PPC32: # %bb.0: # %entry
+; PPC32-NEXT: mtfsb0 30
+; PPC32-NEXT: mtfsb0 31
+; PPC32-NEXT: blr
+;
+; PPC64-LABEL: setrnd_tonearest_tieeven:
+; PPC64: # %bb.0: # %entry
+; PPC64-NEXT: mtfsb0 30
+; PPC64-NEXT: mtfsb0 31
+; PPC64-NEXT: blr
+;
+; PPC64LE-LABEL: setrnd_tonearest_tieeven:
+; PPC64LE: # %bb.0: # %entry
+; PPC64LE-NEXT: mtfsb0 30
+; PPC64LE-NEXT: mtfsb0 31
+; PPC64LE-NEXT: blr
+;
+; DM-LABEL: setrnd_tonearest_tieeven:
+; DM: # %bb.0: # %entry
+; DM-NEXT: mtfsb0 30
+; DM-NEXT: mtfsb0 31
+; DM-NEXT: blr
+entry:
+ call void @llvm.set.rounding(i32 1)
+ ret void
+}
+
+define void @setrnd_toposinf() {
+; PPC32-LABEL: setrnd_toposinf:
+; PPC32: # %bb.0: # %entry
+; PPC32-NEXT: mtfsb1 30
+; PPC32-NEXT: mtfsb0 31
+; PPC32-NEXT: blr
+;
+; PPC64-LABEL: setrnd_toposinf:
+; PPC64: # %bb.0: # %entry
+; PPC64-NEXT: mtfsb1 30
+; PPC64-NEXT: mtfsb0 31
+; PPC64-NEXT: blr
+;
+; PPC64LE-LABEL: setrnd_toposinf:
+; PPC64LE: # %bb.0: # %entry
+; PPC64LE-NEXT: mtfsb1 30
+; PPC64LE-NEXT: mtfsb0 31
+; PPC64LE-NEXT: blr
+;
+; DM-LABEL: setrnd_toposinf:
+; DM: # %bb.0: # %entry
+; DM-NEXT: mtfsb1 30
+; DM-NEXT: mtfsb0 31
+; DM-NEXT: blr
+entry:
+ call void @llvm.set.rounding(i32 2)
+ ret void
+}
+
+define void @setrnd_toneginf() {
+; PPC32-LABEL: setrnd_toneginf:
+; PPC32: # %bb.0: # %entry
+; PPC32-NEXT: mtfsb1 30
+; PPC32-NEXT: mtfsb1 31
+; PPC32-NEXT: blr
+;
+; PPC64-LABEL: setrnd_toneginf:
+; PPC64: # %bb.0: # %entry
+; PPC64-NEXT: mtfsb1 30
+; PPC64-NEXT: mtfsb1 31
+; PPC64-NEXT: blr
+;
+; PPC64LE-LABEL: setrnd_toneginf:
+; PPC64LE: # %bb.0: # %entry
+; PPC64LE-NEXT: mtfsb1 30
+; PPC64LE-NEXT: mtfsb1 31
+; PPC64LE-NEXT: blr
+;
+; DM-LABEL: setrnd_toneginf:
+; DM: # %bb.0: # %entry
+; DM-NEXT: mtfsb1 30
+; DM-NEXT: mtfsb1 31
+; DM-NEXT: blr
+entry:
+ call void @llvm.set.rounding(i32 3)
+ ret void
+}
+
+define void @setrnd_var(i32 %x) {
+; PPC32-LABEL: setrnd_var:
+; PPC32: # %bb.0: # %entry
+; PPC32-NEXT: stwu 1, -16(1)
+; PPC32-NEXT: .cfi_def_cfa_offset 16
+; PPC32-NEXT: mffs 0
+; PPC32-NEXT: stfd 0, 8(1)
+; PPC32-NEXT: clrlwi 4, 3, 30
+; PPC32-NEXT: lwz 5, 12(1)
+; PPC32-NEXT: rlwinm 3, 3, 31, 31, 31
+; PPC32-NEXT: xor 3, 3, 4
+; PPC32-NEXT: xori 3, 3, 1
+; PPC32-NEXT: rlwinm 4, 5, 0, 0, 29
+; PPC32-NEXT: rlwimi 4, 3, 0, 30, 31
+; PPC32-NEXT: stw 4, 12(1)
+; PPC32-NEXT: lfd 0, 8(1)
+; PPC32-NEXT: mtfsf 255, 0
+; PPC32-NEXT: addi 1, 1, 16
+; PPC32-NEXT: blr
+;
+; PPC64-LABEL: setrnd_var:
+; PPC64: # %bb.0: # %entry
+; PPC64-NEXT: mffs 0
+; PPC64-NEXT: stfd 0, -16(1)
+; PPC64-NEXT: ld 5, -16(1)
+; PPC64-NEXT: clrlwi 4, 3, 30
+; PPC64-NEXT: rlwinm 3, 3, 31, 31, 31
+; PPC64-NEXT: xor 3, 3, 4
+; PPC64-NEXT: xori 3, 3, 1
+; PPC64-NEXT: clrldi 3, 3, 32
+; PPC64-NEXT: rldicr 4, 5, 0, 61
+; PPC64-NEXT: or 3, 4, 3
+; PPC64-NEXT: std 3, -8(1)
+; PPC64-NEXT: lfd 0, -8(1)
+; PPC64-NEXT: mtfsf 255, 0
+; PPC64-NEXT: blr
+;
+; PPC64LE-LABEL: setrnd_var:
+; PPC64LE: # %bb.0: # %entry
+; PPC64LE-NEXT: mffs 0
+; PPC64LE-NEXT: stfd 0, -16(1)
+; PPC64LE-NEXT: clrlwi 4, 3, 30
+; PPC64LE-NEXT: rlwinm 3, 3, 31, 31, 31
+; PPC64LE-NEXT: xor 3, 3, 4
+; PPC64LE-NEXT: ld 4, -16(1)
+; PPC64LE-NEXT: xori 3, 3, 1
+; PPC64LE-NEXT: clrldi 3, 3, 32
+; PPC64LE-NEXT: rldicr 4, 4, 0, 61
+; PPC64LE-NEXT: or 3, 4, 3
+; PPC64LE-NEXT: std 3, -8(1)
+; PPC64LE-NEXT: lfd 0, -8(1)
+; PPC64LE-NEXT: mtfsf 255, 0
+; PPC64LE-NEXT: blr
+;
+; DM-LABEL: setrnd_var:
+; DM: # %bb.0: # %entry
+; DM-NEXT: clrlwi 4, 3, 30
+; DM-NEXT: rlwinm 3, 3, 31, 31, 31
+; DM-NEXT: xor 3, 3, 4
+; DM-NEXT: xori 3, 3, 1
+; DM-NEXT: clrldi 3, 3, 32
+; DM-NEXT: mffs 0
+; DM-NEXT: mffprd 4, 0
+; DM-NEXT: rldicr 4, 4, 0, 61
+; DM-NEXT: or 3, 4, 3
+; DM-NEXT: mtfprd 0, 3
+; DM-NEXT: mtfsf 255, 0
+; DM-NEXT: blr
+entry:
+ call void @llvm.set.rounding(i32 %x)
+ ret void
+}
+
+declare i32 @llvm.get.rounding() #0
+declare void @llvm.set.rounding(i32) #0
+
+attributes #0 = { nounwind }
>From f1c1a5c14147c69a5e2731e8ebe0febad6a12c4a Mon Sep 17 00:00:00 2001
From: Qiu Chaofan <qiucofan at cn.ibm.com>
Date: Thu, 16 Nov 2023 15:25:08 +0800
Subject: [PATCH 2/6] Exclude SPE
---
llvm/lib/Target/PowerPC/PPCISelLowering.cpp | 5 ++---
llvm/test/CodeGen/PowerPC/frounds.ll | 4 ++--
2 files changed, 4 insertions(+), 5 deletions(-)
diff --git a/llvm/lib/Target/PowerPC/PPCISelLowering.cpp b/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
index 5efc9ba487a710..3be34f89c51d04 100644
--- a/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
+++ b/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
@@ -420,14 +420,13 @@ PPCTargetLowering::PPCTargetLowering(const PPCTargetMachine &TM,
} else {
setOperationAction(ISD::FMA , MVT::f64, Legal);
setOperationAction(ISD::FMA , MVT::f32, Legal);
+ setOperationAction(ISD::GET_ROUNDING, MVT::i32, Custom);
+ setOperationAction(ISD::SET_ROUNDING, MVT::Other, Custom);
}
if (Subtarget.hasSPE())
setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
- setOperationAction(ISD::GET_ROUNDING, MVT::i32, Custom);
- setOperationAction(ISD::SET_ROUNDING, MVT::Other, Custom);
-
// If we're enabling GP optimizations, use hardware square root
if (!Subtarget.hasFSQRT() &&
!(TM.Options.UnsafeFPMath && Subtarget.hasFRSQRTE() &&
diff --git a/llvm/test/CodeGen/PowerPC/frounds.ll b/llvm/test/CodeGen/PowerPC/frounds.ll
index 1944af1687d0ef..538de2938aba81 100644
--- a/llvm/test/CodeGen/PowerPC/frounds.ll
+++ b/llvm/test/CodeGen/PowerPC/frounds.ll
@@ -233,14 +233,14 @@ define void @setrnd_var(i32 %x) {
; PPC64LE-LABEL: setrnd_var:
; PPC64LE: # %bb.0: # %entry
; PPC64LE-NEXT: mffs 0
-; PPC64LE-NEXT: stfd 0, -16(1)
; PPC64LE-NEXT: clrlwi 4, 3, 30
; PPC64LE-NEXT: rlwinm 3, 3, 31, 31, 31
+; PPC64LE-NEXT: stfd 0, -16(1)
; PPC64LE-NEXT: xor 3, 3, 4
; PPC64LE-NEXT: ld 4, -16(1)
; PPC64LE-NEXT: xori 3, 3, 1
-; PPC64LE-NEXT: clrldi 3, 3, 32
; PPC64LE-NEXT: rldicr 4, 4, 0, 61
+; PPC64LE-NEXT: clrldi 3, 3, 32
; PPC64LE-NEXT: or 3, 4, 3
; PPC64LE-NEXT: std 3, -8(1)
; PPC64LE-NEXT: lfd 0, -8(1)
>From a2c14908060849a83b0ac000c96fa6a9251e811b Mon Sep 17 00:00:00 2001
From: Qiu Chaofan <qiucofan at cn.ibm.com>
Date: Mon, 27 Nov 2023 17:42:07 +0800
Subject: [PATCH 3/6] Use assert instead of unreachable
---
llvm/lib/Target/PowerPC/PPCISelLowering.cpp | 3 +--
1 file changed, 1 insertion(+), 2 deletions(-)
diff --git a/llvm/lib/Target/PowerPC/PPCISelLowering.cpp b/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
index 8cc794ad375de6..32faa60970605a 100644
--- a/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
+++ b/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
@@ -8910,8 +8910,7 @@ SDValue PPCTargetLowering::LowerSET_ROUNDING(SDValue Op,
// If requested mode is constant, just use simpler mtfsb.
if (auto *CVal = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
uint64_t Mode = CVal->getZExtValue();
- if (Mode >= 4)
- llvm_unreachable("Unsupported rounding mode!");
+ assert(Model < 4 && "Unsupported rounding mode!");
unsigned InternalRnd = Mode ^ (~(Mode >> 1) & 1);
SDNode *SetHi = DAG.getMachineNode(
(InternalRnd & 2) ? PPC::MTFSB1 : PPC::MTFSB0, Dl, MVT::Other,
>From 00a5ae8fac889f14ead068f0259540788bd125fb Mon Sep 17 00:00:00 2001
From: Qiu Chaofan <qiucofan at cn.ibm.com>
Date: Mon, 27 Nov 2023 17:49:49 +0800
Subject: [PATCH 4/6] Fixup
---
llvm/lib/Target/PowerPC/PPCISelLowering.cpp | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/llvm/lib/Target/PowerPC/PPCISelLowering.cpp b/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
index 32faa60970605a..c56d8d65c75bdd 100644
--- a/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
+++ b/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
@@ -8910,7 +8910,7 @@ SDValue PPCTargetLowering::LowerSET_ROUNDING(SDValue Op,
// If requested mode is constant, just use simpler mtfsb.
if (auto *CVal = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
uint64_t Mode = CVal->getZExtValue();
- assert(Model < 4 && "Unsupported rounding mode!");
+ assert(Mode < 4 && "Unsupported rounding mode!");
unsigned InternalRnd = Mode ^ (~(Mode >> 1) & 1);
SDNode *SetHi = DAG.getMachineNode(
(InternalRnd & 2) ? PPC::MTFSB1 : PPC::MTFSB0, Dl, MVT::Other,
>From 228d1845d9bf407712ea251ff350f7d6a952b2f2 Mon Sep 17 00:00:00 2001
From: Qiu Chaofan <qiucofan at cn.ibm.com>
Date: Wed, 10 Jan 2024 17:19:47 +0800
Subject: [PATCH 5/6] Exploit P9 mffscrn
---
llvm/lib/Target/PowerPC/PPCISelLowering.cpp | 76 +++++++++-----
llvm/test/CodeGen/PowerPC/frounds.ll | 108 ++++++++++++++++++--
2 files changed, 149 insertions(+), 35 deletions(-)
diff --git a/llvm/lib/Target/PowerPC/PPCISelLowering.cpp b/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
index 454e3c78cafa9e..16dcb973d6c967 100644
--- a/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
+++ b/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
@@ -8906,11 +8906,17 @@ SDValue PPCTargetLowering::LowerSET_ROUNDING(SDValue Op,
EVT PtrVT = getPointerTy(MF.getDataLayout());
SDValue Chain = Op.getOperand(0);
- // If requested mode is constant, just use simpler mtfsb.
+ // If requested mode is constant, just use simpler mtfsb/mffscrni
if (auto *CVal = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
uint64_t Mode = CVal->getZExtValue();
assert(Mode < 4 && "Unsupported rounding mode!");
unsigned InternalRnd = Mode ^ (~(Mode >> 1) & 1);
+ if (Subtarget.isISA3_0())
+ return SDValue(
+ DAG.getMachineNode(
+ PPC::MFFSCRNI, Dl, {MVT::f64, MVT::Other},
+ {DAG.getConstant(InternalRnd, Dl, MVT::i32, true), Chain}),
+ 1);
SDNode *SetHi = DAG.getMachineNode(
(InternalRnd & 2) ? PPC::MTFSB1 : PPC::MTFSB0, Dl, MVT::Other,
{DAG.getConstant(30, Dl, MVT::i32, true), Chain});
@@ -8931,43 +8937,59 @@ SDValue PPCTargetLowering::LowerSET_ROUNDING(SDValue Op,
DAG.getNode(ISD::SRL, Dl, MVT::i32, SrcFlag, One),
MVT::i32),
One));
- SDValue MFFS = DAG.getNode(PPCISD::MFFS, Dl, {MVT::f64, MVT::Other}, Chain);
- Chain = MFFS.getValue(1);
+ // For Power9, there's faster mffscrn, and we don't need to read FPSCR
+ SDValue MFFS;
+ if (!Subtarget.isISA3_0()) {
+ MFFS = DAG.getNode(PPCISD::MFFS, Dl, {MVT::f64, MVT::Other}, Chain);
+ Chain = MFFS.getValue(1);
+ }
SDValue NewFPSCR;
if (isTypeLegal(MVT::i64)) {
- // Set the last two bits (rounding mode) of bitcasted FPSCR.
- NewFPSCR = DAG.getNode(
- ISD::OR, Dl, MVT::i64,
- DAG.getNode(ISD::AND, Dl, MVT::i64,
- DAG.getNode(ISD::BITCAST, Dl, MVT::i64, MFFS),
- DAG.getNOT(Dl, DAG.getConstant(3, Dl, MVT::i64), MVT::i64)),
- DAG.getNode(ISD::ZERO_EXTEND, Dl, MVT::i64, DstFlag));
+ if (Subtarget.isISA3_0())
+ NewFPSCR = DAG.getAnyExtOrTrunc(DstFlag, Dl, MVT::i64);
+ else
+ // Set the last two bits (rounding mode) of bitcasted FPSCR.
+ NewFPSCR = DAG.getNode(
+ ISD::OR, Dl, MVT::i64,
+ DAG.getNode(
+ ISD::AND, Dl, MVT::i64,
+ DAG.getNode(ISD::BITCAST, Dl, MVT::i64, MFFS),
+ DAG.getNOT(Dl, DAG.getConstant(3, Dl, MVT::i64), MVT::i64)),
+ DAG.getNode(ISD::ZERO_EXTEND, Dl, MVT::i64, DstFlag));
NewFPSCR = DAG.getNode(ISD::BITCAST, Dl, MVT::f64, NewFPSCR);
} else {
// In 32-bit mode, store f64, load and update the lower half.
int SSFI = MF.getFrameInfo().CreateStackObject(8, Align(8), false);
SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
- Chain = DAG.getStore(Chain, Dl, MFFS, StackSlot, MachinePointerInfo());
- SDValue Addr;
- if (Subtarget.isLittleEndian())
- Addr = StackSlot;
- else
- Addr = DAG.getNode(ISD::ADD, Dl, PtrVT, StackSlot,
- DAG.getConstant(4, Dl, PtrVT));
- SDValue Tmp = DAG.getLoad(MVT::i32, Dl, Chain, Addr, MachinePointerInfo());
- Chain = Tmp.getValue(1);
-
- Tmp = DAG.getNode(
- ISD::OR, Dl, MVT::i32,
- DAG.getNode(ISD::AND, Dl, MVT::i32, Tmp,
- DAG.getNOT(Dl, DAG.getConstant(3, Dl, MVT::i32), MVT::i32)),
- DstFlag);
-
- Chain = DAG.getStore(Chain, Dl, Tmp, Addr, MachinePointerInfo());
+ SDValue Addr = Subtarget.isLittleEndian()
+ ? StackSlot
+ : DAG.getNode(ISD::ADD, Dl, PtrVT, StackSlot,
+ DAG.getConstant(4, Dl, PtrVT));
+ if (Subtarget.isISA3_0()) {
+ Chain = DAG.getStore(Chain, Dl, DstFlag, Addr, MachinePointerInfo());
+ } else {
+ Chain = DAG.getStore(Chain, Dl, MFFS, StackSlot, MachinePointerInfo());
+ SDValue Tmp =
+ DAG.getLoad(MVT::i32, Dl, Chain, Addr, MachinePointerInfo());
+ Chain = Tmp.getValue(1);
+
+ Tmp = DAG.getNode(
+ ISD::OR, Dl, MVT::i32,
+ DAG.getNode(
+ ISD::AND, Dl, MVT::i32, Tmp,
+ DAG.getNOT(Dl, DAG.getConstant(3, Dl, MVT::i32), MVT::i32)),
+ DstFlag);
+
+ Chain = DAG.getStore(Chain, Dl, Tmp, Addr, MachinePointerInfo());
+ }
NewFPSCR =
DAG.getLoad(MVT::f64, Dl, Chain, StackSlot, MachinePointerInfo());
Chain = NewFPSCR.getValue(1);
}
+ if (Subtarget.isISA3_0())
+ return SDValue(DAG.getMachineNode(PPC::MFFSCRN, Dl, {MVT::f64, MVT::Other},
+ {NewFPSCR, Chain}),
+ 1);
SDValue Zero = DAG.getConstant(0, Dl, MVT::i32, true);
SDNode *MTFSF = DAG.getMachineNode(
PPC::MTFSF, Dl, MVT::Other,
diff --git a/llvm/test/CodeGen/PowerPC/frounds.ll b/llvm/test/CodeGen/PowerPC/frounds.ll
index 538de2938aba81..de94aab839f8bc 100644
--- a/llvm/test/CodeGen/PowerPC/frounds.ll
+++ b/llvm/test/CodeGen/PowerPC/frounds.ll
@@ -5,14 +5,17 @@
; RUN: -check-prefix=PPC64
; RUN: llc -verify-machineinstrs < %s -mtriple=powerpc64le -mattr=-direct-move \
; RUN: | FileCheck %s -check-prefix=PPC64LE
+; RUN: llc -verify-machineinstrs < %s -mtriple=ppc32-- -mcpu=pwr9 \
+; RUN: | FileCheck %s -check-prefix=P9_32
+; RUN: llc -verify-machineinstrs < %s -mtriple=powerpc64le -mcpu=pwr9 \
+; RUN: | FileCheck %s -check-prefix=P9
; RUN: llc -verify-machineinstrs < %s -mtriple=powerpc64le | FileCheck %s \
; RUN: -check-prefix=DM
-define i32 @foo() {
+define i32 @foo() #0 {
; PPC32-LABEL: foo:
; PPC32: # %bb.0: # %entry
; PPC32-NEXT: stwu 1, -32(1)
-; PPC32-NEXT: .cfi_def_cfa_offset 32
; PPC32-NEXT: mffs 0
; PPC32-NEXT: stfd 0, 16(1)
; PPC32-NEXT: lwz 3, 20(1)
@@ -51,6 +54,33 @@ define i32 @foo() {
; PPC64LE-NEXT: stw 3, -4(1)
; PPC64LE-NEXT: blr
;
+; P9_32-LABEL: foo:
+; P9_32: # %bb.0: # %entry
+; P9_32-NEXT: stwu 1, -32(1)
+; P9_32-NEXT: mffs 0
+; P9_32-NEXT: stfd 0, 16(1)
+; P9_32-NEXT: lwz 3, 20(1)
+; P9_32-NEXT: clrlwi 4, 3, 30
+; P9_32-NEXT: not 3, 3
+; P9_32-NEXT: rlwinm 3, 3, 31, 31, 31
+; P9_32-NEXT: xor 3, 4, 3
+; P9_32-NEXT: stw 3, 24(1)
+; P9_32-NEXT: stw 3, 28(1)
+; P9_32-NEXT: addi 1, 1, 32
+; P9_32-NEXT: blr
+;
+; P9-LABEL: foo:
+; P9: # %bb.0: # %entry
+; P9-NEXT: mffs 0
+; P9-NEXT: mffprd 3, 0
+; P9-NEXT: clrlwi 4, 3, 30
+; P9-NEXT: not 3, 3
+; P9-NEXT: rlwinm 3, 3, 31, 31, 31
+; P9-NEXT: xor 3, 4, 3
+; P9-NEXT: stw 3, -8(1)
+; P9-NEXT: stw 3, -4(1)
+; P9-NEXT: blr
+;
; DM-LABEL: foo:
; DM: # %bb.0: # %entry
; DM-NEXT: mffs 0
@@ -77,7 +107,7 @@ return: ; preds = %entry
ret i32 %retval3
}
-define void @setrnd_tozero() {
+define void @setrnd_tozero() #0 {
; PPC32-LABEL: setrnd_tozero:
; PPC32: # %bb.0: # %entry
; PPC32-NEXT: mtfsb0 30
@@ -96,6 +126,16 @@ define void @setrnd_tozero() {
; PPC64LE-NEXT: mtfsb1 31
; PPC64LE-NEXT: blr
;
+; P9_32-LABEL: setrnd_tozero:
+; P9_32: # %bb.0: # %entry
+; P9_32-NEXT: mffscrni 0, 1
+; P9_32-NEXT: blr
+;
+; P9-LABEL: setrnd_tozero:
+; P9: # %bb.0: # %entry
+; P9-NEXT: mffscrni 0, 1
+; P9-NEXT: blr
+;
; DM-LABEL: setrnd_tozero:
; DM: # %bb.0: # %entry
; DM-NEXT: mtfsb0 30
@@ -106,7 +146,7 @@ entry:
ret void
}
-define void @setrnd_tonearest_tieeven() {
+define void @setrnd_tonearest_tieeven() #0 {
; PPC32-LABEL: setrnd_tonearest_tieeven:
; PPC32: # %bb.0: # %entry
; PPC32-NEXT: mtfsb0 30
@@ -125,6 +165,16 @@ define void @setrnd_tonearest_tieeven() {
; PPC64LE-NEXT: mtfsb0 31
; PPC64LE-NEXT: blr
;
+; P9_32-LABEL: setrnd_tonearest_tieeven:
+; P9_32: # %bb.0: # %entry
+; P9_32-NEXT: mffscrni 0, 0
+; P9_32-NEXT: blr
+;
+; P9-LABEL: setrnd_tonearest_tieeven:
+; P9: # %bb.0: # %entry
+; P9-NEXT: mffscrni 0, 0
+; P9-NEXT: blr
+;
; DM-LABEL: setrnd_tonearest_tieeven:
; DM: # %bb.0: # %entry
; DM-NEXT: mtfsb0 30
@@ -135,7 +185,7 @@ entry:
ret void
}
-define void @setrnd_toposinf() {
+define void @setrnd_toposinf() #0 {
; PPC32-LABEL: setrnd_toposinf:
; PPC32: # %bb.0: # %entry
; PPC32-NEXT: mtfsb1 30
@@ -154,6 +204,16 @@ define void @setrnd_toposinf() {
; PPC64LE-NEXT: mtfsb0 31
; PPC64LE-NEXT: blr
;
+; P9_32-LABEL: setrnd_toposinf:
+; P9_32: # %bb.0: # %entry
+; P9_32-NEXT: mffscrni 0, 2
+; P9_32-NEXT: blr
+;
+; P9-LABEL: setrnd_toposinf:
+; P9: # %bb.0: # %entry
+; P9-NEXT: mffscrni 0, 2
+; P9-NEXT: blr
+;
; DM-LABEL: setrnd_toposinf:
; DM: # %bb.0: # %entry
; DM-NEXT: mtfsb1 30
@@ -164,7 +224,7 @@ entry:
ret void
}
-define void @setrnd_toneginf() {
+define void @setrnd_toneginf() #0 {
; PPC32-LABEL: setrnd_toneginf:
; PPC32: # %bb.0: # %entry
; PPC32-NEXT: mtfsb1 30
@@ -183,6 +243,16 @@ define void @setrnd_toneginf() {
; PPC64LE-NEXT: mtfsb1 31
; PPC64LE-NEXT: blr
;
+; P9_32-LABEL: setrnd_toneginf:
+; P9_32: # %bb.0: # %entry
+; P9_32-NEXT: mffscrni 0, 3
+; P9_32-NEXT: blr
+;
+; P9-LABEL: setrnd_toneginf:
+; P9: # %bb.0: # %entry
+; P9-NEXT: mffscrni 0, 3
+; P9-NEXT: blr
+;
; DM-LABEL: setrnd_toneginf:
; DM: # %bb.0: # %entry
; DM-NEXT: mtfsb1 30
@@ -193,11 +263,10 @@ entry:
ret void
}
-define void @setrnd_var(i32 %x) {
+define void @setrnd_var(i32 %x) #0 {
; PPC32-LABEL: setrnd_var:
; PPC32: # %bb.0: # %entry
; PPC32-NEXT: stwu 1, -16(1)
-; PPC32-NEXT: .cfi_def_cfa_offset 16
; PPC32-NEXT: mffs 0
; PPC32-NEXT: stfd 0, 8(1)
; PPC32-NEXT: clrlwi 4, 3, 30
@@ -247,6 +316,29 @@ define void @setrnd_var(i32 %x) {
; PPC64LE-NEXT: mtfsf 255, 0
; PPC64LE-NEXT: blr
;
+; P9_32-LABEL: setrnd_var:
+; P9_32: # %bb.0: # %entry
+; P9_32-NEXT: stwu 1, -16(1)
+; P9_32-NEXT: clrlwi 4, 3, 30
+; P9_32-NEXT: rlwinm 3, 3, 31, 31, 31
+; P9_32-NEXT: xor 3, 3, 4
+; P9_32-NEXT: xori 3, 3, 1
+; P9_32-NEXT: stw 3, 12(1)
+; P9_32-NEXT: lfd 0, 8(1)
+; P9_32-NEXT: mffscrn 0, 0
+; P9_32-NEXT: addi 1, 1, 16
+; P9_32-NEXT: blr
+;
+; P9-LABEL: setrnd_var:
+; P9: # %bb.0: # %entry
+; P9-NEXT: clrlwi 4, 3, 30
+; P9-NEXT: rlwinm 3, 3, 31, 31, 31
+; P9-NEXT: xor 3, 3, 4
+; P9-NEXT: xori 3, 3, 1
+; P9-NEXT: mtfprd 0, 3
+; P9-NEXT: mffscrn 0, 0
+; P9-NEXT: blr
+;
; DM-LABEL: setrnd_var:
; DM: # %bb.0: # %entry
; DM-NEXT: clrlwi 4, 3, 30
>From 9636dea38f886bbd9c4f197fbfa343ea59a88b12 Mon Sep 17 00:00:00 2001
From: Qiu Chaofan <qiucofan at cn.ibm.com>
Date: Wed, 10 Jan 2024 17:27:58 +0800
Subject: [PATCH 6/6] Address comments
---
llvm/lib/Target/PowerPC/PPCISelLowering.cpp | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/llvm/lib/Target/PowerPC/PPCISelLowering.cpp b/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
index 16dcb973d6c967..70ab531a6ee07a 100644
--- a/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
+++ b/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
@@ -8944,7 +8944,7 @@ SDValue PPCTargetLowering::LowerSET_ROUNDING(SDValue Op,
Chain = MFFS.getValue(1);
}
SDValue NewFPSCR;
- if (isTypeLegal(MVT::i64)) {
+ if (Subtarget.isPPC64()) {
if (Subtarget.isISA3_0())
NewFPSCR = DAG.getAnyExtOrTrunc(DstFlag, Dl, MVT::i64);
else
More information about the cfe-commits
mailing list