phoebewang wrote:
> Looks like this causes a significant compile-time regression, but only for
> ReleaseLTO-g:
> https://llvm-compile-time-tracker.com/compare.php?from=1e19f0f9d92b5e9c43d53893e387341835d3d96b&to=c72a751dabff4260dcc309e48008941d51b31d21&stat=instructions:u
>
> I wouldn't really
nikic wrote:
Looks like this causes a significant compile-time regression, but only for
ReleaseLTO-g:
https://llvm-compile-time-tracker.com/compare.php?from=1e19f0f9d92b5e9c43d53893e387341835d3d96b&to=c72a751dabff4260dcc309e48008941d51b31d21&stat=instructions:u
I wouldn't really expect this ch
@@ -919,23 +1017,66 @@ bool X86LowerAMXCast::optimizeAMXCastFromPhi(
return true;
}
+static Value *getShapeFromAMXIntrinsic(Value *Inst, unsigned ShapeIdx,
+ bool IsRow) {
+ if (!isAMXIntrinsic(Inst))
+return nullptr;
+
+ auto *II
https://github.com/fzou1 approved this pull request.
LGTM
https://github.com/llvm/llvm-project/pull/113532
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
https://github.com/phoebewang closed
https://github.com/llvm/llvm-project/pull/113532
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
https://github.com/fzou1 deleted
https://github.com/llvm/llvm-project/pull/113532
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
https://github.com/phoebewang edited
https://github.com/llvm/llvm-project/pull/113532
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
@@ -0,0 +1,248 @@
+/* ===--- amxtransposeintrin.h - AMX_TRANSPOSE intrinsics -*- C++
-*-===
+ *
+ * Part of the LLVM Project, under the Apache License v2.0 with LLVM
Exceptions.
+ * See https://llvm.org/LICENSE.txt for license information.
+ * SPDX-License-Identifier: Ap
@@ -0,0 +1,248 @@
+/* ===--- amxtransposeintrin.h - AMX_TRANSPOSE intrinsics -*- C++
-*-===
+ *
+ * Part of the LLVM Project, under the Apache License v2.0 with LLVM
Exceptions.
+ * See https://llvm.org/LICENSE.txt for license information.
+ * SPDX-License-Identifier: Ap
@@ -121,12 +137,96 @@ static Instruction
*getFirstNonAllocaInTheEntryBlock(Function &F) {
llvm_unreachable("No terminator in the entry block!");
}
-static std::pair getShape(IntrinsicInst *II, unsigned OpNo) {
+class ShapeCalculator {
+private:
+ TargetMachine *TM = nullpt
@@ -623,6 +623,37 @@ struct X86Operand final : public MCParsedAsmOperand {
Inst.addOperand(MCOperand::createReg(Reg));
}
+ bool isTILEPair() const {
+return Kind == Register &&
+ X86MCRegisterClasses[X86::TILERegClassID].contains(getReg());
---
@@ -34,9 +34,31 @@ class ShapeT {
if (MRI)
deduceImm(MRI);
}
+ // When ShapeT has mult shapes, we only use Shapes (never use Row and Col)
phoebewang wrote:
Done.
https://github.com/llvm/llvm-project/pull/113532
___
@@ -34,9 +34,31 @@ class ShapeT {
if (MRI)
deduceImm(MRI);
}
+ // When ShapeT has mult shapes, we only use Shapes (never use Row and Col)
+ // and ImmShapes. Due to the most case is only one shape (just simply use
+ // Shape.Row or Shape.Col), so here we don't me
@@ -16920,6 +16920,58 @@ Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned
BuiltinID,
// instruction, but it will create a memset that won't be optimized away.
return Builder.CreateMemSet(Ops[0], Ops[1], Ops[2], Align(1), true);
}
+ // Corresponding to intrisics w
@@ -0,0 +1,248 @@
+/* ===--- amxtransposeintrin.h - AMX_TRANSPOSE intrinsics -*- C++
-*-===
+ *
+ * Part of the LLVM Project, under the Apache License v2.0 with LLVM
Exceptions.
+ * See https://llvm.org/LICENSE.txt for license information.
+ * SPDX-License-Identifier: Ap
@@ -0,0 +1,248 @@
+/* ===--- amxtransposeintrin.h - AMX_TRANSPOSE intrinsics -*- C++
-*-===
+ *
+ * Part of the LLVM Project, under the Apache License v2.0 with LLVM
Exceptions.
+ * See https://llvm.org/LICENSE.txt for license information.
+ * SPDX-License-Identifier: Ap
@@ -34,9 +34,31 @@ class ShapeT {
if (MRI)
deduceImm(MRI);
}
+ // When ShapeT has mult shapes, we only use Shapes (never use Row and Col)
+ // and ImmShapes. Due to the most case is only one shape (just simply use
+ // Shape.Row or Shape.Col), so here we don't me
@@ -623,6 +623,37 @@ struct X86Operand final : public MCParsedAsmOperand {
Inst.addOperand(MCOperand::createReg(Reg));
}
+ bool isTILEPair() const {
+return Kind == Register &&
+ X86MCRegisterClasses[X86::TILERegClassID].contains(getReg());
---
@@ -0,0 +1,248 @@
+/* ===--- amxtransposeintrin.h - AMX_TRANSPOSE intrinsics -*- C++
-*-===
+ *
+ * Part of the LLVM Project, under the Apache License v2.0 with LLVM
Exceptions.
+ * See https://llvm.org/LICENSE.txt for license information.
+ * SPDX-License-Identifier: Ap
@@ -919,23 +1017,66 @@ bool X86LowerAMXCast::optimizeAMXCastFromPhi(
return true;
}
+static Value *getShapeFromAMXIntrinsic(Value *Inst, unsigned ShapeIdx,
+ bool IsRow) {
+ if (!isAMXIntrinsic(Inst))
+return nullptr;
+
+ auto *II
@@ -34,9 +34,31 @@ class ShapeT {
if (MRI)
deduceImm(MRI);
}
+ // When ShapeT has mult shapes, we only use Shapes (never use Row and Col)
fzou1 wrote:
mult -> multiple
https://github.com/llvm/llvm-project/pull/113532
_
@@ -121,12 +137,96 @@ static Instruction
*getFirstNonAllocaInTheEntryBlock(Function &F) {
llvm_unreachable("No terminator in the entry block!");
}
-static std::pair getShape(IntrinsicInst *II, unsigned OpNo) {
+class ShapeCalculator {
+private:
+ TargetMachine *TM = nullpt
@@ -16920,6 +16920,58 @@ Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned
BuiltinID,
// instruction, but it will create a memset that won't be optimized away.
return Builder.CreateMemSet(Ops[0], Ops[1], Ops[2], Align(1), true);
}
+ // Corresponding to intrisics w
@@ -0,0 +1,248 @@
+/* ===--- amxtransposeintrin.h - AMX_TRANSPOSE intrinsics -*- C++
-*-===
+ *
+ * Part of the LLVM Project, under the Apache License v2.0 with LLVM
Exceptions.
+ * See https://llvm.org/LICENSE.txt for license information.
+ * SPDX-License-Identifier: Ap
@@ -80,28 +80,41 @@ INITIALIZE_PASS_BEGIN(X86FastTileConfig, DEBUG_TYPE,
INITIALIZE_PASS_END(X86FastTileConfig, DEBUG_TYPE,
"Fast Tile Register Configure", false, false)
-static bool isTileDef(MachineRegisterInfo *MRI, MachineInstr &MI) {
+static unsigned g
@@ -80,28 +80,41 @@ INITIALIZE_PASS_BEGIN(X86FastTileConfig, DEBUG_TYPE,
INITIALIZE_PASS_END(X86FastTileConfig, DEBUG_TYPE,
"Fast Tile Register Configure", false, false)
-static bool isTileDef(MachineRegisterInfo *MRI, MachineInstr &MI) {
+static unsigned g
@@ -34,9 +34,31 @@ class ShapeT {
if (MRI)
deduceImm(MRI);
}
+ // When ShapeT has mult shapes, we only use Shapes (never use Row and Col)
+ // and ImmShapes. Due to the most case is only one shape (just simply use
+ // Shape.Row or Shape.Col), so here we don't me
@@ -568,6 +568,131 @@ bool X86ExpandPseudo::expandMI(MachineBasicBlock &MBB,
MI.setDesc(TII->get(Opc));
return true;
}
+ // TILEPAIRLOAD is just for TILEPair spill, we don't have corresponding
+ // AMX instruction to support it. So, split it to 2 load instructions:
@@ -568,6 +568,131 @@ bool X86ExpandPseudo::expandMI(MachineBasicBlock &MBB,
MI.setDesc(TII->get(Opc));
return true;
}
+ // TILEPAIRLOAD is just for TILEPair spill, we don't have corresponding
+ // AMX instruction to support it. So, split it to 2 load instructions:
@@ -80,28 +80,41 @@ INITIALIZE_PASS_BEGIN(X86FastTileConfig, DEBUG_TYPE,
INITIALIZE_PASS_END(X86FastTileConfig, DEBUG_TYPE,
"Fast Tile Register Configure", false, false)
-static bool isTileDef(MachineRegisterInfo *MRI, MachineInstr &MI) {
+static unsigned g
@@ -34,9 +34,31 @@ class ShapeT {
if (MRI)
deduceImm(MRI);
}
+ // When ShapeT has mult shapes, we only use Shapes (never use Row and Col)
+ // and ImmShapes. Due to the most case is only one shape (just simply use
+ // Shape.Row or Shape.Col), so here we don't me
llvmbot wrote:
@llvm/pr-subscribers-mc
Author: Phoebe Wang (phoebewang)
Changes
Ref.: https://cdrdv2.intel.com/v1/dl/getContent/671368
---
Patch is 184.18 KiB, truncated to 20.00 KiB below, full version:
https://github.com/llvm/llvm-project/pull/113532.diff
57 Files Affected:
- (modi
llvmbot wrote:
@llvm/pr-subscribers-llvm-ir
Author: Phoebe Wang (phoebewang)
Changes
Ref.: https://cdrdv2.intel.com/v1/dl/getContent/671368
---
Patch is 184.18 KiB, truncated to 20.00 KiB below, full version:
https://github.com/llvm/llvm-project/pull/113532.diff
57 Files Affected:
-
llvmbot wrote:
@llvm/pr-subscribers-tablegen
Author: Phoebe Wang (phoebewang)
Changes
Ref.: https://cdrdv2.intel.com/v1/dl/getContent/671368
---
Patch is 184.18 KiB, truncated to 20.00 KiB below, full version:
https://github.com/llvm/llvm-project/pull/113532.diff
57 Files Affected:
-
llvmbot wrote:
@llvm/pr-subscribers-clang
@llvm/pr-subscribers-clang-codegen
Author: Phoebe Wang (phoebewang)
Changes
Ref.: https://cdrdv2.intel.com/v1/dl/getContent/671368
---
Patch is 184.18 KiB, truncated to 20.00 KiB below, full version:
https://github.com/llvm/llvm-project/pull/113
35 matches
Mail list logo
