[RISCV][MC] Add support for experimental Zfbfmin extension

[asb]

Summary

As documented, this extension includes FLH, FSH, FMV.H.X, and FMH.X.H as defined in Zfh/Zfhmin, but doesn't require either extension.

No Zfbfinxmin has been defined (though you would expect one in the future, for symmetry with Zfhinxmin). See issue https://github.com/riscv/riscv-bfloat16/issues/27.

[craig.topper]

LGTM

[asb]

Rebase and use the new encoding suggested in https://github.com/riscv/riscv-bfloat16/issues/33 - hoping for the commits that alter the encoding to be pushed to the spec repo and a new PDF (with new version) to be generated ahead of landing this.

[joshua-arch1]

This comment was removed by joshua-arch1.

[joshua-arch1]

I'm wondering whether it is appropriate to just use FPR16 for the destination of fcvt.bf16.s? The destination is in BF16 format instead of simple FP16. Your implemention looks like just replacing fcvt.h.s with fcvt.bf16.s. Do we need to define a new register class?

[craig.topper]

In D147610#4294247, @joshua-arch1 wrote:

I'm wondering whether it is appropriate to just use FPR16 for the destination of fcvt.bf16.s? The destination is in BF16 format instead of simple FP16. Your implemention looks like just replacing fcvt.h.s with fcvt.bf16.s. Do we need to define a new register class?

Registers classes only distinguished by the registers in them and what their alignment and spill size are. It doesn't define anything about the format of the register. A BF16 specific register class would be identical in those properties to the FPR16 register class.

[joshua-arch1]

In D147610#4294260, @craig.topper wrote:

In D147610#4294247, @joshua-arch1 wrote:

I'm wondering whether it is appropriate to just use FPR16 for the destination of fcvt.bf16.s? The destination is in BF16 format instead of simple FP16. Your implemention looks like just replacing fcvt.h.s with fcvt.bf16.s. Do we need to define a new register class?

Registers classes only distinguished by the registers in them and what their alignment and spill size are. It doesn't define anything about the format of the register. A BF16 specific register class would be identical in those properties to the FPR16 register class.

Do you have any plan for code generation of vector instructions with corresponding intrinsics?

[craig.topper]

In D147610#4294322, @joshua-arch1 wrote:

In D147610#4294260, @craig.topper wrote:

In D147610#4294247, @joshua-arch1 wrote:

I'm wondering whether it is appropriate to just use FPR16 for the destination of fcvt.bf16.s? The destination is in BF16 format instead of simple FP16. Your implemention looks like just replacing fcvt.h.s with fcvt.bf16.s. Do we need to define a new register class?

Registers classes only distinguished by the registers in them and what their alignment and spill size are. It doesn't define anything about the format of the register. A BF16 specific register class would be identical in those properties to the FPR16 register class.

Do you have any plan for code generation of these instructions with corresponding intrinsics?

The RVI toolchain SIG is supposed to be setting up a task group to define intrinsics for all extensions.

[joshua-arch1]

The RVI toolchain SIG is supposed to be setting up a task group to define intrinsics for all extensions.

Where should I discuss this intrinsic issue right now?

[craig.topper]

In D147610#4297533, @joshua-arch1 wrote:

The RVI toolchain SIG is supposed to be setting up a task group to define intrinsics for all extensions.

Where should I discuss this intrinsic issue right now?

I guess https://github.com/riscv-non-isa/riscv-c-api-doc

[asb]

Now updated to reflect v0.6 of the spec (was previously blocked on a tagged PDF being uploaded) and should be ready for final review and merge https://github.com/riscv/riscv-bfloat16/releases/tag/main

[reames]

LGTM

[joshua-arch1]

Hi, I'm adding support for __bf16 in clang, which will rely on FeatureStdExtZfbfmin. Can you merge this patch?

clang/test/Preprocessor/riscv-target-features.c

 // CHECK-NOT: __riscv_zvksed {{.*$}}
 // CHECK-NOT: __riscv_zvksh {{.*$}}
 // CHECK-NOT: __riscv_zicond {{.*$}}
+// CHECK-NOT: __riscv_zfbfmin {{.*$}}
 
 // RUN: %clang -target riscv32-unknown-linux-gnu -march=rv32i -x c -E -dM %s \
 // RUN: -o - | FileCheck %s

 // RUN: -march=rv64i_zicond1p0 -x c -E -dM %s \
 // RUN: -o - | FileCheck --check-prefix=CHECK-ZICOND-EXT %s
 // CHECK-ZICOND-EXT: __riscv_zicond  1000000{{$}}
+
+// RUN: %clang -target riscv32 -menable-experimental-extensions \
+// RUN: -march=rv32izfbfmin0p2 -x c -E -dM %s \
+// RUN: -o - | FileCheck --check-prefix=CHECK-ZFBFMIN-EXT %s
+// RUN: %clang -target riscv64 -menable-experimental-extensions \
+// RUN: -march=rv64izfbfmin0p2 -x c -E -dM %s \
+// RUN: -o - | FileCheck --check-prefix=CHECK-ZFBFMIN-EXT %s
+// CHECK-ZFBFMIN-EXT: __riscv_zfbfmin 2000{{$}}

llvm/docs/RISCVUsage.rst

 ``experimental-zfa``
   LLVM implements a subset of `0.1 draft specification <https://github.com/riscv/riscv-isa-manual/releases/download/draft-20221119-5234c63/riscv-spec.pdf>`_ (see Chapter 25).
 
+``experimental-zfbfmin``
+  LLVM implements assembler support for the `20230322 draft specification <https://github.com/riscv/riscv-bfloat16/releases/tag/20230322>`_.
+
 ``experimental-zicond``
   LLVM implements the `1.0-rc1 draft specification <https://github.com/riscv/riscv-zicond/releases/tag/v1.0-rc1`>_.
 

llvm/docs/ReleaseNotes.rst

 * I, F, D, and A extension versions have been update to the 20191214 spec versions.
   New version I2.1, F2.2, D2.2, A2.1. This should not impact code generation.
   Immpacts versions accepted in ``-march`` and reported in ELF attributes.
+* Assembly support was added for the experimental Zfbfmin (scalar BF16
+  conversions) extension.
 
 Changes to the WebAssembly Backend
 ----------------------------------

llvm/lib/Support/RISCVISAInfo.cpp

     {"zcd", RISCVExtensionVersion{1, 0}},
     {"zcf", RISCVExtensionVersion{1, 0}},
     {"zfa", RISCVExtensionVersion{0, 1}},
+    {"zfbfmin", RISCVExtensionVersion{0, 2}},
     {"zicond", RISCVExtensionVersion{1, 0}},
     {"zvfh", RISCVExtensionVersion{0, 1}},
     {"ztso", RISCVExtensionVersion{0, 1}},

llvm/lib/Target/RISCV/RISCVFeatures.td

                                 AssemblerPredicate<(all_of FeatureStdExtZicond),
                                 "'Zicond' (Integer Conditional Operations)">;
 
+def FeatureStdExtZfbfmin
+    : SubtargetFeature<"experimental-zfbfmin", "HasStdExtZfbfmin", "true",
+                       "'Zfbfmin' (Scalar BF16 Converts)",
+                       [FeatureStdExtF]>;
+def HasStdExtZfbfmin : Predicate<"Subtarget->hasStdExtZfbfmin()">,
+                                 AssemblerPredicate<(all_of FeatureStdExtZfbfmin),
+                                 "'Zfbfmin' (Scalar BF16 Converts)">;
+
+def HasHalfFPLoadStoreMove
+    : Predicate<"Subtarget->hasHalfFPLoadStoreMove()">,
+                AssemblerPredicate<(any_of FeatureStdExtZfh, FeatureStdExtZfhmin, FeatureStdExtZfbfmin),
+                                    "'Zfh' (Half-Precision Floating-Point) or "
+                                    "'Zfhmin' (Half-Precision Floating-Point Minimal) or "
+                                    "'Zfbfmin' (Scalar BF16 Converts)">;
+
 //===----------------------------------------------------------------------===//
 // Vendor extensions
 //===----------------------------------------------------------------------===//

llvm/lib/Target/RISCV/RISCVInstrInfo.td

 include "RISCVInstrInfoZk.td"
 include "RISCVInstrInfoV.td"
 include "RISCVInstrInfoZfa.td"
+include "RISCVInstrInfoZfbfmin.td"
 include "RISCVInstrInfoZfh.td"
 include "RISCVInstrInfoZicbo.td"
 include "RISCVInstrInfoZicond.td"

llvm/lib/Target/RISCV/RISCVInstrInfoZfbfmin.td

+//===-- RISCVInstrInfoZfbfmin.td - 'Zfbfmin' instructions --*- tablegen -*-===//
+//
+// 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: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file describes the RISC-V instructions from the standard 'Zfbfmin'
+// extension, providing scalar conversion instructions for BFloat16.
+// This version is still experimental as the 'Zfbfmin' extension hasn't been
+// ratified yet.
+//
+//===----------------------------------------------------------------------===//
+
+//===----------------------------------------------------------------------===//
+// Instructions
+//===----------------------------------------------------------------------===//
+
+let Predicates = [HasStdExtZfbfmin] in {
+def FCVT_BF16_S : FPUnaryOp_r_frm<0b0100010, 0b01000, FPR16, FPR32, "fcvt.bf16.s">,
+                  Sched<[WriteFCvtF32ToF16, ReadFCvtF32ToF16]>;
+def FCVT_S_BF16 : FPUnaryOp_r_frm<0b0001000, 0b00110, FPR32, FPR16, "fcvt.s.bf16">,
+                  Sched<[WriteFCvtF32ToF16, ReadFCvtF32ToF16]>;
+} // Predicates = [HasStdExtZfbfmin]

llvm/lib/Target/RISCV/RISCVInstrInfoZfh.td

 // Instructions
 //===----------------------------------------------------------------------===//
 
-let Predicates = [HasStdExtZfhOrZfhmin] in {
+let Predicates = [HasHalfFPLoadStoreMove] in {
 def FLH : FPLoad_r<0b001, "flh", FPR16, WriteFLD16>;
 
 // Operands for stores are in the order srcreg, base, offset rather than
 // reflecting the order these fields are specified in the instruction
 // encoding.
 def FSH : FPStore_r<0b001, "fsh", FPR16, WriteFST16>;
-} // Predicates = [HasStdExtZfhOrZfhmin]
+} // Predicates = [HasHalfFPLoadStoreMove]
 
 let SchedRW = [WriteFMA16, ReadFMA16, ReadFMA16, ReadFMA16] in {
 defm FMADD_H  : FPFMA_rrr_frm_m<OPC_MADD,  0b10, "fmadd.h",  HINX>;

 defm FCVT_S_H : FPUnaryOp_r_m<0b0100000, 0b00010, 0b000, FHINXmin, "fcvt.s.h">,
                Sched<[WriteFCvtF16ToF32, ReadFCvtF16ToF32]>;
 
-let Predicates = [HasStdExtZfhOrZfhmin] in {
+let Predicates = [HasHalfFPLoadStoreMove] in {
 let mayRaiseFPException = 0, IsSignExtendingOpW = 1 in
 def FMV_X_H : FPUnaryOp_r<0b1110010, 0b00000, 0b000, GPR, FPR16, "fmv.x.h">,
               Sched<[WriteFMovF16ToI16, ReadFMovF16ToI16]>;

 let mayRaiseFPException = 0 in
 def FMV_H_X : FPUnaryOp_r<0b1111010, 0b00000, 0b000, FPR16, GPR, "fmv.h.x">,
               Sched<[WriteFMovI16ToF16, ReadFMovI16ToF16]>;
-} // Predicates = [HasStdExtZfhOrZfhmin]
+} // Predicates = [HasHalfFPLoadStoreMove]
 
 let SchedRW = [WriteFCmp16, ReadFCmp16, ReadFCmp16] in {
 defm FEQ_H : FPCmp_rr_m<0b1010010, 0b010, "feq.h", HINX, /*Commutable*/1>;

llvm/lib/Target/RISCV/RISCVSubtarget.h

   bool hasStdExtCOrZca() const { return HasStdExtC || HasStdExtZca; }
   bool hasStdExtZvl() const { return ZvlLen != 0; }
   bool hasStdExtZfhOrZfhmin() const { return HasStdExtZfh || HasStdExtZfhmin; }
+  bool hasHalfFPLoadStoreMove() const {
+    return HasStdExtZfh || HasStdExtZfhmin || HasStdExtZfbfmin;
+  }
   bool is64Bit() const { return IsRV64; }
   MVT getXLenVT() const { return XLenVT; }
   unsigned getXLen() const { return XLen; }

llvm/test/CodeGen/RISCV/attributes.ll

 ; RUN: llc -mtriple=riscv32 -mattr=+zve32x -mattr=+experimental-zvksed %s -o - | FileCheck --check-prefix=RV32ZVKSED %s
 ; RUN: llc -mtriple=riscv32 -mattr=+zve32x -mattr=+experimental-zvksh %s -o - | FileCheck --check-prefix=RV32ZVKSH %s
 ; RUN: llc -mtriple=riscv32 -mattr=+experimental-zicond %s -o - | FileCheck --check-prefix=RV32ZICOND %s
+; RUN: llc -mtriple=riscv32 -mattr=+experimental-zfbfmin %s -o - | FileCheck --check-prefixes=CHECK,RV32ZFBFMIN %s
 
 ; RUN: llc -mtriple=riscv64 %s -o - | FileCheck %s
 ; RUN: llc -mtriple=riscv64 -mattr=+m %s -o - | FileCheck --check-prefixes=CHECK,RV64M %s

 ; RUN: llc -mtriple=riscv64 -mattr=+zve32x -mattr=+experimental-zvksed %s -o - | FileCheck --check-prefix=RV64ZVKSED %s
 ; RUN: llc -mtriple=riscv64 -mattr=+zve32x -mattr=+experimental-zvksh %s -o - | FileCheck --check-prefix=RV64ZVKSH %s
 ; RUN: llc -mtriple=riscv64 -mattr=+experimental-zicond %s -o - | FileCheck --check-prefix=RV64ZICOND %s
+; RUN: llc -mtriple=riscv64 -mattr=+experimental-zfbfmin %s -o - | FileCheck --check-prefixes=CHECK,RV64ZFBFMIN %s
 
 ; CHECK: .attribute 4, 16
 

 ; RV32ZVKSED: .attribute 5, "rv32i2p1_zicsr2p0_zve32x1p0_zvksed0p3_zvl32b1p0"
 ; RV32ZVKSH: .attribute 5, "rv32i2p1_zicsr2p0_zve32x1p0_zvksh0p3_zvl32b1p0"
 ; RV32ZICOND: .attribute 5, "rv32i2p1_zicond1p0"
+; RV32ZFBFMIN: .attribute 5, "rv32i2p1_f2p2_zicsr2p0_zfbfmin0p2"
 
 ; RV64M: .attribute 5, "rv64i2p1_m2p0"
 ; RV64ZMMUL: .attribute 5, "rv64i2p1_zmmul1p0"

 ; RV64ZVKSED: .attribute 5, "rv64i2p1_zicsr2p0_zve32x1p0_zvksed0p3_zvl32b1p0"
 ; RV64ZVKSH: .attribute 5, "rv64i2p1_zicsr2p0_zve32x1p0_zvksh0p3_zvl32b1p0"
 ; RV64ZICOND: .attribute 5, "rv64i2p1_zicond1p0"
+; RV64ZFBFMIN: .attribute 5, "rv64i2p1_f2p2_zicsr2p0_zfbfmin0p2"
 
 define i32 @addi(i32 %a) {
   %1 = add i32 %a, 1

llvm/test/MC/RISCV/attribute-arch.s

 
 .attribute arch, "rv32izicond1p0"
 # CHECK: attribute      5, "rv32i2p1_zicond1p0"
+
+.attribute arch, "rv32if_zfbfmin0p2"
+# CHECK: .attribute     5, "rv32i2p1_f2p2_zicsr2p0_zfbfmin0p2"

llvm/test/MC/RISCV/rv32zfbfmin-invalid.s

+# RUN: not llvm-mc -triple riscv32 -mattr=+experimental-zfbfmin,+d < %s 2>&1 | \
+# RUN:   FileCheck %s
+# RUN: not llvm-mc -triple riscv64 -mattr=+experimental-zfbfmin,+d < %s 2>&1 | \
+# RUN:   FileCheck %s
+
+# Out of range immediates
+## simm12
+flh ft1, -2049(a0) # CHECK: :[[@LINE]]:10: error: operand must be a symbol with %lo/%pcrel_lo/%tprel_lo modifier or an integer in the range [-2048, 2047]
+fsh ft2, 2048(a1) # CHECK: :[[@LINE]]:10: error: operand must be a symbol with %lo/%pcrel_lo/%tprel_lo modifier or an integer in the range [-2048, 2047]
+
+# Memory operand not formatted correctly
+flh ft1, a0, -200 # CHECK: :[[@LINE]]:14: error: invalid operand for instruction
+
+# Invalid register names
+flh ft15, 100(a0) # CHECK: :[[@LINE]]:5: error: invalid operand for instruction
+flh ft1, 100(a10) # CHECK: :[[@LINE]]:14: error: expected register
+
+# Integer registers where FP regs are expected
+fmv.x.h fs7, a2 # CHECK: :[[@LINE]]:9: error: invalid operand for instruction
+
+# FP registers where integer regs are expected
+fmv.h.x a8, ft2 # CHECK: :[[@LINE]]:9: error: invalid operand for instruction
+
+# Attempting to use fcvt instructions from zfhmin
+fcvt.s.h fa0, ft0 # CHECK: [[@LINE]]:1: error: instruction requires the following: 'Zfh' (Half-Precision Floating-Point) or 'Zfhmin' (Half-Precision Floating-Point Minimal)
+fcvt.h.s ft2, fa2 # CHECK: [[@LINE]]:1: error: instruction requires the following: 'Zfh' (Half-Precision Floating-Point) or 'Zfhmin' (Half-Precision Floating-Point Minimal)
+fcvt.d.h fa0, ft0 # CHECK: [[@LINE]]:1: error: instruction requires the following: 'Zfh' (Half-Precision Floating-Point) or 'Zfhmin' (Half-Precision Floating-Point Minimal)
+fcvt.h.d ft2, fa2 # CHECK: [[@LINE]]:1: error: instruction requires the following: 'Zfh' (Half-Precision Floating-Point) or 'Zfhmin' (Half-Precision Floating-Point Minimal)

llvm/test/MC/RISCV/rv32zfbfmin-valid.s

+# RUN: llvm-mc %s -triple=riscv32 -mattr=+experimental-zfbfmin,+f -riscv-no-aliases -show-encoding \
+# RUN:     | FileCheck -check-prefixes=CHECK-ASM,CHECK-ASM-AND-OBJ %s
+# RUN: llvm-mc %s -triple=riscv64 -mattr=+experimental-zfbfmin,+f -riscv-no-aliases -show-encoding \
+# RUN:     | FileCheck -check-prefixes=CHECK-ASM,CHECK-ASM-AND-OBJ %s
+# RUN: llvm-mc -filetype=obj -triple=riscv32 -mattr=+experimental-zfbfmin,+d < %s \
+# RUN:     | llvm-objdump --mattr=+experimental-zfbfmin,+f -M no-aliases -d -r - \
+# RUN:     | FileCheck --check-prefix=CHECK-ASM-AND-OBJ %s
+# RUN: llvm-mc -filetype=obj -triple=riscv64 -mattr=+experimental-zfbfmin,+d < %s \
+# RUN:     | llvm-objdump --mattr=+experimental-zfbfmin,+f -M no-aliases -d -r - \
+# RUN:     | FileCheck --check-prefix=CHECK-ASM-AND-OBJ %s
+
+# CHECK-ASM-AND-OBJ: flh ft0, 12(a0)
+# CHECK-ASM: encoding: [0x07,0x10,0xc5,0x00]
+flh f0, 12(a0)
+# CHECK-ASM-AND-OBJ: flh ft1, 4(ra)
+# CHECK-ASM: encoding: [0x87,0x90,0x40,0x00]
+flh f1, +4(ra)
+# CHECK-ASM-AND-OBJ: flh ft2, -2048(a3)
+# CHECK-ASM: encoding: [0x07,0x91,0x06,0x80]
+flh f2, -2048(x13)
+# CHECK-ASM-AND-OBJ: flh ft3, -2048(s1)
+# CHECK-ASM: encoding: [0x87,0x91,0x04,0x80]
+flh f3, %lo(2048)(s1)
+# CHECK-ASM-AND-OBJ: flh ft4, 2047(s2)
+# CHECK-ASM: encoding: [0x07,0x12,0xf9,0x7f]
+flh f4, 2047(s2)
+# CHECK-ASM-AND-OBJ: flh ft5, 0(s3)
+# CHECK-ASM: encoding: [0x87,0x92,0x09,0x00]
+flh f5, 0(s3)
+
+# CHECK-ASM-AND-OBJ: fsh ft6, 2047(s4)
+# CHECK-ASM: encoding: [0xa7,0x1f,0x6a,0x7e]
+fsh f6, 2047(s4)
+# CHECK-ASM-AND-OBJ: fsh ft7, -2048(s5)
+# CHECK-ASM: encoding: [0x27,0x90,0x7a,0x80]
+fsh f7, -2048(s5)
+# CHECK-ASM-AND-OBJ: fsh fs0, -2048(s6)
+# CHECK-ASM: encoding: [0x27,0x10,0x8b,0x80]
+fsh f8, %lo(2048)(s6)
+# CHECK-ASM-AND-OBJ: fsh fs1, 999(s7)
+# CHECK-ASM: encoding: [0xa7,0x93,0x9b,0x3e]
+fsh f9, 999(s7)
+
+# CHECK-ASM-AND-OBJ: fmv.x.h a2, fs7
+# CHECK-ASM: encoding: [0x53,0x86,0x0b,0xe4]
+fmv.x.h a2, fs7
+# CHECK-ASM-AND-OBJ: fmv.h.x ft1, a6
+# CHECK-ASM: encoding: [0xd3,0x00,0x08,0xf4]
+fmv.h.x ft1, a6
+
+# CHECK-ASM-AND-OBJ: fcvt.s.bf16 fa0, ft0
+# CHECK-ASM: encoding: [0x53,0x75,0x60,0x10]
+fcvt.s.bf16 fa0, ft0
+# CHECK-ASM-AND-OBJ: fcvt.bf16.s ft2, fa2
+# CHECK-ASM: encoding: [0x53,0x71,0x86,0x44]
+fcvt.bf16.s ft2, fa2

llvm/test/MC/RISCV/rv64zhinx-invalid.s

 # RUN: not llvm-mc -triple riscv64 -mattr=+zhinx %s 2>&1 | FileCheck %s
 
 # Not support float registers
-flh fa4, 12(sp) # CHECK: :[[@LINE]]:1: error: instruction requires the following: 'Zfh' (Half-Precision Floating-Point) or 'Zfhmin' (Half-Precision Floating-Point Minimal){{$}}
+flh fa4, 12(sp) # CHECK: :[[@LINE]]:1: error: instruction requires the following: 'Zfh' (Half-Precision Floating-Point) or 'Zfhmin' (Half-Precision Floating-Point Minimal) or 'Zfbfmin' (Scalar BF16 Converts){{$}}
 
 # Invalid instructions
 fsh a5, 12(sp) # CHECK: :[[@LINE]]:5: error: invalid operand for instruction

llvm/test/MC/RISCV/rv64zhinxmin-invalid.s

 # RUN: not llvm-mc -triple riscv64 -mattr=+zhinxmin %s 2>&1 | FileCheck %s
 
 # Not support float registers
-flh fa4, 12(sp) # CHECK: :[[@LINE]]:1: error: instruction requires the following: 'Zfh' (Half-Precision Floating-Point) or 'Zfhmin' (Half-Precision Floating-Point Minimal){{$}}
+flh fa4, 12(sp) # CHECK: :[[@LINE]]:1: error: instruction requires the following: 'Zfh' (Half-Precision Floating-Point) or 'Zfhmin' (Half-Precision Floating-Point Minimal) or 'Zfbfmin' (Scalar BF16 Converts){{$}}
 
 # Invalid instructions
 fsh a5, 12(sp) # CHECK: :[[@LINE]]:5: error: invalid operand for instruction