[RISCV] Support Zvfbfa codegen (#161158)

- [RISCV][llvm] Support Zvfbfa codegen and vsetvli insertion
- [RISCV][clang] Support Zvfbfa C intrinsics
The original PR is split into 2 PRs, this is codegen PR and other is
here: #164094

Co-authored-by: Craig Topper <[email protected]>

Author: 4vtomat

llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp

@@ -699,7 +699,8 @@ class VSETVLIInfo {
            "Can't encode VTYPE for uninitialized or unknown");
     if (TWiden != 0)
       return RISCVVType::encodeXSfmmVType(SEW, TWiden, AltFmt);
-    return RISCVVType::encodeVTYPE(VLMul, SEW, TailAgnostic, MaskAgnostic);
+    return RISCVVType::encodeVTYPE(VLMul, SEW, TailAgnostic, MaskAgnostic,
+                                   AltFmt);
   }
 
   bool hasSEWLMULRatioOnly() const { return SEWLMULRatioOnly; }

llvm/lib/Target/RISCV/RISCVInstrInfo.cpp

@@ -3775,11 +3775,13 @@ std::string RISCVInstrInfo::createMIROperandComment(
 
 #define CASE_VFMA_OPCODE_VV(OP)                                                \
   CASE_VFMA_OPCODE_LMULS_MF4(OP, VV, E16):                                     \
+  case CASE_VFMA_OPCODE_LMULS_MF4(OP##_ALT, VV, E16):                          \
   case CASE_VFMA_OPCODE_LMULS_MF2(OP, VV, E32):                                \
   case CASE_VFMA_OPCODE_LMULS_M1(OP, VV, E64)
 
 #define CASE_VFMA_SPLATS(OP)                                                   \
   CASE_VFMA_OPCODE_LMULS_MF4(OP, VFPR16, E16):                                 \
+  case CASE_VFMA_OPCODE_LMULS_MF4(OP##_ALT, VFPR16, E16):                      \
   case CASE_VFMA_OPCODE_LMULS_MF2(OP, VFPR32, E32):                            \
   case CASE_VFMA_OPCODE_LMULS_M1(OP, VFPR64, E64)
 // clang-format on
@@ -4003,11 +4005,13 @@ bool RISCVInstrInfo::findCommutedOpIndices(const MachineInstr &MI,
 
 #define CASE_VFMA_CHANGE_OPCODE_VV(OLDOP, NEWOP)                               \
   CASE_VFMA_CHANGE_OPCODE_LMULS_MF4(OLDOP, NEWOP, VV, E16)                     \
+  CASE_VFMA_CHANGE_OPCODE_LMULS_MF4(OLDOP##_ALT, NEWOP##_ALT, VV, E16)         \
   CASE_VFMA_CHANGE_OPCODE_LMULS_MF2(OLDOP, NEWOP, VV, E32)                     \
   CASE_VFMA_CHANGE_OPCODE_LMULS_M1(OLDOP, NEWOP, VV, E64)
 
 #define CASE_VFMA_CHANGE_OPCODE_SPLATS(OLDOP, NEWOP)                           \
   CASE_VFMA_CHANGE_OPCODE_LMULS_MF4(OLDOP, NEWOP, VFPR16, E16)                 \
+  CASE_VFMA_CHANGE_OPCODE_LMULS_MF4(OLDOP##_ALT, NEWOP##_ALT, VFPR16, E16)     \
   CASE_VFMA_CHANGE_OPCODE_LMULS_MF2(OLDOP, NEWOP, VFPR32, E32)                 \
   CASE_VFMA_CHANGE_OPCODE_LMULS_M1(OLDOP, NEWOP, VFPR64, E64)
 // clang-format on
@@ -4469,6 +4473,20 @@ bool RISCVInstrInfo::simplifyInstruction(MachineInstr &MI) const {
   CASE_FP_WIDEOP_CHANGE_OPCODE_COMMON(OP, M2, E32)                             \
   CASE_FP_WIDEOP_CHANGE_OPCODE_COMMON(OP, M4, E16)                             \
   CASE_FP_WIDEOP_CHANGE_OPCODE_COMMON(OP, M4, E32)                             \
+
+#define CASE_FP_WIDEOP_OPCODE_LMULS_ALT(OP)                                    \
+  CASE_FP_WIDEOP_OPCODE_COMMON(OP, MF4, E16):                                  \
+  case CASE_FP_WIDEOP_OPCODE_COMMON(OP, MF2, E16):                             \
+  case CASE_FP_WIDEOP_OPCODE_COMMON(OP, M1, E16):                              \
+  case CASE_FP_WIDEOP_OPCODE_COMMON(OP, M2, E16):                              \
+  case CASE_FP_WIDEOP_OPCODE_COMMON(OP, M4, E16)
+
+#define CASE_FP_WIDEOP_CHANGE_OPCODE_LMULS_ALT(OP)                             \
+  CASE_FP_WIDEOP_CHANGE_OPCODE_COMMON(OP, MF4, E16)                            \
+  CASE_FP_WIDEOP_CHANGE_OPCODE_COMMON(OP, MF2, E16)                            \
+  CASE_FP_WIDEOP_CHANGE_OPCODE_COMMON(OP, M1, E16)                             \
+  CASE_FP_WIDEOP_CHANGE_OPCODE_COMMON(OP, M2, E16)                             \
+  CASE_FP_WIDEOP_CHANGE_OPCODE_COMMON(OP, M4, E16)
 // clang-format on
 
 MachineInstr *RISCVInstrInfo::convertToThreeAddress(MachineInstr &MI,
@@ -4478,6 +4496,8 @@ MachineInstr *RISCVInstrInfo::convertToThreeAddress(MachineInstr &MI,
   switch (MI.getOpcode()) {
   default:
     return nullptr;
+  case CASE_FP_WIDEOP_OPCODE_LMULS_ALT(FWADD_ALT_WV):
+  case CASE_FP_WIDEOP_OPCODE_LMULS_ALT(FWSUB_ALT_WV):
   case CASE_FP_WIDEOP_OPCODE_LMULS(FWADD_WV):
   case CASE_FP_WIDEOP_OPCODE_LMULS(FWSUB_WV): {
     assert(RISCVII::hasVecPolicyOp(MI.getDesc().TSFlags) &&
@@ -4494,6 +4514,8 @@ MachineInstr *RISCVInstrInfo::convertToThreeAddress(MachineInstr &MI,
       llvm_unreachable("Unexpected opcode");
     CASE_FP_WIDEOP_CHANGE_OPCODE_LMULS(FWADD_WV)
     CASE_FP_WIDEOP_CHANGE_OPCODE_LMULS(FWSUB_WV)
+    CASE_FP_WIDEOP_CHANGE_OPCODE_LMULS_ALT(FWADD_ALT_WV)
+    CASE_FP_WIDEOP_CHANGE_OPCODE_LMULS_ALT(FWSUB_ALT_WV)
     }
     // clang-format on