[RISCV] Separate addend from FMA operands to support cascade FMA. NFC. (#70241)

This PR separate addend from FMA operands to support cascade FMA. In
some microarchitectures (e.g., ARM cortex-a72 and XiangShan-NanHu), FP
multiply-accumulate pipelines support late-forwarding of accumulate
operands, which reduces the latency of a sequence of multiply-accumulate
instructions.
See also #70232.

Author: dtcxzyw

llvm/lib/Target/RISCV/RISCVInstrInfoD.td

@@ -78,7 +78,7 @@ def FSD : FPStore_r<0b011, "fsd", FPR64, WriteFST64>;
 } // Predicates = [HasStdExtD]
 
 foreach Ext = DExts in {
-  let SchedRW = [WriteFMA64, ReadFMA64, ReadFMA64, ReadFMA64] in {
+  let SchedRW = [WriteFMA64, ReadFMA64, ReadFMA64, ReadFMA64Addend] in {
     defm FMADD_D  : FPFMA_rrr_frm_m<OPC_MADD,  0b01, "fmadd.d",  Ext>;
     defm FMSUB_D  : FPFMA_rrr_frm_m<OPC_MSUB,  0b01, "fmsub.d",  Ext>;
     defm FNMSUB_D : FPFMA_rrr_frm_m<OPC_NMSUB, 0b01, "fnmsub.d", Ext>;

llvm/lib/Target/RISCV/RISCVInstrInfoF.td

@@ -302,7 +302,7 @@ def FSW : FPStore_r<0b010, "fsw", FPR32, WriteFST32>;
 } // Predicates = [HasStdExtF]
 
 foreach Ext = FExts in {
-  let SchedRW = [WriteFMA32, ReadFMA32, ReadFMA32, ReadFMA32] in {
+  let SchedRW = [WriteFMA32, ReadFMA32, ReadFMA32, ReadFMA32Addend] in {
     defm FMADD_S  : FPFMA_rrr_frm_m<OPC_MADD,  0b00, "fmadd.s",  Ext>;
     defm FMSUB_S  : FPFMA_rrr_frm_m<OPC_MSUB,  0b00, "fmsub.s",  Ext>;
     defm FNMSUB_S : FPFMA_rrr_frm_m<OPC_NMSUB, 0b00, "fnmsub.s", Ext>;

llvm/lib/Target/RISCV/RISCVInstrInfoZfh.td

@@ -85,7 +85,7 @@ def FSH : FPStore_r<0b001, "fsh", FPR16, WriteFST16>;
 } // Predicates = [HasHalfFPLoadStoreMove]
 
 foreach Ext = ZfhExts in {
-  let SchedRW = [WriteFMA16, ReadFMA16, ReadFMA16, ReadFMA16] in {
+  let SchedRW = [WriteFMA16, ReadFMA16, ReadFMA16, ReadFMA16Addend] in {
     defm FMADD_H  : FPFMA_rrr_frm_m<OPC_MADD,  0b10, "fmadd.h",  Ext>;
     defm FMSUB_H  : FPFMA_rrr_frm_m<OPC_MSUB,  0b10, "fmsub.h",  Ext>;
     defm FNMSUB_H : FPFMA_rrr_frm_m<OPC_NMSUB, 0b10, "fnmsub.h", Ext>;

llvm/lib/Target/RISCV/RISCVSchedRocket.td

@@ -206,7 +206,9 @@ def : ReadAdvance<ReadFAdd64, 0>;
 def : ReadAdvance<ReadFMul32, 0>;
 def : ReadAdvance<ReadFMul64, 0>;
 def : ReadAdvance<ReadFMA32, 0>;
+def : ReadAdvance<ReadFMA32Addend, 0>;
 def : ReadAdvance<ReadFMA64, 0>;
+def : ReadAdvance<ReadFMA64Addend, 0>;
 def : ReadAdvance<ReadFDiv32, 0>;
 def : ReadAdvance<ReadFDiv64, 0>;
 def : ReadAdvance<ReadFSqrt32, 0>;

llvm/lib/Target/RISCV/RISCVSchedSiFive7.td

@@ -933,10 +933,13 @@ def : ReadAdvance<ReadFAdd32, 0>;
 def : ReadAdvance<ReadFAdd64, 0>;
 def : ReadAdvance<ReadFMul16, 0>;
 def : ReadAdvance<ReadFMA16, 0>;
+def : ReadAdvance<ReadFMA16Addend, 0>;
 def : ReadAdvance<ReadFMul32, 0>;
 def : ReadAdvance<ReadFMul64, 0>;
 def : ReadAdvance<ReadFMA32, 0>;
+def : ReadAdvance<ReadFMA32Addend, 0>;
 def : ReadAdvance<ReadFMA64, 0>;
+def : ReadAdvance<ReadFMA64Addend, 0>;
 def : ReadAdvance<ReadFDiv16, 0>;
 def : ReadAdvance<ReadFDiv32, 0>;
 def : ReadAdvance<ReadFDiv64, 0>;

llvm/lib/Target/RISCV/RISCVSchedSyntacoreSCR1.td

@@ -164,7 +164,9 @@ def : ReadAdvance<ReadFAdd64, 0>;
 def : ReadAdvance<ReadFMul32, 0>;
 def : ReadAdvance<ReadFMul64, 0>;
 def : ReadAdvance<ReadFMA32, 0>;
+def : ReadAdvance<ReadFMA32Addend, 0>;
 def : ReadAdvance<ReadFMA64, 0>;
+def : ReadAdvance<ReadFMA64Addend, 0>;
 def : ReadAdvance<ReadFDiv32, 0>;
 def : ReadAdvance<ReadFDiv64, 0>;
 def : ReadAdvance<ReadFSqrt32, 0>;

llvm/lib/Target/RISCV/RISCVSchedule.td

@@ -150,8 +150,11 @@ def ReadFMul16      : SchedRead;    // 16-bit floating point multiply
 def ReadFMul32      : SchedRead;    // 32-bit floating point multiply
 def ReadFMul64      : SchedRead;    // 64-bit floating point multiply
 def ReadFMA16       : SchedRead;    // 16-bit floating point fused multiply-add
+def ReadFMA16Addend : SchedRead;    // 16-bit floating point fused multiply-add (addend)
 def ReadFMA32       : SchedRead;    // 32-bit floating point fused multiply-add
+def ReadFMA32Addend : SchedRead;    // 32-bit floating point fused multiply-add (addend)
 def ReadFMA64       : SchedRead;    // 64-bit floating point fused multiply-add
+def ReadFMA64Addend : SchedRead;    // 64-bit floating point fused multiply-add (addend)
 def ReadFDiv16      : SchedRead;    // 16-bit floating point divide
 def ReadFDiv32      : SchedRead;    // 32-bit floating point divide
 def ReadFDiv64      : SchedRead;    // 64-bit floating point divide