[RISCV] Generate more W instructons

We rename `TuneNoStripWSuffix` to `TunePreferWInst`.

If all the users of an instruction just use the low 32 bits, we can
convert it to its W variant.

A quick test on Coremark (`-O3 -march=rv64gc`):

|        | W instructions | code size(.text) |
|--------|----------------|------------------|
| before | 302            | 12257            |
| after  | 343            | 12265            |
|        | +13.58%        | +0.065%          |

Reviewers: asb, dtcxzyw, preames, lukel97, michaelmaitland, topperc

Reviewed By: topperc, dtcxzyw

Pull Request: #87237

Author: wangpc-pp

llvm/lib/Target/RISCV/RISCVFeatures.td

@@ -1226,9 +1226,9 @@ def TuneNoSinkSplatOperands
                        "false", "Disable sink splat operands to enable .vx, .vf,"
                        ".wx, and .wf instructions">;
 
-def TuneNoStripWSuffix
-    : SubtargetFeature<"no-strip-w-suffix", "EnableStripWSuffix", "false",
-                       "Disable strip W suffix">;
+def TunePreferWInst
+    : SubtargetFeature<"prefer-w-inst", "PreferWInst", "true",
+                       "Prefer instructions with W suffix">;
 
 def TuneConditionalCompressedMoveFusion
     : SubtargetFeature<"conditional-cmv-fusion", "HasConditionalCompressedMoveFusion",

llvm/lib/Target/RISCV/RISCVOptWInstrs.cpp

@@ -12,15 +12,24 @@
 // extended bits aren't consumed or because the input was already sign extended
 // by an earlier instruction.
 //
-// Then it removes the -w suffix from opw instructions whenever all users are
-// dependent only on the lower word of the result of the instruction.
-// The cases handled are:
-// * addw because c.add has a larger register encoding than c.addw.
-// * addiw because it helps reduce test differences between RV32 and RV64
-//   w/o being a pessimization.
-// * mulw because c.mulw doesn't exist but c.mul does (w/ zcb)
-// * slliw because c.slliw doesn't exist and c.slli does
+// Then:
+// 1. Unless explicit disabled or the target prefers instructions with W suffix,
+//    it removes the -w suffix from opw instructions whenever all users are
+//    dependent only on the lower word of the result of the instruction.
+//    The cases handled are:
+//    * addw because c.add has a larger register encoding than c.addw.
+//    * addiw because it helps reduce test differences between RV32 and RV64
+//      w/o being a pessimization.
+//    * mulw because c.mulw doesn't exist but c.mul does (w/ zcb)
+//    * slliw because c.slliw doesn't exist and c.slli does
 //
+// 2. Or if explicit enabled or the target prefers instructions with W suffix,
+//    it adds the W suffix to the instruction whenever all users are dependent
+//    only on the lower word of the result of the instruction.
+//    The cases handled are:
+//    * add/addi/sub/mul.
+//    * slli with imm < 32.
+//    * ld/lwu.
 //===---------------------------------------------------------------------===//
 
 #include "RISCV.h"
@@ -60,6 +69,8 @@ class RISCVOptWInstrs : public MachineFunctionPass {
                          const RISCVSubtarget &ST, MachineRegisterInfo &MRI);
   bool stripWSuffixes(MachineFunction &MF, const RISCVInstrInfo &TII,
                       const RISCVSubtarget &ST, MachineRegisterInfo &MRI);
+  bool appendWSuffixes(MachineFunction &MF, const RISCVInstrInfo &TII,
+                       const RISCVSubtarget &ST, MachineRegisterInfo &MRI);
 
   void getAnalysisUsage(AnalysisUsage &AU) const override {
     AU.setPreservesCFG();
@@ -672,9 +683,6 @@ bool RISCVOptWInstrs::stripWSuffixes(MachineFunction &MF,
                                      const RISCVInstrInfo &TII,
                                      const RISCVSubtarget &ST,
                                      MachineRegisterInfo &MRI) {
-  if (DisableStripWSuffix || !ST.enableStripWSuffix())
-    return false;
-
   bool MadeChange = false;
   for (MachineBasicBlock &MBB : MF) {
     for (MachineInstr &MI : MBB) {
@@ -698,6 +706,58 @@ bool RISCVOptWInstrs::stripWSuffixes(MachineFunction &MF,
   return MadeChange;
 }
 
+bool RISCVOptWInstrs::appendWSuffixes(MachineFunction &MF,
+                                      const RISCVInstrInfo &TII,
+                                      const RISCVSubtarget &ST,
+                                      MachineRegisterInfo &MRI) {
+  bool MadeChange = false;
+  for (MachineBasicBlock &MBB : MF) {
+    for (MachineInstr &MI : MBB) {
+      unsigned WOpc;
+      // TODO: Add more?
+      switch (MI.getOpcode()) {
+      default:
+        continue;
+      case RISCV::ADD:
+        WOpc = RISCV::ADDW;
+        break;
+      case RISCV::ADDI:
+        WOpc = RISCV::ADDIW;
+        break;
+      case RISCV::SUB:
+        WOpc = RISCV::SUBW;
+        break;
+      case RISCV::MUL:
+        WOpc = RISCV::MULW;
+        break;
+      case RISCV::SLLI:
+        // SLLIW reads the lowest 5 bits, while SLLI reads lowest 6 bits
+        if (MI.getOperand(2).getImm() >= 32)
+          continue;
+        WOpc = RISCV::SLLIW;
+        break;
+      case RISCV::LD:
+      case RISCV::LWU:
+        WOpc = RISCV::LW;
+        break;
+      }
+
+      if (hasAllWUsers(MI, ST, MRI)) {
+        LLVM_DEBUG(dbgs() << "Replacing " << MI);
+        MI.setDesc(TII.get(WOpc));
+        MI.clearFlag(MachineInstr::MIFlag::NoSWrap);
+        MI.clearFlag(MachineInstr::MIFlag::NoUWrap);
+        MI.clearFlag(MachineInstr::MIFlag::IsExact);
+        LLVM_DEBUG(dbgs() << "     with " << MI);
+        ++NumTransformedToWInstrs;
+        MadeChange = true;
+      }
+    }
+  }
+
+  return MadeChange;
+}
+
 bool RISCVOptWInstrs::runOnMachineFunction(MachineFunction &MF) {
   if (skipFunction(MF.getFunction()))
     return false;
@@ -711,7 +771,12 @@ bool RISCVOptWInstrs::runOnMachineFunction(MachineFunction &MF) {
 
   bool MadeChange = false;
   MadeChange |= removeSExtWInstrs(MF, TII, ST, MRI);
-  MadeChange |= stripWSuffixes(MF, TII, ST, MRI);
+
+  if (!(DisableStripWSuffix || ST.preferWInst()))
+    MadeChange |= stripWSuffixes(MF, TII, ST, MRI);
+
+  if (ST.preferWInst())
+    MadeChange |= appendWSuffixes(MF, TII, ST, MRI);
 
   return MadeChange;
 }

llvm/test/CodeGen/RISCV/prefer-w-inst.ll

@@ -0,0 +1,105 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=riscv64 -mattr=+m -verify-machineinstrs < %s \
+; RUN:   | FileCheck -check-prefixes=NO-PREFER-W-INST %s
+; RUN: llc -mtriple=riscv64 -mattr=+m -riscv-disable-strip-w-suffix -verify-machineinstrs < %s \
+; RUN:   | FileCheck -check-prefixes=NO-STRIP %s
+; RUN: llc -mtriple=riscv64 -mattr=+m,+prefer-w-inst -verify-machineinstrs < %s \
+; RUN:   | FileCheck -check-prefixes=PREFER-W-INST %s
+
+define i32 @addiw(i32 %a) {
+; NO-PREFER-W-INST-LABEL: addiw:
+; NO-PREFER-W-INST:       # %bb.0:
+; NO-PREFER-W-INST-NEXT:    lui a1, 1
+; NO-PREFER-W-INST-NEXT:    addi a1, a1, -1
+; NO-PREFER-W-INST-NEXT:    addw a0, a0, a1
+; NO-PREFER-W-INST-NEXT:    ret
+;
+; NO-STRIP-LABEL: addiw:
+; NO-STRIP:       # %bb.0:
+; NO-STRIP-NEXT:    lui a1, 1
+; NO-STRIP-NEXT:    addiw a1, a1, -1
+; NO-STRIP-NEXT:    addw a0, a0, a1
+; NO-STRIP-NEXT:    ret
+;
+; PREFER-W-INST-LABEL: addiw:
+; PREFER-W-INST:       # %bb.0:
+; PREFER-W-INST-NEXT:    lui a1, 1
+; PREFER-W-INST-NEXT:    addiw a1, a1, -1
+; PREFER-W-INST-NEXT:    addw a0, a0, a1
+; PREFER-W-INST-NEXT:    ret
+  %ret = add i32 %a, 4095
+  ret i32 %ret
+}
+
+define i32 @addw(i32 %a, i32 %b) {
+; NO-PREFER-W-INST-LABEL: addw:
+; NO-PREFER-W-INST:       # %bb.0:
+; NO-PREFER-W-INST-NEXT:    add a0, a0, a1
+; NO-PREFER-W-INST-NEXT:    addiw a0, a0, 1024
+; NO-PREFER-W-INST-NEXT:    ret
+;
+; NO-STRIP-LABEL: addw:
+; NO-STRIP:       # %bb.0:
+; NO-STRIP-NEXT:    addw a0, a0, a1
+; NO-STRIP-NEXT:    addiw a0, a0, 1024
+; NO-STRIP-NEXT:    ret
+;
+; PREFER-W-INST-LABEL: addw:
+; PREFER-W-INST:       # %bb.0:
+; PREFER-W-INST-NEXT:    addw a0, a0, a1
+; PREFER-W-INST-NEXT:    addiw a0, a0, 1024
+; PREFER-W-INST-NEXT:    ret
+  %add = add i32 %a, %b
+  %ret = add i32 %add, 1024
+  ret i32 %ret
+}
+
+define i32 @mulw(i32 %a, i32 %b) {
+; NO-PREFER-W-INST-LABEL: mulw:
+; NO-PREFER-W-INST:       # %bb.0:
+; NO-PREFER-W-INST-NEXT:    mul a1, a0, a1
+; NO-PREFER-W-INST-NEXT:    mul a0, a0, a1
+; NO-PREFER-W-INST-NEXT:    addiw a0, a0, 1024
+; NO-PREFER-W-INST-NEXT:    ret
+;
+; NO-STRIP-LABEL: mulw:
+; NO-STRIP:       # %bb.0:
+; NO-STRIP-NEXT:    mulw a1, a0, a1
+; NO-STRIP-NEXT:    mulw a0, a0, a1
+; NO-STRIP-NEXT:    addiw a0, a0, 1024
+; NO-STRIP-NEXT:    ret
+;
+; PREFER-W-INST-LABEL: mulw:
+; PREFER-W-INST:       # %bb.0:
+; PREFER-W-INST-NEXT:    mulw a1, a0, a1
+; PREFER-W-INST-NEXT:    mulw a0, a0, a1
+; PREFER-W-INST-NEXT:    addiw a0, a0, 1024
+; PREFER-W-INST-NEXT:    ret
+  %mul1 = mul i32 %a, %b
+  %mul = mul i32 %a, %mul1
+  %ret = add i32 %mul, 1024
+  ret i32 %ret
+}
+
+define i32 @slliw(i32 %a) {
+; NO-PREFER-W-INST-LABEL: slliw:
+; NO-PREFER-W-INST:       # %bb.0:
+; NO-PREFER-W-INST-NEXT:    slli a0, a0, 1
+; NO-PREFER-W-INST-NEXT:    addiw a0, a0, 1024
+; NO-PREFER-W-INST-NEXT:    ret
+;
+; NO-STRIP-LABEL: slliw:
+; NO-STRIP:       # %bb.0:
+; NO-STRIP-NEXT:    slliw a0, a0, 1
+; NO-STRIP-NEXT:    addiw a0, a0, 1024
+; NO-STRIP-NEXT:    ret
+;
+; PREFER-W-INST-LABEL: slliw:
+; PREFER-W-INST:       # %bb.0:
+; PREFER-W-INST-NEXT:    slliw a0, a0, 1
+; PREFER-W-INST-NEXT:    addiw a0, a0, 1024
+; PREFER-W-INST-NEXT:    ret
+  %shl = shl i32 %a, 1
+  %ret = add i32 %shl, 1024
+  ret i32 %ret
+}