[RISCV][GISel] Fold G_FCONSTANT 0.0 store into G_CONSTANT x0

llvm/lib/Target/RISCV/GISel/RISCVPostLegalizerCombiner.cpp

@@ -27,6 +27,7 @@
 #include "llvm/CodeGen/MachineDominators.h"
 #include "llvm/CodeGen/MachineFunctionPass.h"
 #include "llvm/CodeGen/TargetPassConfig.h"
+#include "llvm/Support/FormatVariadic.h"
 
 #define GET_GICOMBINER_DEPS
 #include "RISCVGenPostLegalizeGICombiner.inc"
@@ -42,6 +43,50 @@ namespace {
 #include "RISCVGenPostLegalizeGICombiner.inc"
 #undef GET_GICOMBINER_TYPES
 
+/// Match: G_STORE (G_FCONSTANT +0.0), addr
+/// Return the source vreg in MatchInfo if matched.
+bool matchFoldFPZeroStore(MachineInstr &MI, MachineRegisterInfo &MRI,
+                          Register &MatchInfo) {
+  if (MI.getOpcode() != TargetOpcode::G_STORE)
+    return false;
+
+  Register SrcReg = MI.getOperand(0).getReg();
+  if (!SrcReg.isVirtual())
+    return false;
+
+  MachineInstr *Def = MRI.getVRegDef(SrcReg);
+  if (!Def || Def->getOpcode() != TargetOpcode::G_FCONSTANT)
+    return false;
+
+  auto *CFP = Def->getOperand(1).getFPImm();
+  if (!CFP || !CFP->getValueAPF().isPosZero())
+    return false;
+
+  MatchInfo = SrcReg;
+  return true;
+}
+
+/// Apply: rewrite to G_STORE (G_CONSTANT 0 [XLEN]), addr
+void applyFoldFPZeroStore(MachineInstr &MI, MachineRegisterInfo &MRI,
+                          MachineIRBuilder &B, const RISCVSubtarget &STI,
+                          Register &MatchInfo) {
+  const unsigned XLen = STI.getXLen();
+
+  auto Zero = B.buildConstant(LLT::scalar(XLen), 0);
+  MI.getOperand(0).setReg(Zero.getReg(0));
+
+  MachineInstr *Def = MRI.getVRegDef(MatchInfo);
+  if (Def && MRI.use_nodbg_empty(MatchInfo))
+    Def->eraseFromParent();
+
+#ifndef NDEBUG
+  unsigned ValBits = MRI.getType(MatchInfo).getSizeInBits();
+  LLVM_DEBUG(dbgs() << formatv("[{0}] Fold FP zero store -> int zero "
+                               "(XLEN={1}, ValBits={2}):\n  {3}\n",
+                               DEBUG_TYPE, XLen, ValBits, MI));
+#endif
+}
+
 class RISCVPostLegalizerCombinerImpl : public Combiner {
 protected:
   const CombinerHelper Helper;

llvm/lib/Target/RISCV/RISCVCombine.td

@@ -19,11 +19,20 @@ def RISCVO0PreLegalizerCombiner: GICombiner<
   "RISCVO0PreLegalizerCombinerImpl", [optnone_combines]> {
 }
 
+// Rule: fold store (fp +0.0) -> store (int zero [XLEN])
+def fp_zero_store_matchdata : GIDefMatchData<"Register">;
+def fold_fp_zero_store : GICombineRule<
+  (defs root:$root, fp_zero_store_matchdata:$matchinfo),
+  (match (G_STORE $src, $addr):$root,
+         [{ return matchFoldFPZeroStore(*${root}, MRI, ${matchinfo}); }]),
+  (apply [{ applyFoldFPZeroStore(*${root}, MRI, B, STI, ${matchinfo}); }])>;
+
 // Post-legalization combines which are primarily optimizations.
 // TODO: Add more combines.
 def RISCVPostLegalizerCombiner
     : GICombiner<"RISCVPostLegalizerCombinerImpl",
                  [sub_to_add, combines_for_extload, redundant_and,
                   identity_combines, shift_immed_chain,
-                  commute_constant_to_rhs, simplify_neg_minmax]> {
+                  commute_constant_to_rhs, simplify_neg_minmax,
+                  fold_fp_zero_store]> {
 }

llvm/test/CodeGen/RISCV/GlobalISel/store-fp-zero-to-x0.ll

@@ -0,0 +1,196 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -global-isel -mtriple=riscv32 -mattr=+f,+zfh < %s \
+; RUN:   | FileCheck %s --check-prefix=RV32
+; RUN: llc -global-isel -mtriple=riscv64 -mattr=+d,+zfh < %s \
+; RUN:   | FileCheck %s --check-prefix=RV64
+
+define void @zero_f16(ptr %i) {
+; RV32-LABEL: zero_f16:
+; RV32:       # %bb.0: # %entry
+; RV32-NEXT:    sh zero, 0(a0)
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: zero_f16:
+; RV64:       # %bb.0: # %entry
+; RV64-NEXT:    sh zero, 0(a0)
+; RV64-NEXT:    ret
+entry:
+  store half 0.0, ptr %i, align 4
+  ret void
+}
+
+define void @zero_bf16(ptr %i) {
+; RV32-LABEL: zero_bf16:
+; RV32:       # %bb.0: # %entry
+; RV32-NEXT:    sh zero, 0(a0)
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: zero_bf16:
+; RV64:       # %bb.0: # %entry
+; RV64-NEXT:    sh zero, 0(a0)
+; RV64-NEXT:    ret
+entry:
+  store bfloat 0.0, ptr %i, align 4
+  ret void
+}
+
+define void @zero_f32(ptr %i) {
+; RV32-LABEL: zero_f32:
+; RV32:       # %bb.0: # %entry
+; RV32-NEXT:    sw zero, 0(a0)
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: zero_f32:
+; RV64:       # %bb.0: # %entry
+; RV64-NEXT:    sw zero, 0(a0)
+; RV64-NEXT:    ret
+entry:
+  store float 0.0, ptr %i, align 4
+  ret void
+}
+
+
+define void @zero_f64(ptr %i) {
+; RV32-LABEL: zero_f64:
+; RV32:       # %bb.0: # %entry
+; RV32-NEXT:    lui a1, %hi(.LCPI3_0)
+; RV32-NEXT:    addi a1, a1, %lo(.LCPI3_0)
+; RV32-NEXT:    lw a2, 0(a1)
+; RV32-NEXT:    lw a1, 4(a1)
+; RV32-NEXT:    sw a2, 0(a0)
+; RV32-NEXT:    sw a1, 4(a0)
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: zero_f64:
+; RV64:       # %bb.0: # %entry
+; RV64-NEXT:    sd zero, 0(a0)
+; RV64-NEXT:    ret
+entry:
+  store double 0.0, ptr %i, align 8
+  ret void
+}
+
+define void @zero_v1f32(ptr %i) {
+; RV32-LABEL: zero_v1f32:
+; RV32:       # %bb.0: # %entry
+; RV32-NEXT:    sw zero, 0(a0)
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: zero_v1f32:
+; RV64:       # %bb.0: # %entry
+; RV64-NEXT:    sw zero, 0(a0)
+; RV64-NEXT:    ret
+entry:
+  store <1 x float> <float 0.0>, ptr %i, align 8
+  ret void
+}
+
+define void @zero_v2f32(ptr %i) {
+; RV32-LABEL: zero_v2f32:
+; RV32:       # %bb.0: # %entry
+; RV32-NEXT:    sw zero, 0(a0)
+; RV32-NEXT:    sw zero, 4(a0)
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: zero_v2f32:
+; RV64:       # %bb.0: # %entry
+; RV64-NEXT:    sw zero, 0(a0)
+; RV64-NEXT:    sw zero, 4(a0)
+; RV64-NEXT:    ret
+entry:
+  store <2 x float> <float 0.0, float 0.0>, ptr %i, align 8
+  ret void
+}
+
+define void @zero_v4f32(ptr %i) {
+; RV32-LABEL: zero_v4f32:
+; RV32:       # %bb.0: # %entry
+; RV32-NEXT:    sw zero, 0(a0)
+; RV32-NEXT:    sw zero, 4(a0)
+; RV32-NEXT:    sw zero, 8(a0)
+; RV32-NEXT:    sw zero, 12(a0)
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: zero_v4f32:
+; RV64:       # %bb.0: # %entry
+; RV64-NEXT:    sw zero, 0(a0)
+; RV64-NEXT:    sw zero, 4(a0)
+; RV64-NEXT:    sw zero, 8(a0)
+; RV64-NEXT:    sw zero, 12(a0)
+; RV64-NEXT:    ret
+entry:
+  store <4 x float> <float 0.0, float 0.0, float 0.0, float 0.0>, ptr %i, align 8
+  ret void
+}
+
+define void @zero_v1f64(ptr %i) {
+; RV32-LABEL: zero_v1f64:
+; RV32:       # %bb.0: # %entry
+; RV32-NEXT:    lui a1, %hi(.LCPI7_0)
+; RV32-NEXT:    addi a1, a1, %lo(.LCPI7_0)
+; RV32-NEXT:    lw a2, 0(a1)
+; RV32-NEXT:    lw a1, 4(a1)
+; RV32-NEXT:    sw a2, 0(a0)
+; RV32-NEXT:    sw a1, 4(a0)
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: zero_v1f64:
+; RV64:       # %bb.0: # %entry
+; RV64-NEXT:    sd zero, 0(a0)
+; RV64-NEXT:    ret
+entry:
+  store <1 x double> <double 0.0>, ptr %i, align 8
+  ret void
+}
+
+define void @zero_v2f64(ptr %i) {
+; RV32-LABEL: zero_v2f64:
+; RV32:       # %bb.0: # %entry
+; RV32-NEXT:    lui a1, %hi(.LCPI8_0)
+; RV32-NEXT:    addi a1, a1, %lo(.LCPI8_0)
+; RV32-NEXT:    lw a2, 0(a1)
+; RV32-NEXT:    lw a1, 4(a1)
+; RV32-NEXT:    sw a2, 0(a0)
+; RV32-NEXT:    sw a1, 4(a0)
+; RV32-NEXT:    sw a2, 8(a0)
+; RV32-NEXT:    sw a1, 12(a0)
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: zero_v2f64:
+; RV64:       # %bb.0: # %entry
+; RV64-NEXT:    sd zero, 0(a0)
+; RV64-NEXT:    sd zero, 8(a0)
+; RV64-NEXT:    ret
+entry:
+  store <2 x double> <double 0.0, double 0.0>, ptr %i, align 8
+  ret void
+}
+
+define void @zero_v4f64(ptr %i) {
+; RV32-LABEL: zero_v4f64:
+; RV32:       # %bb.0: # %entry
+; RV32-NEXT:    lui a1, %hi(.LCPI9_0)
+; RV32-NEXT:    addi a1, a1, %lo(.LCPI9_0)
+; RV32-NEXT:    lw a2, 0(a1)
+; RV32-NEXT:    lw a1, 4(a1)
+; RV32-NEXT:    sw a2, 0(a0)
+; RV32-NEXT:    sw a1, 4(a0)
+; RV32-NEXT:    sw a2, 8(a0)
+; RV32-NEXT:    sw a1, 12(a0)
+; RV32-NEXT:    sw a2, 16(a0)
+; RV32-NEXT:    sw a1, 20(a0)
+; RV32-NEXT:    sw a2, 24(a0)
+; RV32-NEXT:    sw a1, 28(a0)
+; RV32-NEXT:    ret
+;
+; RV64-LABEL: zero_v4f64:
+; RV64:       # %bb.0: # %entry
+; RV64-NEXT:    sd zero, 0(a0)
+; RV64-NEXT:    sd zero, 8(a0)
+; RV64-NEXT:    sd zero, 16(a0)
+; RV64-NEXT:    sd zero, 24(a0)
+; RV64-NEXT:    ret
+entry:
+  store <4 x double> <double 0.0, double 0.0, double 0.0, double 0.0>, ptr %i, align 8
+  ret void
+}