[RISCV] Add optimization for memset inline

Author: BoyaoWang430

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

@@ -1664,7 +1664,11 @@ RISCVTargetLowering::RISCVTargetLowering(const TargetMachine &TM,
   PredictableSelectIsExpensive = Subtarget.predictableSelectIsExpensive();
 
   MaxStoresPerMemsetOptSize = Subtarget.getMaxStoresPerMemset(/*OptSize=*/true);
-  MaxStoresPerMemset = Subtarget.getMaxStoresPerMemset(/*OptSize=*/false);
+  MaxStoresPerMemset = Subtarget.hasVInstructions()
+                           ? (Subtarget.getRealMinVLen() / 8 *
+                              Subtarget.getMaxLMULForFixedLengthVectors() /
+                              (Subtarget.is64Bit() ? 8 : 4))
+                           : Subtarget.getMaxStoresPerMemset(/*OptSize=*/false);
 
   MaxGluedStoresPerMemcpy = Subtarget.getMaxGluedStoresPerMemcpy();
   MaxStoresPerMemcpyOptSize = Subtarget.getMaxStoresPerMemcpy(/*OptSize=*/true);
@@ -23804,6 +23808,20 @@ EVT RISCVTargetLowering::getOptimalMemOpType(
   if (MinVLenInBytes <= RISCV::RVVBytesPerBlock)
     return MVT::Other;
 
+  // If Op size is greater than LMUL8 memory operation, we don't support inline
+  // of memset. Return EVT based on Op size to avoid redundant splitting and
+  // merging operations if Op size is no greater than LMUL8 memory operation.
+  if (Op.isMemset()) {
+    if (Op.size() >
+        Subtarget.getMaxLMULForFixedLengthVectors() * MinVLenInBytes)
+      return MVT::Other;
+    if (Subtarget.hasVInstructionsI64() && Op.size() % 8 == 0)
+      return EVT::getVectorVT(Context, MVT::i64, Op.size() / 8);
+    if (Op.size() % 4 == 0)
+      return EVT::getVectorVT(Context, MVT::i32, Op.size() / 4);
+    return EVT::getVectorVT(Context, MVT::i8, Op.size());
+  }
+
   // Prefer i8 for non-zero memset as it allows us to avoid materializing
   // a large scalar constant and instead use vmv.v.x/i to do the
   // broadcast.  For everything else, prefer ELenVT to minimize VL and thus

llvm/test/CodeGen/RISCV/pr135206.ll

@@ -12,9 +12,6 @@ define i1 @foo() nounwind "probe-stack"="inline-asm" "target-features"="+v" {
 ; CHECK-NEXT:    addi sp, sp, -2032
 ; CHECK-NEXT:    sd ra, 2024(sp) # 8-byte Folded Spill
 ; CHECK-NEXT:    sd s0, 2016(sp) # 8-byte Folded Spill
-; CHECK-NEXT:    sd s1, 2008(sp) # 8-byte Folded Spill
-; CHECK-NEXT:    sd s2, 2000(sp) # 8-byte Folded Spill
-; CHECK-NEXT:    sd s3, 1992(sp) # 8-byte Folded Spill
 ; CHECK-NEXT:    lui a0, 7
 ; CHECK-NEXT:    sub t1, sp, a0
 ; CHECK-NEXT:    lui t2, 1
@@ -24,8 +21,9 @@ define i1 @foo() nounwind "probe-stack"="inline-asm" "target-features"="+v" {
 ; CHECK-NEXT:    bne sp, t1, .LBB0_1
 ; CHECK-NEXT:  # %bb.2:
 ; CHECK-NEXT:    addi sp, sp, -2048
-; CHECK-NEXT:    addi sp, sp, -96
+; CHECK-NEXT:    addi sp, sp, -48
 ; CHECK-NEXT:    csrr t1, vlenb
+; CHECK-NEXT:    slli t1, t1, 2
 ; CHECK-NEXT:    lui t2, 1
 ; CHECK-NEXT:  .LBB0_3: # =>This Inner Loop Header: Depth=1
 ; CHECK-NEXT:    sub sp, sp, t2
@@ -34,45 +32,34 @@ define i1 @foo() nounwind "probe-stack"="inline-asm" "target-features"="+v" {
 ; CHECK-NEXT:    bge t1, t2, .LBB0_3
 ; CHECK-NEXT:  # %bb.4:
 ; CHECK-NEXT:    sub sp, sp, t1
-; CHECK-NEXT:    li a0, 86
-; CHECK-NEXT:    addi s0, sp, 48
-; CHECK-NEXT:    addi s1, sp, 32
-; CHECK-NEXT:    addi s2, sp, 16
-; CHECK-NEXT:    lui a1, 353637
-; CHECK-NEXT:    vsetivli zero, 16, e8, m1, ta, ma
+; CHECK-NEXT:    lui a0, 353637
+; CHECK-NEXT:    addi a0, a0, 1622
+; CHECK-NEXT:    slli a1, a0, 32
+; CHECK-NEXT:    add a0, a0, a1
+; CHECK-NEXT:    vsetivli zero, 7, e64, m4, ta, ma
 ; CHECK-NEXT:    vmv.v.x v8, a0
 ; CHECK-NEXT:    lui a0, 8
-; CHECK-NEXT:    addi a0, a0, 32
+; CHECK-NEXT:    addi a0, a0, 16
 ; CHECK-NEXT:    add a0, sp, a0
-; CHECK-NEXT:    vs1r.v v8, (a0) # vscale x 8-byte Folded Spill
-; CHECK-NEXT:    addi a0, a1, 1622
-; CHECK-NEXT:    vse8.v v8, (s0)
-; CHECK-NEXT:    vse8.v v8, (s1)
-; CHECK-NEXT:    vse8.v v8, (s2)
-; CHECK-NEXT:    slli a1, a0, 32
-; CHECK-NEXT:    add s3, a0, a1
-; CHECK-NEXT:    sd s3, 64(sp)
+; CHECK-NEXT:    vs4r.v v8, (a0) # vscale x 32-byte Folded Spill
+; CHECK-NEXT:    addi s0, sp, 16
+; CHECK-NEXT:    vse64.v v8, (s0)
 ; CHECK-NEXT:    call bar
 ; CHECK-NEXT:    lui a0, 8
-; CHECK-NEXT:    addi a0, a0, 32
+; CHECK-NEXT:    addi a0, a0, 16
 ; CHECK-NEXT:    add a0, sp, a0
-; CHECK-NEXT:    vl1r.v v8, (a0) # vscale x 8-byte Folded Reload
-; CHECK-NEXT:    vsetivli zero, 16, e8, m1, ta, ma
-; CHECK-NEXT:    vse8.v v8, (s0)
-; CHECK-NEXT:    vse8.v v8, (s1)
-; CHECK-NEXT:    vse8.v v8, (s2)
-; CHECK-NEXT:    sd s3, 64(sp)
+; CHECK-NEXT:    vl4r.v v8, (a0) # vscale x 32-byte Folded Reload
+; CHECK-NEXT:    vsetivli zero, 7, e64, m4, ta, ma
+; CHECK-NEXT:    vse64.v v8, (s0)
 ; CHECK-NEXT:    li a0, 0
 ; CHECK-NEXT:    csrr a1, vlenb
+; CHECK-NEXT:    slli a1, a1, 2
 ; CHECK-NEXT:    add sp, sp, a1
 ; CHECK-NEXT:    lui a1, 8
-; CHECK-NEXT:    addi a1, a1, -1952
+; CHECK-NEXT:    addi a1, a1, -2000
 ; CHECK-NEXT:    add sp, sp, a1
 ; CHECK-NEXT:    ld ra, 2024(sp) # 8-byte Folded Reload
 ; CHECK-NEXT:    ld s0, 2016(sp) # 8-byte Folded Reload
-; CHECK-NEXT:    ld s1, 2008(sp) # 8-byte Folded Reload
-; CHECK-NEXT:    ld s2, 2000(sp) # 8-byte Folded Reload
-; CHECK-NEXT:    ld s3, 1992(sp) # 8-byte Folded Reload
 ; CHECK-NEXT:    addi sp, sp, 2032
 ; CHECK-NEXT:    ret
   %1 = alloca %"buff", align 8