[RISCV] Implement isMultiStoresCheaperThanBitsMerge hook

[asb]

Summary

Grabs the same logic and reasoning from the X86 implementation of the hook. The benefit is slightly less clear for when the soft float ABI is used (i.e. there's no transfer from an FPR to a GPR), but I've opted not to gate it based on ABI - feedback welcome.

I've got no reason to believe this is a particularly meaningful optimisation (i.e. I spotted the hook and figured the X86 logic applied to RISC-V rather than seeing this issue in real code), but it still seems marginally better for these test cases.

On the off chance you want to apply locally and experiment, the base test file is in D140409 (I've not committed directly, in case people aren't keen on adding this hook).

[asb]

Rebase and ping.

[asb]

Ping.

[reames]

LGTM

This looks entirely reasonable. I don't think this is worth gating on ABI. If we did gate it, we'd probably do so on the conversion cost - which would in turn indirectly dependent on ABI. We can revisit that in the future if we ever care about the non-F non-D configuration performance in this specific case.

llvm/lib/Target/RISCV/RISCVISelLowering.h

public:
   bool shouldFoldSelectWithIdentityConstant(unsigned BinOpcode,
                                             EVT VT) const override;
 
   /// Return true if the given shuffle mask can be codegen'd directly, or if it
   /// should be stack expanded.
   bool isShuffleMaskLegal(ArrayRef<int> M, EVT VT) const override;
 
   bool hasBitPreservingFPLogic(EVT VT) const override;
+  bool isMultiStoresCheaperThanBitsMerge(EVT LTy, EVT HTy) const override {
+    // If the pair to store is a mixture of float and int values, we will
+    // save two bitwise instructions and one float-to-int instruction and
+    // increase one store instruction. There is potentially a more
+    // significant benefit because it avoids the float->int domain switch
+    // for input value. So It is more likely a win.
+    if ((LTy.isFloatingPoint() && HTy.isInteger()) ||
+        (LTy.isInteger() && HTy.isFloatingPoint()))
+      return true;
+    // If the pair only contains int values, we will save two bitwise
+    // instructions and increase one store instruction (costing one more
+    // store buffer). Since the benefit is more blurred so we leave
+    // such pair out until we get testcase to prove it is a win.
+    return false;
+  }
   bool
   shouldExpandBuildVectorWithShuffles(EVT VT,
                                       unsigned DefinedValues) const override;
 
   // Provide custom lowering hooks for some operations.
   SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
   void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
                           SelectionDAG &DAG) const override;

llvm/test/CodeGen/RISCV/split-store.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -mtriple=riscv32 -verify-machineinstrs < %s \
 ; RUN:   | FileCheck %s -check-prefixes=RV32-RV64,RV32
 ; RUN: llc -mtriple=riscv32 -mattr=+d -verify-machineinstrs < %s \
 ; RUN:   | FileCheck %s -check-prefixes=RV32D-RV64D,RV32D
 ; RUN: llc -mtriple=riscv64 -verify-machineinstrs < %s \
 ; RUN:   | FileCheck %s -check-prefixes=RV32-RV64,RV64
 ; RUN: llc -mtriple=riscv64 -mattr=+d -verify-machineinstrs < %s \
 ; RUN:   | FileCheck %s -check-prefixes=RV32D-RV64D,RV64D
 
 define void @int32_float_pair(i32 %tmp1, float %tmp2, ptr %ref.tmp) {
-; RV32-LABEL: int32_float_pair:
-; RV32:       # %bb.0:
-; RV32-NEXT:    sw a1, 4(a2)
-; RV32-NEXT:    sw a0, 0(a2)
-; RV32-NEXT:    ret
-;
-; RV32D-LABEL: int32_float_pair:
-; RV32D:       # %bb.0:
-; RV32D-NEXT:    fsw fa0, 4(a1)
-; RV32D-NEXT:    sw a0, 0(a1)
-; RV32D-NEXT:    ret
-;
-; RV64-LABEL: int32_float_pair:
-; RV64:       # %bb.0:
-; RV64-NEXT:    slli a1, a1, 32
-; RV64-NEXT:    slli a0, a0, 32
-; RV64-NEXT:    srli a0, a0, 32
-; RV64-NEXT:    or a0, a1, a0
-; RV64-NEXT:    sd a0, 0(a2)
-; RV64-NEXT:    ret
-;
-; RV64D-LABEL: int32_float_pair:
-; RV64D:       # %bb.0:
-; RV64D-NEXT:    fmv.x.w a2, fa0
-; RV64D-NEXT:    slli a2, a2, 32
-; RV64D-NEXT:    slli a0, a0, 32
-; RV64D-NEXT:    srli a0, a0, 32
-; RV64D-NEXT:    or a0, a2, a0
-; RV64D-NEXT:    sd a0, 0(a1)
-; RV64D-NEXT:    ret
+; RV32-RV64-LABEL: int32_float_pair:
+; RV32-RV64:       # %bb.0:
+; RV32-RV64-NEXT:    sw a0, 0(a2)
+; RV32-RV64-NEXT:    sw a1, 4(a2)
+; RV32-RV64-NEXT:    ret
+;
+; RV32D-RV64D-LABEL: int32_float_pair:
+; RV32D-RV64D:       # %bb.0:
+; RV32D-RV64D-NEXT:    sw a0, 0(a1)
+; RV32D-RV64D-NEXT:    fsw fa0, 4(a1)
+; RV32D-RV64D-NEXT:    ret
   %t0 = bitcast float %tmp2 to i32
   %t1 = zext i32 %t0 to i64
   %t2 = shl nuw i64 %t1, 32
   %t3 = zext i32 %tmp1 to i64
   %t4 = or i64 %t2, %t3
   store i64 %t4, ptr %ref.tmp, align 8
   ret void
 }
 
 define void @float_int32_pair(float %tmp1, i32 %tmp2, ptr %ref.tmp) {
-; RV32-LABEL: float_int32_pair:
-; RV32:       # %bb.0:
-; RV32-NEXT:    sw a1, 4(a2)
-; RV32-NEXT:    sw a0, 0(a2)
-; RV32-NEXT:    ret
-;
-; RV32D-LABEL: float_int32_pair:
-; RV32D:       # %bb.0:
-; RV32D-NEXT:    sw a0, 4(a1)
-; RV32D-NEXT:    fsw fa0, 0(a1)
-; RV32D-NEXT:    ret
-;
-; RV64-LABEL: float_int32_pair:
-; RV64:       # %bb.0:
-; RV64-NEXT:    slli a1, a1, 32
-; RV64-NEXT:    slli a0, a0, 32
-; RV64-NEXT:    srli a0, a0, 32
-; RV64-NEXT:    or a0, a1, a0
-; RV64-NEXT:    sd a0, 0(a2)
-; RV64-NEXT:    ret
-;
-; RV64D-LABEL: float_int32_pair:
-; RV64D:       # %bb.0:
-; RV64D-NEXT:    fmv.x.w a2, fa0
-; RV64D-NEXT:    slli a0, a0, 32
-; RV64D-NEXT:    slli a2, a2, 32
-; RV64D-NEXT:    srli a2, a2, 32
-; RV64D-NEXT:    or a0, a0, a2
-; RV64D-NEXT:    sd a0, 0(a1)
-; RV64D-NEXT:    ret
+; RV32-RV64-LABEL: float_int32_pair:
+; RV32-RV64:       # %bb.0:
+; RV32-RV64-NEXT:    sw a0, 0(a2)
+; RV32-RV64-NEXT:    sw a1, 4(a2)
+; RV32-RV64-NEXT:    ret
+;
+; RV32D-RV64D-LABEL: float_int32_pair:
+; RV32D-RV64D:       # %bb.0:
+; RV32D-RV64D-NEXT:    fsw fa0, 0(a1)
+; RV32D-RV64D-NEXT:    sw a0, 4(a1)
+; RV32D-RV64D-NEXT:    ret
   %t0 = bitcast float %tmp1 to i32
   %t1 = zext i32 %tmp2 to i64
   %t2 = shl nuw i64 %t1, 32
   %t3 = zext i32 %t0 to i64
   %t4 = or i64 %t2, %t3
   store i64 %t4, ptr %ref.tmp, align 8
   ret void
 }
 
 define void @int16_float_pair(i16 signext %tmp1, float %tmp2, ptr %ref.tmp) {
 ; RV32-LABEL: int16_float_pair:
 ; RV32:       # %bb.0:
 ; RV32-NEXT:    slli a0, a0, 16
 ; RV32-NEXT:    srli a0, a0, 16
-; RV32-NEXT:    sw a1, 4(a2)
 ; RV32-NEXT:    sw a0, 0(a2)
+; RV32-NEXT:    sw a1, 4(a2)
 ; RV32-NEXT:    ret
 ;
 ; RV32D-LABEL: int16_float_pair:
 ; RV32D:       # %bb.0:
 ; RV32D-NEXT:    slli a0, a0, 16
 ; RV32D-NEXT:    srli a0, a0, 16
-; RV32D-NEXT:    fsw fa0, 4(a1)
 ; RV32D-NEXT:    sw a0, 0(a1)
+; RV32D-NEXT:    fsw fa0, 4(a1)
 ; RV32D-NEXT:    ret
 ;
 ; RV64-LABEL: int16_float_pair:
 ; RV64:       # %bb.0:
-; RV64-NEXT:    slli a1, a1, 32
 ; RV64-NEXT:    slli a0, a0, 48
 ; RV64-NEXT:    srli a0, a0, 48
-; RV64-NEXT:    or a0, a1, a0
-; RV64-NEXT:    sd a0, 0(a2)
+; RV64-NEXT:    sw a0, 0(a2)
+; RV64-NEXT:    sw a1, 4(a2)
 ; RV64-NEXT:    ret
 ;
 ; RV64D-LABEL: int16_float_pair:
 ; RV64D:       # %bb.0:
-; RV64D-NEXT:    fmv.x.w a2, fa0
-; RV64D-NEXT:    slli a2, a2, 32
 ; RV64D-NEXT:    slli a0, a0, 48
 ; RV64D-NEXT:    srli a0, a0, 48
-; RV64D-NEXT:    or a0, a2, a0
-; RV64D-NEXT:    sd a0, 0(a1)
+; RV64D-NEXT:    sw a0, 0(a1)
+; RV64D-NEXT:    fsw fa0, 4(a1)
 ; RV64D-NEXT:    ret
   %t0 = bitcast float %tmp2 to i32
   %t1 = zext i32 %t0 to i64
   %t2 = shl nuw i64 %t1, 32
   %t3 = zext i16 %tmp1 to i64
   %t4 = or i64 %t2, %t3
   store i64 %t4, ptr %ref.tmp, align 8
   ret void
 }
 
 define void @int8_float_pair(i8 signext %tmp1, float %tmp2, ptr %ref.tmp) {
-; RV32-LABEL: int8_float_pair:
-; RV32:       # %bb.0:
-; RV32-NEXT:    andi a0, a0, 255
-; RV32-NEXT:    sw a1, 4(a2)
-; RV32-NEXT:    sw a0, 0(a2)
-; RV32-NEXT:    ret
-;
-; RV32D-LABEL: int8_float_pair:
-; RV32D:       # %bb.0:
-; RV32D-NEXT:    andi a0, a0, 255
-; RV32D-NEXT:    fsw fa0, 4(a1)
-; RV32D-NEXT:    sw a0, 0(a1)
-; RV32D-NEXT:    ret
-;
-; RV64-LABEL: int8_float_pair:
-; RV64:       # %bb.0:
-; RV64-NEXT:    slli a1, a1, 32
-; RV64-NEXT:    andi a0, a0, 255
-; RV64-NEXT:    or a0, a1, a0
-; RV64-NEXT:    sd a0, 0(a2)
-; RV64-NEXT:    ret
-;
-; RV64D-LABEL: int8_float_pair:
-; RV64D:       # %bb.0:
-; RV64D-NEXT:    fmv.x.w a2, fa0
-; RV64D-NEXT:    slli a2, a2, 32
-; RV64D-NEXT:    andi a0, a0, 255
-; RV64D-NEXT:    or a0, a2, a0
-; RV64D-NEXT:    sd a0, 0(a1)
-; RV64D-NEXT:    ret
+; RV32-RV64-LABEL: int8_float_pair:
+; RV32-RV64:       # %bb.0:
+; RV32-RV64-NEXT:    andi a0, a0, 255
+; RV32-RV64-NEXT:    sw a0, 0(a2)
+; RV32-RV64-NEXT:    sw a1, 4(a2)
+; RV32-RV64-NEXT:    ret
+;
+; RV32D-RV64D-LABEL: int8_float_pair:
+; RV32D-RV64D:       # %bb.0:
+; RV32D-RV64D-NEXT:    andi a0, a0, 255
+; RV32D-RV64D-NEXT:    sw a0, 0(a1)
+; RV32D-RV64D-NEXT:    fsw fa0, 4(a1)
+; RV32D-RV64D-NEXT:    ret
   %t0 = bitcast float %tmp2 to i32
   %t1 = zext i32 %t0 to i64
   %t2 = shl nuw i64 %t1, 32
   %t3 = zext i8 %tmp1 to i64
   %t4 = or i64 %t2, %t3
   store i64 %t4, ptr %ref.tmp, align 8
   ret void
 }
; RV64D-NEXT:    ret
   %t2 = shl nuw i64 %t1, 32
   %t3 = zext i32 %tmp1 to i64
   %t4 = or i64 %t2, %t3
   store i64 %t4, ptr %ref.tmp, align 8
   ret void
 }
 
 define void @mbb_int32_float_pair(i32 %tmp1, float %tmp2, ptr %ref.tmp) {
-; RV32-LABEL: mbb_int32_float_pair:
-; RV32:       # %bb.0: # %entry
-; RV32-NEXT:    sw a1, 4(a2)
-; RV32-NEXT:    sw a0, 0(a2)
-; RV32-NEXT:    ret
-;
-; RV32D-LABEL: mbb_int32_float_pair:
-; RV32D:       # %bb.0: # %entry
-; RV32D-NEXT:    fmv.x.w a2, fa0
-; RV32D-NEXT:    sw a2, 4(a1)
-; RV32D-NEXT:    sw a0, 0(a1)
-; RV32D-NEXT:    ret
-;
-; RV64-LABEL: mbb_int32_float_pair:
-; RV64:       # %bb.0: # %entry
-; RV64-NEXT:    slli a1, a1, 32
-; RV64-NEXT:    slli a0, a0, 32
-; RV64-NEXT:    srli a0, a0, 32
-; RV64-NEXT:    or a0, a1, a0
-; RV64-NEXT:    sd a0, 0(a2)
-; RV64-NEXT:    ret
-;
-; RV64D-LABEL: mbb_int32_float_pair:
-; RV64D:       # %bb.0: # %entry
-; RV64D-NEXT:    fmv.x.w a2, fa0
-; RV64D-NEXT:    slli a2, a2, 32
-; RV64D-NEXT:    slli a0, a0, 32
-; RV64D-NEXT:    srli a0, a0, 32
-; RV64D-NEXT:    or a0, a2, a0
-; RV64D-NEXT:    sd a0, 0(a1)
-; RV64D-NEXT:    ret
+; RV32-RV64-LABEL: mbb_int32_float_pair:
+; RV32-RV64:       # %bb.0: # %entry
+; RV32-RV64-NEXT:    sw a0, 0(a2)
+; RV32-RV64-NEXT:    sw a1, 4(a2)
+; RV32-RV64-NEXT:    ret
+;
+; RV32D-RV64D-LABEL: mbb_int32_float_pair:
+; RV32D-RV64D:       # %bb.0: # %entry
+; RV32D-RV64D-NEXT:    sw a0, 0(a1)
+; RV32D-RV64D-NEXT:    fsw fa0, 4(a1)
+; RV32D-RV64D-NEXT:    ret
 entry:
   %t0 = bitcast float %tmp2 to i32
   br label %next
 next:
   %t1 = zext i32 %t0 to i64
   %t2 = shl nuw i64 %t1, 32
   %t3 = zext i32 %tmp1 to i64
   %t4 = or i64 %t2, %t3
   store i64 %t4, ptr %ref.tmp, align 8
   ret void
 }
 
 define void @mbb_int32_float_multi_stores(i32 %tmp1, float %tmp2, ptr %ref.tmp, ptr %ref.tmp1, i1 %cmp) {
-; RV32-LABEL: mbb_int32_float_multi_stores:
-; RV32:       # %bb.0: # %entry
-; RV32-NEXT:    sw a1, 4(a2)
-; RV32-NEXT:    andi a4, a4, 1
-; RV32-NEXT:    sw a0, 0(a2)
-; RV32-NEXT:    beqz a4, .LBB6_2
-; RV32-NEXT:  # %bb.1: # %bb2
-; RV32-NEXT:    sw a0, 0(a3)
-; RV32-NEXT:    sw a1, 4(a3)
-; RV32-NEXT:  .LBB6_2: # %exitbb
-; RV32-NEXT:    ret
-;
-; RV32D-LABEL: mbb_int32_float_multi_stores:
-; RV32D:       # %bb.0: # %entry
-; RV32D-NEXT:    fmv.x.w a4, fa0
-; RV32D-NEXT:    sw a4, 4(a1)
-; RV32D-NEXT:    andi a3, a3, 1
-; RV32D-NEXT:    sw a0, 0(a1)
-; RV32D-NEXT:    beqz a3, .LBB6_2
-; RV32D-NEXT:  # %bb.1: # %bb2
-; RV32D-NEXT:    sw a0, 0(a2)
-; RV32D-NEXT:    sw a4, 4(a2)
-; RV32D-NEXT:  .LBB6_2: # %exitbb
-; RV32D-NEXT:    ret
-;
-; RV64-LABEL: mbb_int32_float_multi_stores:
-; RV64:       # %bb.0: # %entry
-; RV64-NEXT:    slli a1, a1, 32
-; RV64-NEXT:    slli a0, a0, 32
-; RV64-NEXT:    srli a0, a0, 32
-; RV64-NEXT:    or a0, a1, a0
-; RV64-NEXT:    andi a4, a4, 1
-; RV64-NEXT:    sd a0, 0(a2)
-; RV64-NEXT:    beqz a4, .LBB6_2
-; RV64-NEXT:  # %bb.1: # %bb2
-; RV64-NEXT:    sd a0, 0(a3)
-; RV64-NEXT:  .LBB6_2: # %exitbb
-; RV64-NEXT:    ret
-;
-; RV64D-LABEL: mbb_int32_float_multi_stores:
-; RV64D:       # %bb.0: # %entry
-; RV64D-NEXT:    fmv.x.w a4, fa0
-; RV64D-NEXT:    slli a4, a4, 32
-; RV64D-NEXT:    slli a0, a0, 32
-; RV64D-NEXT:    srli a0, a0, 32
-; RV64D-NEXT:    or a0, a4, a0
-; RV64D-NEXT:    andi a3, a3, 1
-; RV64D-NEXT:    sd a0, 0(a1)
-; RV64D-NEXT:    beqz a3, .LBB6_2
-; RV64D-NEXT:  # %bb.1: # %bb2
-; RV64D-NEXT:    sd a0, 0(a2)
-; RV64D-NEXT:  .LBB6_2: # %exitbb
-; RV64D-NEXT:    ret
+; RV32-RV64-LABEL: mbb_int32_float_multi_stores:
+; RV32-RV64:       # %bb.0: # %entry
+; RV32-RV64-NEXT:    andi a4, a4, 1
+; RV32-RV64-NEXT:    sw a0, 0(a2)
+; RV32-RV64-NEXT:    sw a1, 4(a2)
+; RV32-RV64-NEXT:    beqz a4, .LBB6_2
+; RV32-RV64-NEXT:  # %bb.1: # %bb2
+; RV32-RV64-NEXT:    sw a0, 0(a3)
+; RV32-RV64-NEXT:    sw a1, 4(a3)
+; RV32-RV64-NEXT:  .LBB6_2: # %exitbb
+; RV32-RV64-NEXT:    ret
+;
+; RV32D-RV64D-LABEL: mbb_int32_float_multi_stores:
+; RV32D-RV64D:       # %bb.0: # %entry
+; RV32D-RV64D-NEXT:    andi a3, a3, 1
+; RV32D-RV64D-NEXT:    sw a0, 0(a1)
+; RV32D-RV64D-NEXT:    fsw fa0, 4(a1)
+; RV32D-RV64D-NEXT:    beqz a3, .LBB6_2
+; RV32D-RV64D-NEXT:  # %bb.1: # %bb2
+; RV32D-RV64D-NEXT:    sw a0, 0(a2)
+; RV32D-RV64D-NEXT:    fsw fa0, 4(a2)
+; RV32D-RV64D-NEXT:  .LBB6_2: # %exitbb
+; RV32D-RV64D-NEXT:    ret
 entry:
   %t0 = bitcast float %tmp2 to i32
   br label %bb1
 bb1:
   %t1 = zext i32 %t0 to i64
   %t2 = shl nuw i64 %t1, 32
   %t3 = zext i32 %tmp1 to i64
   %t4 = or i64 %t2, %t3
   store i64 %t4, ptr %ref.tmp, align 8
   br i1 %cmp, label %bb2, label %exitbb
 bb2:
   store i64 %t4, ptr %ref.tmp1, align 8
   br label %exitbb
 exitbb:
   ret void
 }
-;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
-; RV32-RV64: {{.*}}
-; RV32D-RV64D: {{.*}}