[RISCV] Add DAGCombine for (SplitF64 (ConstantFP x))

The SplitF64 node is used on RV32D to convert an f64 directly to a pair of i32
(necessary as bitcasting to i64 isn't legal). When performed on a ConstantFP,
this will result in a FP load from the constant pool followed by a store to
the stack and two integer loads from the stack (necessary as there is no way
to directly move between f64 FPRs and i32 GPRs on RV32D). It's always cheaper
to just materialise integers for the lo and hi parts of the FP constant, so do
that instead.

llvm-svn: 357341

Author: asb

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

@@ -633,6 +633,17 @@ SDValue RISCVTargetLowering::PerformDAGCombine(SDNode *N,
       return DCI.CombineTo(N, Op0.getOperand(0), Op0.getOperand(1));
 
     SDLoc DL(N);
+
+    // It's cheaper to materialise two 32-bit integers than to load a double
+    // from the constant pool and transfer it to integer registers through the
+    // stack.
+    if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op0)) {
+      APInt V = C->getValueAPF().bitcastToAPInt();
+      SDValue Lo = DAG.getConstant(V.trunc(32), DL, MVT::i32);
+      SDValue Hi = DAG.getConstant(V.lshr(32).trunc(32), DL, MVT::i32);
+      return DCI.CombineTo(N, Lo, Hi);
+    }
+
     // This is a target-specific version of a DAGCombine performed in
     // DAGCombiner::visitBITCAST. It performs the equivalent of:
     // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)

llvm/test/CodeGen/RISCV/double-calling-conv.ll

@@ -34,18 +34,13 @@ define double @caller_double_inreg() nounwind {
 ; RV32IFD:       # %bb.0:
 ; RV32IFD-NEXT:    addi sp, sp, -16
 ; RV32IFD-NEXT:    sw ra, 12(sp)
-; RV32IFD-NEXT:    lui a0, %hi(.LCPI1_0)
-; RV32IFD-NEXT:    addi a0, a0, %lo(.LCPI1_0)
-; RV32IFD-NEXT:    fld ft0, 0(a0)
-; RV32IFD-NEXT:    lui a0, %hi(.LCPI1_1)
-; RV32IFD-NEXT:    addi a0, a0, %lo(.LCPI1_1)
-; RV32IFD-NEXT:    fld ft1, 0(a0)
-; RV32IFD-NEXT:    fsd ft1, 0(sp)
-; RV32IFD-NEXT:    lw a0, 0(sp)
-; RV32IFD-NEXT:    lw a1, 4(sp)
-; RV32IFD-NEXT:    fsd ft0, 0(sp)
-; RV32IFD-NEXT:    lw a2, 0(sp)
-; RV32IFD-NEXT:    lw a3, 4(sp)
+; RV32IFD-NEXT:    lui a0, 262236
+; RV32IFD-NEXT:    addi a1, a0, 655
+; RV32IFD-NEXT:    lui a0, 377487
+; RV32IFD-NEXT:    addi a0, a0, 1475
+; RV32IFD-NEXT:    lui a2, 262364
+; RV32IFD-NEXT:    addi a3, a2, 655
+; RV32IFD-NEXT:    mv a2, a0
 ; RV32IFD-NEXT:    call callee_double_inreg
 ; RV32IFD-NEXT:    lw ra, 12(sp)
 ; RV32IFD-NEXT:    addi sp, sp, 16
@@ -78,29 +73,24 @@ define double @callee_double_split_reg_stack(i32 %a, i64 %b, i64 %c, double %d,
 define double @caller_double_split_reg_stack() nounwind {
 ; RV32IFD-LABEL: caller_double_split_reg_stack:
 ; RV32IFD:       # %bb.0:
-; RV32IFD-NEXT:    addi sp, sp, -32
-; RV32IFD-NEXT:    sw ra, 28(sp)
-; RV32IFD-NEXT:    lui a0, %hi(.LCPI3_0)
-; RV32IFD-NEXT:    addi a0, a0, %lo(.LCPI3_0)
-; RV32IFD-NEXT:    fld ft0, 0(a0)
-; RV32IFD-NEXT:    fsd ft0, 16(sp)
-; RV32IFD-NEXT:    lw a7, 16(sp)
-; RV32IFD-NEXT:    lw a0, 20(sp)
+; RV32IFD-NEXT:    addi sp, sp, -16
+; RV32IFD-NEXT:    sw ra, 12(sp)
+; RV32IFD-NEXT:    lui a0, 262510
+; RV32IFD-NEXT:    addi a0, a0, 327
 ; RV32IFD-NEXT:    sw a0, 0(sp)
-; RV32IFD-NEXT:    lui a0, %hi(.LCPI3_1)
-; RV32IFD-NEXT:    addi a0, a0, %lo(.LCPI3_1)
-; RV32IFD-NEXT:    fld ft0, 0(a0)
-; RV32IFD-NEXT:    fsd ft0, 16(sp)
-; RV32IFD-NEXT:    lw a5, 16(sp)
-; RV32IFD-NEXT:    lw a6, 20(sp)
+; RV32IFD-NEXT:    lui a0, 262446
+; RV32IFD-NEXT:    addi a6, a0, 327
+; RV32IFD-NEXT:    lui a0, 713032
+; RV32IFD-NEXT:    addi a5, a0, -1311
 ; RV32IFD-NEXT:    addi a0, zero, 1
 ; RV32IFD-NEXT:    addi a1, zero, 2
 ; RV32IFD-NEXT:    mv a2, zero
 ; RV32IFD-NEXT:    addi a3, zero, 3
 ; RV32IFD-NEXT:    mv a4, zero
+; RV32IFD-NEXT:    mv a7, a5
 ; RV32IFD-NEXT:    call callee_double_split_reg_stack
-; RV32IFD-NEXT:    lw ra, 28(sp)
-; RV32IFD-NEXT:    addi sp, sp, 32
+; RV32IFD-NEXT:    lw ra, 12(sp)
+; RV32IFD-NEXT:    addi sp, sp, 16
 ; RV32IFD-NEXT:    ret
   %1 = call double @callee_double_split_reg_stack(i32 1, i64 2, i64 3, double 4.72, double 5.72)
   ret double %1

llvm/test/CodeGen/RISCV/double-imm.ll

@@ -10,14 +10,10 @@ define double @double_imm() nounwind {
 ;
 ; RV32IFD-LABEL: double_imm:
 ; RV32IFD:       # %bb.0:
-; RV32IFD-NEXT:    addi sp, sp, -16
-; RV32IFD-NEXT:    lui a0, %hi(.LCPI0_0)
-; RV32IFD-NEXT:    addi a0, a0, %lo(.LCPI0_0)
-; RV32IFD-NEXT:    fld ft0, 0(a0)
-; RV32IFD-NEXT:    fsd ft0, 8(sp)
-; RV32IFD-NEXT:    lw a0, 8(sp)
-; RV32IFD-NEXT:    lw a1, 12(sp)
-; RV32IFD-NEXT:    addi sp, sp, 16
+; RV32IFD-NEXT:    lui a0, 345155
+; RV32IFD-NEXT:    addi a0, a0, -744
+; RV32IFD-NEXT:    lui a1, 262290
+; RV32IFD-NEXT:    addi a1, a1, 507
 ; RV32IFD-NEXT:    ret
 ;
 ; RV64IFD-LABEL: double_imm:

llvm/test/CodeGen/RISCV/double-previous-failure.ll

@@ -17,24 +17,20 @@ define i32 @main() nounwind {
 ; RV32IFD:       # %bb.0: # %entry
 ; RV32IFD-NEXT:    addi sp, sp, -16
 ; RV32IFD-NEXT:    sw ra, 12(sp)
-; RV32IFD-NEXT:    lui a0, %hi(.LCPI1_0)
-; RV32IFD-NEXT:    addi a0, a0, %lo(.LCPI1_0)
-; RV32IFD-NEXT:    fld ft0, 0(a0)
-; RV32IFD-NEXT:    fsd ft0, 0(sp)
-; RV32IFD-NEXT:    lw a0, 0(sp)
-; RV32IFD-NEXT:    lw a1, 4(sp)
+; RV32IFD-NEXT:    mv a0, zero
+; RV32IFD-NEXT:    lui a1, 262144
 ; RV32IFD-NEXT:    call test
-; RV32IFD-NEXT:    lui a2, %hi(.LCPI1_1)
-; RV32IFD-NEXT:    addi a2, a2, %lo(.LCPI1_1)
-; RV32IFD-NEXT:    fld ft1, 0(a2)
 ; RV32IFD-NEXT:    sw a0, 0(sp)
 ; RV32IFD-NEXT:    sw a1, 4(sp)
 ; RV32IFD-NEXT:    fld ft0, 0(sp)
+; RV32IFD-NEXT:    lui a0, %hi(.LCPI1_0)
+; RV32IFD-NEXT:    addi a0, a0, %lo(.LCPI1_0)
+; RV32IFD-NEXT:    fld ft1, 0(a0)
 ; RV32IFD-NEXT:    flt.d a0, ft0, ft1
 ; RV32IFD-NEXT:    bnez a0, .LBB1_3
 ; RV32IFD-NEXT:  # %bb.1: # %entry
-; RV32IFD-NEXT:    lui a0, %hi(.LCPI1_2)
-; RV32IFD-NEXT:    addi a0, a0, %lo(.LCPI1_2)
+; RV32IFD-NEXT:    lui a0, %hi(.LCPI1_1)
+; RV32IFD-NEXT:    addi a0, a0, %lo(.LCPI1_1)
 ; RV32IFD-NEXT:    fld ft1, 0(a0)
 ; RV32IFD-NEXT:    flt.d a0, ft1, ft0
 ; RV32IFD-NEXT:    xori a0, a0, 1