[LoongArch] Optimize bitcasting vector-extracted fp element to integer

Author: zhaoqi5

llvm/lib/Target/LoongArch/LoongArchFloat32InstrInfo.td

@@ -6,7 +6,7 @@
 //
 //===----------------------------------------------------------------------===//
 //
-// This file describes the baisc single-precision floating-point instructions.
+// This file describes the basic single-precision floating-point instructions.
 //
 //===----------------------------------------------------------------------===//
 

llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td

@@ -1793,6 +1793,12 @@ foreach vt = [v32i8, v16i16, v8i32, v4i64, v8f32, v4f64] in {
   def  : RegRegStPat<store, XVSTX, LASX256, vt>;
 }
 
+// Bitcast float/double element extracted from vector to integer.
+def : Pat<(loongarch_movfr2gr_s_la64 (f32 (vector_extract v8f32:$xj, uimm3:$imm))),
+          (XVPICKVE2GR_W v8f32:$xj, uimm3:$imm)>;
+def : Pat<(i64 (bitconvert (f64 (vector_extract v4f64:$xj, uimm2:$imm)))),
+          (XVPICKVE2GR_D v4f64:$xj, uimm2:$imm)>;
+
 // Vector extraction with constant index.
 def : Pat<(i64 (vector_extract v32i8:$xj, uimm4:$imm)),
           (VPICKVE2GR_B (EXTRACT_SUBREG v32i8:$xj, sub_128), uimm4:$imm)>;

llvm/lib/Target/LoongArch/LoongArchLSXInstrInfo.td

@@ -1993,6 +1993,12 @@ foreach vt = [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64] in {
   def  : RegRegStPat<store, VSTX, LSX128, vt>;
 }
 
+// Bitcast float/double element extracted from vector to integer.
+def : Pat<(loongarch_movfr2gr_s_la64 (f32 (vector_extract v4f32:$vj, uimm2:$imm))),
+          (VPICKVE2GR_W v4f32:$vj, uimm2:$imm)>;
+def : Pat<(i64 (bitconvert (f64 (vector_extract v2f64:$vj, uimm1:$imm)))),
+          (VPICKVE2GR_D v2f64:$vj, uimm1:$imm)>;
+
 // Vector extraction with constant index.
 def : Pat<(i64 (vector_extract v16i8:$vj, uimm4:$imm)),
           (VPICKVE2GR_B v16i8:$vj, uimm4:$imm)>;

llvm/test/CodeGen/LoongArch/lasx/ir-instruction/bitcast-extract-element.ll

@@ -5,8 +5,6 @@ define i32 @bitcast_extract_v8f32(<8 x float> %a) nounwind {
 ; CHECK-LABEL: bitcast_extract_v8f32:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    xvpickve2gr.w $a0, $xr0, 7
-; CHECK-NEXT:    movgr2fr.w $fa0, $a0
-; CHECK-NEXT:    movfr2gr.s $a0, $fa0
 ; CHECK-NEXT:    ret
 entry:
   %b = extractelement <8 x float> %a, i32 7
@@ -18,8 +16,6 @@ define i64 @bitcast_extract_v4f64(<4 x double> %a) nounwind {
 ; CHECK-LABEL: bitcast_extract_v4f64:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    xvpickve2gr.d $a0, $xr0, 3
-; CHECK-NEXT:    movgr2fr.d $fa0, $a0
-; CHECK-NEXT:    movfr2gr.d $a0, $fa0
 ; CHECK-NEXT:    ret
 entry:
   %b = extractelement <4 x double> %a, i32 3

llvm/test/CodeGen/LoongArch/lsx/ir-instruction/bitcast-extract-element.ll

@@ -4,8 +4,7 @@
 define i32 @bitcast_extract_v4f32(<4 x float> %a) nounwind {
 ; CHECK-LABEL: bitcast_extract_v4f32:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vreplvei.w $vr0, $vr0, 3
-; CHECK-NEXT:    movfr2gr.s $a0, $fa0
+; CHECK-NEXT:    vpickve2gr.w $a0, $vr0, 3
 ; CHECK-NEXT:    ret
 entry:
   %b = extractelement <4 x float> %a, i32 3
@@ -16,8 +15,7 @@ entry:
 define i64 @bitcast_extract_v2f64(<2 x double> %a) nounwind {
 ; CHECK-LABEL: bitcast_extract_v2f64:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vreplvei.d $vr0, $vr0, 1
-; CHECK-NEXT:    movfr2gr.d $a0, $fa0
+; CHECK-NEXT:    vpickve2gr.d $a0, $vr0, 1
 ; CHECK-NEXT:    ret
 entry:
   %b = extractelement <2 x double> %a, i32 1