[LoongArch] Optimize inserting bitcasted integer element or bitcasting extracted fp element (#147043)

zhaoqi5 · web-flow · commit e74082703e22 · 2025-07-17T19:21:24.000+08:00
diff --git a/llvm/lib/Target/LoongArch/LoongArchFloat32InstrInfo.td b/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.
 //
 //===----------------------------------------------------------------------===//
 
diff --git a/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td b/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td
@@ -1651,6 +1651,10 @@ def : Pat<(vector_insert v8i32:$xd, GRLenVT:$rj, uimm3:$imm),
           (XVINSGR2VR_W v8i32:$xd, GRLenVT:$rj, uimm3:$imm)>;
 def : Pat<(vector_insert v4i64:$xd, GRLenVT:$rj, uimm2:$imm),
           (XVINSGR2VR_D v4i64:$xd, GRLenVT:$rj, uimm2:$imm)>;
+def : Pat<(vector_insert v8f32:$vd, (loongarch_movgr2fr_w_la64 GPR:$rj), uimm3:$imm),
+          (XVINSGR2VR_W $vd, $rj, uimm3:$imm)>;
+def : Pat<(vector_insert v4f64:$vd, (f64 (bitconvert i64:$rj)), uimm2:$imm),
+          (XVINSGR2VR_D $vd, $rj, uimm2:$imm)>;
 def : Pat<(vector_insert v8f32:$xd, (f32 (vector_extract v8f32:$xj, uimm3:$imm1)), uimm3:$imm2),
           (XVINSGR2VR_W $xd, (XVPICKVE2GR_W v8f32:$xj, uimm3:$imm1), uimm3:$imm2)>;
 def : Pat<(vector_insert v4f64:$xd, (f64 (vector_extract v4f64:$xj, uimm2:$imm1)), uimm2:$imm2),
@@ -1851,6 +1855,12 @@ 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.
 foreach imm = 16...31 in {
   defvar Imm = !and(imm, 15);
diff --git a/llvm/lib/Target/LoongArch/LoongArchLSXInstrInfo.td b/llvm/lib/Target/LoongArch/LoongArchLSXInstrInfo.td
@@ -1838,6 +1838,10 @@ def : Pat<(vector_insert v4i32:$vd, GRLenVT:$rj, uimm2:$imm),
           (VINSGR2VR_W v4i32:$vd, GRLenVT:$rj, uimm2:$imm)>;
 def : Pat<(vector_insert v2i64:$vd, GRLenVT:$rj, uimm1:$imm),
           (VINSGR2VR_D v2i64:$vd, GRLenVT:$rj, uimm1:$imm)>;
+def : Pat<(vector_insert v4f32:$vd, (loongarch_movgr2fr_w_la64 GPR:$rj), uimm2:$imm),
+          (VINSGR2VR_W $vd, $rj, uimm2:$imm)>;
+def : Pat<(vector_insert v2f64:$vd, (f64 (bitconvert i64:$rj)), uimm1:$imm),
+          (VINSGR2VR_D $vd, $rj, uimm1:$imm)>;
 def : Pat<(vector_insert v4f32:$vd, FPR32:$fj, uimm2:$imm),
           (VINSGR2VR_W $vd, (COPY_TO_REGCLASS FPR32:$fj, GPR), uimm2:$imm)>;
 def : Pat<(vector_insert v2f64:$vd, FPR64:$fj, uimm1:$imm),
@@ -2036,6 +2040,12 @@ 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)>;
diff --git a/llvm/test/CodeGen/LoongArch/lasx/ir-instruction/bitcast-extract-element.ll b/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
diff --git a/llvm/test/CodeGen/LoongArch/lasx/ir-instruction/insert-bitcast-element.ll b/llvm/test/CodeGen/LoongArch/lasx/ir-instruction/insert-bitcast-element.ll
@@ -4,8 +4,6 @@
 define <8 x float> @insert_bitcast_v8f32(<8 x float> %a, i32 %b) nounwind {
 ; CHECK-LABEL: insert_bitcast_v8f32:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    movgr2fr.w $fa1, $a0
-; CHECK-NEXT:    movfr2gr.s $a0, $fa1
 ; CHECK-NEXT:    xvinsgr2vr.w $xr0, $a0, 1
 ; CHECK-NEXT:    ret
 entry:
@@ -17,8 +15,6 @@ entry:
 define <4 x double> @insert_bitcast_v4f64(<4 x double> %a, i64 %b) nounwind {
 ; CHECK-LABEL: insert_bitcast_v4f64:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    movgr2fr.d $fa1, $a0
-; CHECK-NEXT:    movfr2gr.d $a0, $fa1
 ; CHECK-NEXT:    xvinsgr2vr.d $xr0, $a0, 1
 ; CHECK-NEXT:    ret
 entry:
diff --git a/llvm/test/CodeGen/LoongArch/lsx/ir-instruction/bitcast-extract-element.ll b/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
diff --git a/llvm/test/CodeGen/LoongArch/lsx/ir-instruction/insert-bitcast-element.ll b/llvm/test/CodeGen/LoongArch/lsx/ir-instruction/insert-bitcast-element.ll
@@ -4,8 +4,6 @@
 define <4 x float> @insert_bitcast_v4f32(<4 x float> %a, i32 %b) nounwind {
 ; CHECK-LABEL: insert_bitcast_v4f32:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    movgr2fr.w $fa1, $a0
-; CHECK-NEXT:    movfr2gr.s $a0, $fa1
 ; CHECK-NEXT:    vinsgr2vr.w $vr0, $a0, 1
 ; CHECK-NEXT:    ret
 entry:
@@ -17,8 +15,6 @@ entry:
 define <2 x double> @insert_bitcast_v2f64(<2 x double> %a, i64 %b) nounwind {
 ; CHECK-LABEL: insert_bitcast_v2f64:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    movgr2fr.d $fa1, $a0
-; CHECK-NEXT:    movfr2gr.d $a0, $fa1
 ; CHECK-NEXT:    vinsgr2vr.d $vr0, $a0, 1
 ; CHECK-NEXT:    ret
 entry:
