[X86] Use X86FixupInstTunings to select between (V)MOVSS/D and (V)BLENDPS/D (#143895)

Fix #142588
Following @RKSimon’s suggestion, the transformation applies only when
the blend mask is exactly 1, indicating that the instruction behaves
like a move. Additionally, the conversion will only be performed when
optimizing for size or when the target prefers MOVSS/D over BLENDPS/D
for performance reasons.

The switch-case instructions were identified with GPT O.O .

Co-authored-by: Simon Pilgrim <[email protected]>

Author: houngkoungting

llvm/lib/Target/X86/X86FixupInstTuning.cpp

@@ -222,7 +222,22 @@ bool X86FixupInstTuningPass::processInstruction(
     return ProcessUNPCKToIntDomain(NewOpc);
   };
 
+  auto ProcessBLENDToMOV = [&](unsigned MovOpc) -> bool {
+    if (MI.getOperand(NumOperands - 1).getImm() != 1)
+      return false;
+    bool Force = MF.getFunction().hasOptSize();
+    if (!Force && !NewOpcPreferable(MovOpc))
+      return false;
+    MI.setDesc(TII->get(MovOpc));
+    MI.removeOperand(NumOperands - 1);
+    return true;
+  };
+
   switch (Opc) {
+  case X86::VBLENDPSrri:
+    return ProcessBLENDToMOV(X86::VMOVSSrr);
+  case X86::VBLENDPDrri:
+    return ProcessBLENDToMOV(X86::VMOVSDrr);
   case X86::VPERMILPDri:
     return ProcessVPERMILPDri(X86::VSHUFPDrri);
   case X86::VPERMILPDYri:

llvm/test/CodeGen/X86/2012-01-12-extract-sv.ll

@@ -11,7 +11,7 @@ define void @endless_loop() {
 ; AVX1-NEXT:    vxorps %xmm2, %xmm2, %xmm2
 ; AVX1-NEXT:    vblendps {{.*#+}} ymm1 = ymm2[0,1,2,3,4,5,6],ymm1[7]
 ; AVX1-NEXT:    vxorps %xmm2, %xmm2, %xmm2
-; AVX1-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm2[1,2,3]
+; AVX1-NEXT:    vmovss {{.*#+}} xmm0 = xmm0[0],xmm2[1,2,3]
 ; AVX1-NEXT:    vmovaps %ymm0, (%eax)
 ; AVX1-NEXT:    vmovaps %ymm1, (%eax)
 ; AVX1-NEXT:    vzeroupper
@@ -21,7 +21,7 @@ define void @endless_loop() {
 ; AVX2:       # %bb.0: # %entry
 ; AVX2-NEXT:    vbroadcastss (%eax), %xmm0
 ; AVX2-NEXT:    vxorps %xmm1, %xmm1, %xmm1
-; AVX2-NEXT:    vblendps {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]
+; AVX2-NEXT:    vmovss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]
 ; AVX2-NEXT:    vbroadcastss %xmm0, %ymm0
 ; AVX2-NEXT:    vxorps %xmm2, %xmm2, %xmm2
 ; AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3,4,5,6],ymm0[7]

llvm/test/CodeGen/X86/avx-insertelt.ll

@@ -94,7 +94,7 @@ define <8 x float> @insert_f32_firstelt_of_high_subvector(<8 x float> %x, float
 ; AVX-LABEL: insert_f32_firstelt_of_high_subvector:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm2
-; AVX-NEXT:    vblendps {{.*#+}} xmm1 = xmm1[0],xmm2[1,2,3]
+; AVX-NEXT:    vmovss {{.*#+}} xmm1 = xmm1[0],xmm2[1,2,3]
 ; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX-NEXT:    retq
 ;
@@ -202,9 +202,9 @@ define <4 x i64> @insert_i64_firstelt_of_high_subvector(<4 x i64> %x, i64 %s) {
 define <8 x float> @insert_f32_firstelts(<8 x float> %x, float %s) {
 ; AVX-LABEL: insert_f32_firstelts:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vblendps {{.*#+}} xmm2 = xmm1[0],xmm0[1,2,3]
+; AVX-NEXT:    vmovss {{.*#+}} xmm2 = xmm1[0],xmm0[1,2,3]
 ; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3]
+; AVX-NEXT:    vmovss {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3]
 ; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm2, %ymm0
 ; AVX-NEXT:    retq
 ;

llvm/test/CodeGen/X86/avx512-intrinsics-fast-isel.ll

@@ -1843,7 +1843,7 @@ define <2 x double> @test_mm_cvtu64_sd(<2 x double> %__A, i64 %__B) {
 ; X86-NEXT:    vsubpd {{\.?LCPI[0-9]+_[0-9]+}}, %xmm1, %xmm1
 ; X86-NEXT:    vshufpd {{.*#+}} xmm2 = xmm1[1,0]
 ; X86-NEXT:    vaddsd %xmm1, %xmm2, %xmm1
-; X86-NEXT:    vblendpd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
+; X86-NEXT:    vmovsd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: test_mm_cvtu64_sd:
@@ -1891,7 +1891,7 @@ define <4 x float> @test_mm_cvtu64_ss(<4 x float> %__A, i64 %__B) {
 ; X86-NEXT:    fadds {{\.?LCPI[0-9]+_[0-9]+}}(,%eax,4)
 ; X86-NEXT:    fstps {{[0-9]+}}(%esp)
 ; X86-NEXT:    vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero
-; X86-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3]
+; X86-NEXT:    vmovss {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3]
 ; X86-NEXT:    movl %ebp, %esp
 ; X86-NEXT:    popl %ebp
 ; X86-NEXT:    .cfi_def_cfa %esp, 4

llvm/test/CodeGen/X86/avx512-intrinsics-upgrade.ll

@@ -10483,7 +10483,7 @@ define <4 x float>@test_int_x86_avx512_maskz_vfmadd_ss_rm(<4 x float> %x0, <4 x
 ; CHECK-LABEL: test_int_x86_avx512_maskz_vfmadd_ss_rm:
 ; CHECK:       ## %bb.0:
 ; CHECK-NEXT:    vxorps %xmm1, %xmm1, %xmm1 ## encoding: [0xc5,0xf0,0x57,0xc9]
-; CHECK-NEXT:    vblendps $1, %xmm1, %xmm0, %xmm0 ## encoding: [0xc4,0xe3,0x79,0x0c,0xc1,0x01]
+; CHECK-NEXT:    vmovss %xmm1, %xmm0, %xmm0 ## encoding: [0xc5,0xfa,0x10,0xc1]
 ; CHECK-NEXT:    ## xmm0 = xmm1[0],xmm0[1,2,3]
 ; CHECK-NEXT:    ret{{[l|q]}} ## encoding: [0xc3]
   %q = load float, ptr %ptr_b

llvm/test/CodeGen/X86/avx512-intrinsics.ll

@@ -6505,7 +6505,7 @@ define <4 x float>@test_int_x86_avx512_maskz_vfmadd_ss_rm(<4 x float> %x0, <4 x
 ; CHECK-LABEL: test_int_x86_avx512_maskz_vfmadd_ss_rm:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vxorps %xmm1, %xmm1, %xmm1
-; CHECK-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3]
+; CHECK-NEXT:    vmovss {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3]
 ; CHECK-NEXT:    ret{{[l|q]}}
   %q = load float, ptr %ptr_b
   %vecinit.i = insertelement <4 x float> undef, float %q, i32 0

llvm/test/CodeGen/X86/avx512copy-intrinsics.ll

@@ -11,7 +11,7 @@ define <4 x i32> @test_mm_move_epi32(<4 x i32> %a0) nounwind {
 ; NOAVX512MOVZXC-LABEL: test_mm_move_epi32:
 ; NOAVX512MOVZXC:       # %bb.0:
 ; NOAVX512MOVZXC-NEXT:    vxorps %xmm1, %xmm1, %xmm1 # encoding: [0xc5,0xf0,0x57,0xc9]
-; NOAVX512MOVZXC-NEXT:    vblendps $1, %xmm0, %xmm1, %xmm0 # encoding: [0xc4,0xe3,0x71,0x0c,0xc0,0x01]
+; NOAVX512MOVZXC-NEXT:    vmovss %xmm0, %xmm1, %xmm0 # encoding: [0xc5,0xf2,0x10,0xc0]
 ; NOAVX512MOVZXC-NEXT:    # xmm0 = xmm0[0],xmm1[1,2,3]
 ; NOAVX512MOVZXC-NEXT:    retq # encoding: [0xc3]
   %res = shufflevector <4 x i32> %a0, <4 x i32> zeroinitializer, <4 x i32> <i32 0, i32 4, i32 4, i32 4>

llvm/test/CodeGen/X86/build-vector-512.ll

@@ -578,7 +578,7 @@ define <16 x float> @test_buildvector_16f32_2_var(float %a0, float %a1) {
 ; AVX-32-NEXT:    vpmovsxbd {{.*#+}} xmm1 = [0,17,0,0]
 ; AVX-32-NEXT:    vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero
 ; AVX-32-NEXT:    vpermi2ps %zmm0, %zmm2, %zmm1
-; AVX-32-NEXT:    vblendps {{.*#+}} xmm3 = xmm2[0],xmm0[1,2,3]
+; AVX-32-NEXT:    vmovss {{.*#+}} xmm3 = xmm2[0],xmm0[1,2,3]
 ; AVX-32-NEXT:    vinsertf128 $1, %xmm3, %ymm1, %ymm1
 ; AVX-32-NEXT:    vinsertps {{.*#+}} xmm3 = xmm0[0,1,2],xmm2[0]
 ; AVX-32-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[2,3]
@@ -626,7 +626,7 @@ define <16 x float> @test_buildvector_16f32_2_load(ptr %p0, ptr %p1) {
 ; AVX-32-NEXT:    vbroadcastss (%ecx), %xmm1
 ; AVX-32-NEXT:    vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero
 ; AVX-32-NEXT:    vpermi2ps %zmm1, %zmm2, %zmm0
-; AVX-32-NEXT:    vblendps {{.*#+}} xmm3 = xmm2[0],xmm1[1,2,3]
+; AVX-32-NEXT:    vmovss {{.*#+}} xmm3 = xmm2[0],xmm1[1,2,3]
 ; AVX-32-NEXT:    vinsertf128 $1, %xmm3, %ymm0, %ymm0
 ; AVX-32-NEXT:    vinsertps {{.*#+}} xmm3 = xmm1[0,1,2],xmm2[0]
 ; AVX-32-NEXT:    vinsertps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[2,3]
@@ -640,7 +640,7 @@ define <16 x float> @test_buildvector_16f32_2_load(ptr %p0, ptr %p1) {
 ; AVX-64-NEXT:    vbroadcastss (%rdi), %xmm1
 ; AVX-64-NEXT:    vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero
 ; AVX-64-NEXT:    vpermi2ps %zmm1, %zmm2, %zmm0
-; AVX-64-NEXT:    vblendps {{.*#+}} xmm3 = xmm2[0],xmm1[1,2,3]
+; AVX-64-NEXT:    vmovss {{.*#+}} xmm3 = xmm2[0],xmm1[1,2,3]
 ; AVX-64-NEXT:    vinsertf128 $1, %xmm3, %ymm0, %ymm0
 ; AVX-64-NEXT:    vinsertps {{.*#+}} xmm3 = xmm1[0,1,2],xmm2[0]
 ; AVX-64-NEXT:    vinsertps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[2,3]

llvm/test/CodeGen/X86/buildvec-extract.ll

@@ -42,7 +42,7 @@ define <2 x i64> @extract0_i32_zext_insert0_i64_zero(<4 x i32> %x) {
 ; AVX-LABEL: extract0_i32_zext_insert0_i64_zero:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vxorps %xmm1, %xmm1, %xmm1
-; AVX-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
+; AVX-NEXT:    vmovss {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
 ; AVX-NEXT:    retq
   %e = extractelement <4 x i32> %x, i32 0
   %z = zext i32 %e to i64
@@ -85,7 +85,7 @@ define <2 x i64> @extract1_i32_zext_insert0_i64_zero(<4 x i32> %x) {
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
 ; AVX-NEXT:    vxorps %xmm1, %xmm1, %xmm1
-; AVX-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
+; AVX-NEXT:    vmovss {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
 ; AVX-NEXT:    retq
   %e = extractelement <4 x i32> %x, i32 1
   %z = zext i32 %e to i64
@@ -130,7 +130,7 @@ define <2 x i64> @extract2_i32_zext_insert0_i64_zero(<4 x i32> %x) {
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[2,3,2,3]
 ; AVX-NEXT:    vxorps %xmm1, %xmm1, %xmm1
-; AVX-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
+; AVX-NEXT:    vmovss {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
 ; AVX-NEXT:    retq
   %e = extractelement <4 x i32> %x, i32 2
   %z = zext i32 %e to i64

llvm/test/CodeGen/X86/canonicalize-vars-f16-type.ll

@@ -51,7 +51,7 @@ define void @v_test_canonicalize__half(half addrspace(1)* %out) nounwind {
 ; AVX512-NEXT:    vcvtph2ps %xmm1, %xmm1
 ; AVX512-NEXT:    vmulss %xmm0, %xmm1, %xmm0
 ; AVX512-NEXT:    vxorps %xmm1, %xmm1, %xmm1
-; AVX512-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
+; AVX512-NEXT:    vmovss {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
 ; AVX512-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
 ; AVX512-NEXT:    vpextrw $0, %xmm0, (%rdi)
 ; AVX512-NEXT:    retq
@@ -149,7 +149,7 @@ define half @complex_canonicalize_fmul_half(half %a, half %b) nounwind {
 ; AVX512-NEXT:    vcvtph2ps %xmm2, %xmm2
 ; AVX512-NEXT:    vmulss %xmm2, %xmm0, %xmm0
 ; AVX512-NEXT:    vxorps %xmm2, %xmm2, %xmm2
-; AVX512-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm2[1,2,3]
+; AVX512-NEXT:    vmovss {{.*#+}} xmm0 = xmm0[0],xmm2[1,2,3]
 ; AVX512-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
 ; AVX512-NEXT:    vcvtph2ps %xmm0, %xmm0
 ; AVX512-NEXT:    vsubss %xmm1, %xmm0, %xmm0
@@ -235,12 +235,12 @@ define void @v_test_canonicalize_v2half(<2 x half> addrspace(1)* %out) nounwind
 ; AVX512-NEXT:    vcvtph2ps %xmm2, %xmm2
 ; AVX512-NEXT:    vmulss %xmm1, %xmm2, %xmm2
 ; AVX512-NEXT:    vxorps %xmm3, %xmm3, %xmm3
-; AVX512-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0],xmm3[1,2,3]
+; AVX512-NEXT:    vmovss {{.*#+}} xmm2 = xmm2[0],xmm3[1,2,3]
 ; AVX512-NEXT:    vcvtps2ph $4, %xmm2, %xmm2
 ; AVX512-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
 ; AVX512-NEXT:    vcvtph2ps %xmm0, %xmm0
 ; AVX512-NEXT:    vmulss %xmm1, %xmm0, %xmm0
-; AVX512-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm3[1,2,3]
+; AVX512-NEXT:    vmovss {{.*#+}} xmm0 = xmm0[0],xmm3[1,2,3]
 ; AVX512-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
 ; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
 ; AVX512-NEXT:    vmovd %xmm0, (%rdi)