release/18.x: [X86] Add missing subvector_subreg_lowering for BF16 (#83720)

[llvmbot]

Backport ff72c83

Requested by: @phoebewang

[llvmbot]

@FreddyLeaf What do you think about merging this PR to the release branch?

[llvmbot]

@llvm/pr-subscribers-backend-x86

Author: None (llvmbot)

Changes

Backport ff72c83

Requested by: @phoebewang

---

Full diff: https://github.com/llvm/llvm-project/pull/83758.diff

3 Files Affected:

• (modified) llvm/lib/Target/X86/X86InstrVecCompiler.td (+3)
• (modified) llvm/test/CodeGen/X86/avx512bf16-vl-intrinsics.ll (+22)
• (modified) llvm/test/CodeGen/X86/bfloat.ll (+3-4)

diff --git a/llvm/lib/Target/X86/X86InstrVecCompiler.td b/llvm/lib/Target/X86/X86InstrVecCompiler.td
index bbd19cf8d5b25e..461b2badc13134 100644
--- a/llvm/lib/Target/X86/X86InstrVecCompiler.td
+++ b/llvm/lib/Target/X86/X86InstrVecCompiler.td
@@ -83,6 +83,7 @@ defm : subvector_subreg_lowering<VR128, v2f64, VR256, v4f64,  sub_xmm>;
 defm : subvector_subreg_lowering<VR128, v8i16, VR256, v16i16, sub_xmm>;
 defm : subvector_subreg_lowering<VR128, v16i8, VR256, v32i8,  sub_xmm>;
 defm : subvector_subreg_lowering<VR128, v8f16, VR256, v16f16, sub_xmm>;
+defm : subvector_subreg_lowering<VR128, v8bf16, VR256, v16bf16, sub_xmm>;
 
 // A 128-bit subvector extract from the first 512-bit vector position is a
 // subregister copy that needs no instruction. Likewise, a 128-bit subvector
@@ -95,6 +96,7 @@ defm : subvector_subreg_lowering<VR128, v2f64, VR512, v8f64,  sub_xmm>;
 defm : subvector_subreg_lowering<VR128, v8i16, VR512, v32i16, sub_xmm>;
 defm : subvector_subreg_lowering<VR128, v16i8, VR512, v64i8,  sub_xmm>;
 defm : subvector_subreg_lowering<VR128, v8f16, VR512, v32f16, sub_xmm>;
+defm : subvector_subreg_lowering<VR128, v8bf16, VR512, v32bf16, sub_xmm>;
 
 // A 128-bit subvector extract from the first 512-bit vector position is a
 // subregister copy that needs no instruction. Likewise, a 128-bit subvector
@@ -107,6 +109,7 @@ defm : subvector_subreg_lowering<VR256, v4f64,  VR512, v8f64,  sub_ymm>;
 defm : subvector_subreg_lowering<VR256, v16i16, VR512, v32i16, sub_ymm>;
 defm : subvector_subreg_lowering<VR256, v32i8,  VR512, v64i8,  sub_ymm>;
 defm : subvector_subreg_lowering<VR256, v16f16, VR512, v32f16, sub_ymm>;
+defm : subvector_subreg_lowering<VR256, v16bf16, VR512, v32bf16, sub_ymm>;
 
 
 // If we're inserting into an all zeros vector, just use a plain move which
diff --git a/llvm/test/CodeGen/X86/avx512bf16-vl-intrinsics.ll b/llvm/test/CodeGen/X86/avx512bf16-vl-intrinsics.ll
index 0826faa1071b01..482713e12d15c7 100644
--- a/llvm/test/CodeGen/X86/avx512bf16-vl-intrinsics.ll
+++ b/llvm/test/CodeGen/X86/avx512bf16-vl-intrinsics.ll
@@ -381,3 +381,25 @@ entry:
   %1 = shufflevector <8 x bfloat> %0, <8 x bfloat> undef, <16 x i32> zeroinitializer
   ret <16 x bfloat> %1
 }
+
+define <16 x i32> @pr83358() {
+; X86-LABEL: pr83358:
+; X86:       # %bb.0:
+; X86-NEXT:    vcvtneps2bf16y {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0 # encoding: [0x62,0xf2,0x7e,0x28,0x72,0x05,A,A,A,A]
+; X86-NEXT:    # fixup A - offset: 6, value: {{\.?LCPI[0-9]+_[0-9]+}}, kind: FK_Data_4
+; X86-NEXT:    vshufi64x2 $0, %zmm0, %zmm0, %zmm0 # encoding: [0x62,0xf3,0xfd,0x48,0x43,0xc0,0x00]
+; X86-NEXT:    # zmm0 = zmm0[0,1,0,1,0,1,0,1]
+; X86-NEXT:    retl # encoding: [0xc3]
+;
+; X64-LABEL: pr83358:
+; X64:       # %bb.0:
+; X64-NEXT:    vcvtneps2bf16y {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0 # encoding: [0x62,0xf2,0x7e,0x28,0x72,0x05,A,A,A,A]
+; X64-NEXT:    # fixup A - offset: 6, value: {{\.?LCPI[0-9]+_[0-9]+}}-4, kind: reloc_riprel_4byte
+; X64-NEXT:    vshufi64x2 $0, %zmm0, %zmm0, %zmm0 # encoding: [0x62,0xf3,0xfd,0x48,0x43,0xc0,0x00]
+; X64-NEXT:    # zmm0 = zmm0[0,1,0,1,0,1,0,1]
+; X64-NEXT:    retq # encoding: [0xc3]
+  %1 = call <8 x bfloat> @llvm.x86.avx512bf16.cvtneps2bf16.256(<8 x float> <float 1.000000e+00, float 2.000000e+00, float 3.000000e+00, float 4.000000e+00, float 5.000000e+00, float 6.000000e+00, float 7.000000e+00, float 8.000000e+00>)
+  %2 = bitcast <8 x bfloat> %1 to <4 x i32>
+  %3 = shufflevector <4 x i32> %2, <4 x i32> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
+  ret <16 x i32> %3
+}
diff --git a/llvm/test/CodeGen/X86/bfloat.ll b/llvm/test/CodeGen/X86/bfloat.ll
index f2d3c4fb34199e..0042d477f3b364 100644
--- a/llvm/test/CodeGen/X86/bfloat.ll
+++ b/llvm/test/CodeGen/X86/bfloat.ll
@@ -511,7 +511,7 @@ define void @fold_ext_trunc(ptr %pa, ptr %pc) nounwind {
 define bfloat @fold_ext_trunc2(bfloat %a) nounwind {
 ; X86-LABEL: fold_ext_trunc2:
 ; X86:       # %bb.0:
-; X86-NEXT:    vmovsh {{[0-9]+}}(%esp), %xmm0
+; X86-NEXT:    vmovsh {{.*#+}} xmm0 = mem[0],zero,zero,zero,zero,zero,zero,zero
 ; X86-NEXT:    retl
 ;
 ; CHECK-LABEL: fold_ext_trunc2:
@@ -934,8 +934,8 @@ define <8 x bfloat> @addv(<8 x bfloat> %a, <8 x bfloat> %b) nounwind {
 define <2 x bfloat> @pr62997(bfloat %a, bfloat %b) {
 ; X86-LABEL: pr62997:
 ; X86:       # %bb.0:
-; X86-NEXT:    vmovsh {{[0-9]+}}(%esp), %xmm0
-; X86-NEXT:    vmovsh {{[0-9]+}}(%esp), %xmm1
+; X86-NEXT:    vmovsh {{.*#+}} xmm0 = mem[0],zero,zero,zero,zero,zero,zero,zero
+; X86-NEXT:    vmovsh {{.*#+}} xmm1 = mem[0],zero,zero,zero,zero,zero,zero,zero
 ; X86-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
 ; X86-NEXT:    retl
 ;
@@ -2423,7 +2423,6 @@ define <16 x bfloat> @fptrunc_v16f32(<16 x float> %a) nounwind {
 ; AVXNC-LABEL: fptrunc_v16f32:
 ; AVXNC:       # %bb.0:
 ; AVXNC-NEXT:    {vex} vcvtneps2bf16 %ymm0, %xmm0
-; AVXNC-NEXT:    vinsertf128 $0, %xmm0, %ymm0, %ymm0
 ; AVXNC-NEXT:    {vex} vcvtneps2bf16 %ymm1, %xmm1
 ; AVXNC-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVXNC-NEXT:    retq

[FreddyLeaf]

LGTM

[phoebewang]

Some change related on trunk code, I'll create a manual cherry-pick.