| ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py |
| ; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=avx512fp16 -O3 | FileCheck %s --check-prefixes=CHECK |
| ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx512fp16 -O3 | FileCheck %s --check-prefixes=CHECK |
| |
| |
| declare <8 x i64> @llvm.experimental.constrained.fptosi.v8i64.v8f16(<8 x half>, metadata) |
| declare <8 x i64> @llvm.experimental.constrained.fptoui.v8i64.v8f16(<8 x half>, metadata) |
| declare <16 x i32> @llvm.experimental.constrained.fptosi.v16i32.v16f16(<16 x half>, metadata) |
| declare <16 x i32> @llvm.experimental.constrained.fptoui.v16i32.v16f16(<16 x half>, metadata) |
| declare <32 x i16> @llvm.experimental.constrained.fptosi.v32i16.v32f16(<32 x half>, metadata) |
| declare <32 x i16> @llvm.experimental.constrained.fptoui.v32i16.v32f16(<32 x half>, metadata) |
| declare <32 x i8> @llvm.experimental.constrained.fptosi.v32i8.v32f16(<32 x half>, metadata) |
| declare <32 x i8> @llvm.experimental.constrained.fptoui.v32i8.v32f16(<32 x half>, metadata) |
| declare <32 x i1> @llvm.experimental.constrained.fptosi.v32i1.v32f16(<32 x half>, metadata) |
| declare <32 x i1> @llvm.experimental.constrained.fptoui.v32i1.v32f16(<32 x half>, metadata) |
| |
| define <8 x i64> @strict_vector_fptosi_v8f16_to_v8i64(<8 x half> %a) #0 { |
| ; CHECK-LABEL: strict_vector_fptosi_v8f16_to_v8i64: |
| ; CHECK: # %bb.0: |
| ; CHECK-NEXT: vcvttph2qq %xmm0, %zmm0 |
| ; CHECK-NEXT: ret{{[l|q]}} |
| %ret = call <8 x i64> @llvm.experimental.constrained.fptosi.v8i64.v8f16(<8 x half> %a, |
| metadata !"fpexcept.strict") #0 |
| ret <8 x i64> %ret |
| } |
| |
| define <8 x i64> @strict_vector_fptoui_v8f16_to_v8i64(<8 x half> %a) #0 { |
| ; CHECK-LABEL: strict_vector_fptoui_v8f16_to_v8i64: |
| ; CHECK: # %bb.0: |
| ; CHECK-NEXT: vcvttph2uqq %xmm0, %zmm0 |
| ; CHECK-NEXT: ret{{[l|q]}} |
| %ret = call <8 x i64> @llvm.experimental.constrained.fptoui.v8i64.v8f16(<8 x half> %a, |
| metadata !"fpexcept.strict") #0 |
| ret <8 x i64> %ret |
| } |
| |
| define <16 x i32> @strict_vector_fptosi_v16f16_to_v16i32(<16 x half> %a) #0 { |
| ; CHECK-LABEL: strict_vector_fptosi_v16f16_to_v16i32: |
| ; CHECK: # %bb.0: |
| ; CHECK-NEXT: vcvttph2dq %ymm0, %zmm0 |
| ; CHECK-NEXT: ret{{[l|q]}} |
| %ret = call <16 x i32> @llvm.experimental.constrained.fptosi.v16i32.v16f16(<16 x half> %a, |
| metadata !"fpexcept.strict") #0 |
| ret <16 x i32> %ret |
| } |
| |
| define <16 x i32> @strict_vector_fptoui_v16f16_to_v16i32(<16 x half> %a) #0 { |
| ; CHECK-LABEL: strict_vector_fptoui_v16f16_to_v16i32: |
| ; CHECK: # %bb.0: |
| ; CHECK-NEXT: vcvttph2udq %ymm0, %zmm0 |
| ; CHECK-NEXT: ret{{[l|q]}} |
| %ret = call <16 x i32> @llvm.experimental.constrained.fptoui.v16i32.v16f16(<16 x half> %a, |
| metadata !"fpexcept.strict") #0 |
| ret <16 x i32> %ret |
| } |
| |
| define <32 x i16> @strict_vector_fptosi_v32f16_to_v32i16(<32 x half> %a) #0 { |
| ; CHECK-LABEL: strict_vector_fptosi_v32f16_to_v32i16: |
| ; CHECK: # %bb.0: |
| ; CHECK-NEXT: vcvttph2w %zmm0, %zmm0 |
| ; CHECK-NEXT: ret{{[l|q]}} |
| %ret = call <32 x i16> @llvm.experimental.constrained.fptosi.v32i16.v32f16(<32 x half> %a, |
| metadata !"fpexcept.strict") #0 |
| ret <32 x i16> %ret |
| } |
| |
| define <32 x i16> @strict_vector_fptoui_v32f16_to_v32i16(<32 x half> %a) #0 { |
| ; CHECK-LABEL: strict_vector_fptoui_v32f16_to_v32i16: |
| ; CHECK: # %bb.0: |
| ; CHECK-NEXT: vcvttph2uw %zmm0, %zmm0 |
| ; CHECK-NEXT: ret{{[l|q]}} |
| %ret = call <32 x i16> @llvm.experimental.constrained.fptoui.v32i16.v32f16(<32 x half> %a, |
| metadata !"fpexcept.strict") #0 |
| ret <32 x i16> %ret |
| } |
| |
| define <32 x i8> @strict_vector_fptosi_v32f16_to_v32i8(<32 x half> %a) #0 { |
| ; CHECK-LABEL: strict_vector_fptosi_v32f16_to_v32i8: |
| ; CHECK: # %bb.0: |
| ; CHECK-NEXT: vcvttph2w %zmm0, %zmm0 |
| ; CHECK-NEXT: vpmovwb %zmm0, %ymm0 |
| ; CHECK-NEXT: ret{{[l|q]}} |
| %ret = call <32 x i8> @llvm.experimental.constrained.fptosi.v32i8.v32f16(<32 x half> %a, |
| metadata !"fpexcept.strict") #0 |
| ret <32 x i8> %ret |
| } |
| |
| define <32 x i8> @strict_vector_fptoui_v32f16_to_v32i8(<32 x half> %a) #0 { |
| ; CHECK-LABEL: strict_vector_fptoui_v32f16_to_v32i8: |
| ; CHECK: # %bb.0: |
| ; CHECK-NEXT: vcvttph2w %zmm0, %zmm0 |
| ; CHECK-NEXT: vpmovwb %zmm0, %ymm0 |
| ; CHECK-NEXT: ret{{[l|q]}} |
| %ret = call <32 x i8> @llvm.experimental.constrained.fptoui.v32i8.v32f16(<32 x half> %a, |
| metadata !"fpexcept.strict") #0 |
| ret <32 x i8> %ret |
| } |
| |
| define <32 x i1> @strict_vector_fptosi_v32f16_to_v32i1(<32 x half> %a) #0 { |
| ; CHECK-LABEL: strict_vector_fptosi_v32f16_to_v32i1: |
| ; CHECK: # %bb.0: |
| ; CHECK-NEXT: vcvttph2w %zmm0, %zmm0 |
| ; CHECK-NEXT: vpmovw2m %zmm0, %k0 |
| ; CHECK-NEXT: vpmovm2b %k0, %ymm0 |
| ; CHECK-NEXT: ret{{[l|q]}} |
| %ret = call <32 x i1> @llvm.experimental.constrained.fptosi.v32i1.v32f16(<32 x half> %a, |
| metadata !"fpexcept.strict") #0 |
| ret <32 x i1> %ret |
| } |
| |
| define <32 x i1> @strict_vector_fptoui_v32f16_to_v32i1(<32 x half> %a) #0 { |
| ; CHECK-LABEL: strict_vector_fptoui_v32f16_to_v32i1: |
| ; CHECK: # %bb.0: |
| ; CHECK-NEXT: vcvttph2w %zmm0, %zmm0 |
| ; CHECK-NEXT: vpsllw $15, %zmm0, %zmm0 |
| ; CHECK-NEXT: vpmovw2m %zmm0, %k0 |
| ; CHECK-NEXT: vpmovm2b %k0, %ymm0 |
| ; CHECK-NEXT: ret{{[l|q]}} |
| %ret = call <32 x i1> @llvm.experimental.constrained.fptoui.v32i1.v32f16(<32 x half> %a, |
| metadata !"fpexcept.strict") #0 |
| ret <32 x i1> %ret |
| } |
| |
| attributes #0 = { strictfp } |