diff options
Diffstat (limited to 'llvm/test/CodeGen/AMDGPU/fptrunc.ll')
| -rw-r--r-- | llvm/test/CodeGen/AMDGPU/fptrunc.ll | 1173 |
1 files changed, 994 insertions, 179 deletions
diff --git a/llvm/test/CodeGen/AMDGPU/fptrunc.ll b/llvm/test/CodeGen/AMDGPU/fptrunc.ll index 2bd3659..4f8eab1 100644 --- a/llvm/test/CodeGen/AMDGPU/fptrunc.ll +++ b/llvm/test/CodeGen/AMDGPU/fptrunc.ll @@ -3,17 +3,15 @@ ; RUN: llc -mtriple=amdgcn -mcpu=tonga -global-isel=0 -mattr=-flat-for-global < %s | FileCheck -check-prefixes=VI-SDAG,VI-SAFE-SDAG %s ; RUN: llc -mtriple=amdgcn -mcpu=tonga -global-isel=1 -mattr=-flat-for-global < %s | FileCheck -check-prefixes=VI-GISEL,VI-SAFE-GISEL %s ; RUN: llc -mtriple=amdgcn -mcpu=tonga -global-isel=0 -mattr=-flat-for-global -enable-unsafe-fp-math < %s | FileCheck -check-prefixes=VI-SDAG,VI-UNSAFE-SDAG %s -; RUN: llc -mtriple=amdgcn -mcpu=tonga -global-isel=1 -mattr=-flat-for-global -enable-unsafe-fp-math < %s | FileCheck -check-prefixes=VI-GISEL,VI-UNSAFE-GISEL %s ; RUN: llc -mtriple=amdgcn -mcpu=gfx1030 -global-isel=0 -mattr=-flat-for-global < %s | FileCheck -check-prefixes=GFX10-SDAG,GFX10-SAFE-SDAG %s ; RUN: llc -mtriple=amdgcn -mcpu=gfx1030 -global-isel=1 -mattr=-flat-for-global < %s | FileCheck -check-prefixes=GFX10-GISEL,GFX10-SAFE-GISEL %s ; RUN: llc -mtriple=amdgcn -mcpu=gfx1030 -global-isel=0 -mattr=-flat-for-global -enable-unsafe-fp-math < %s | FileCheck -check-prefixes=GFX10-SDAG,GFX10-UNSAFE-SDAG %s -; RUN: llc -mtriple=amdgcn -mcpu=gfx1030 -global-isel=1 -mattr=-flat-for-global -enable-unsafe-fp-math < %s | FileCheck -check-prefixes=GFX10-GISEL,GFX10-UNSAFE-GISEL %s ; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -global-isel=0 -mattr=-flat-for-global < %s | FileCheck -check-prefixes=GFX11-SDAG,GFX11-SAFE-SDAG %s ; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -global-isel=1 -mattr=-flat-for-global < %s | FileCheck -check-prefixes=GFX11-GISEL,GFX11-SAFE-GISEL %s ; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -global-isel=0 -mattr=-flat-for-global,+real-true16 -enable-unsafe-fp-math < %s | FileCheck -check-prefixes=GFX11-SDAG,GFX11-UNSAFE-DAG-TRUE16 %s ; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -global-isel=0 -mattr=-flat-for-global,-real-true16 -enable-unsafe-fp-math < %s | FileCheck -check-prefixes=GFX11-SDAG,GFX11-UNSAFE-DAG-FAKE16 %s -; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -global-isel=1 -mattr=-flat-for-global,+real-true16 -enable-unsafe-fp-math < %s | FileCheck -check-prefixes=GFX11-GISEL,GFX11-UNSAFE-GISEL-TRUE16 %s -; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -global-isel=1 -mattr=-flat-for-global,-real-true16 -enable-unsafe-fp-math < %s | FileCheck -check-prefixes=GFX11-GISEL,GFX11-UNSAFE-GISEL-FAKE16 %s +; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -global-isel=1 -mattr=-flat-for-global,+real-true16 < %s | FileCheck -check-prefixes=GFX11-GISEL,GFX11-UNSAFE-GISEL-TRUE16 %s +; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -global-isel=1 -mattr=-flat-for-global,-real-true16 < %s | FileCheck -check-prefixes=GFX11-GISEL,GFX11-UNSAFE-GISEL-FAKE16 %s define amdgpu_kernel void @fptrunc_f64_to_f32(ptr addrspace(1) %out, double %in) { ; SI-LABEL: fptrunc_f64_to_f32: @@ -94,6 +92,85 @@ define amdgpu_kernel void @fptrunc_f64_to_f32(ptr addrspace(1) %out, double %in) ret void } +define amdgpu_kernel void @fptrunc_f64_to_f32_afn(ptr addrspace(1) %out, double %in) { +; SI-LABEL: fptrunc_f64_to_f32_afn: +; SI: ; %bb.0: +; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 +; SI-NEXT: s_mov_b32 s7, 0xf000 +; SI-NEXT: s_mov_b32 s6, -1 +; SI-NEXT: s_waitcnt lgkmcnt(0) +; SI-NEXT: s_mov_b32 s4, s0 +; SI-NEXT: s_mov_b32 s5, s1 +; SI-NEXT: v_cvt_f32_f64_e32 v0, s[2:3] +; SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; SI-NEXT: s_endpgm +; +; VI-SDAG-LABEL: fptrunc_f64_to_f32_afn: +; VI-SDAG: ; %bb.0: +; VI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 +; VI-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; VI-SDAG-NEXT: s_mov_b32 s6, -1 +; VI-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[2:3] +; VI-SDAG-NEXT: s_mov_b32 s4, s0 +; VI-SDAG-NEXT: s_mov_b32 s5, s1 +; VI-SDAG-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; VI-SDAG-NEXT: s_endpgm +; +; VI-GISEL-LABEL: fptrunc_f64_to_f32_afn: +; VI-GISEL: ; %bb.0: +; VI-GISEL-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v0, s[2:3] +; VI-GISEL-NEXT: s_mov_b32 s2, -1 +; VI-GISEL-NEXT: s_mov_b32 s3, 0xf000 +; VI-GISEL-NEXT: buffer_store_dword v0, off, s[0:3], 0 +; VI-GISEL-NEXT: s_endpgm +; +; GFX10-SDAG-LABEL: fptrunc_f64_to_f32_afn: +; GFX10-SDAG: ; %bb.0: +; GFX10-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 +; GFX10-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[2:3] +; GFX10-SDAG-NEXT: s_mov_b32 s3, 0x31016000 +; GFX10-SDAG-NEXT: s_mov_b32 s2, -1 +; GFX10-SDAG-NEXT: buffer_store_dword v0, off, s[0:3], 0 +; GFX10-SDAG-NEXT: s_endpgm +; +; GFX10-GISEL-LABEL: fptrunc_f64_to_f32_afn: +; GFX10-GISEL: ; %bb.0: +; GFX10-GISEL-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 +; GFX10-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v0, s[2:3] +; GFX10-GISEL-NEXT: s_mov_b32 s2, -1 +; GFX10-GISEL-NEXT: s_mov_b32 s3, 0x31016000 +; GFX10-GISEL-NEXT: buffer_store_dword v0, off, s[0:3], 0 +; GFX10-GISEL-NEXT: s_endpgm +; +; GFX11-SDAG-LABEL: fptrunc_f64_to_f32_afn: +; GFX11-SDAG: ; %bb.0: +; GFX11-SDAG-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[2:3] +; GFX11-SDAG-NEXT: s_mov_b32 s3, 0x31016000 +; GFX11-SDAG-NEXT: s_mov_b32 s2, -1 +; GFX11-SDAG-NEXT: buffer_store_b32 v0, off, s[0:3], 0 +; GFX11-SDAG-NEXT: s_endpgm +; +; GFX11-GISEL-LABEL: fptrunc_f64_to_f32_afn: +; GFX11-GISEL: ; %bb.0: +; GFX11-GISEL-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v0, s[2:3] +; GFX11-GISEL-NEXT: s_mov_b32 s2, -1 +; GFX11-GISEL-NEXT: s_mov_b32 s3, 0x31016000 +; GFX11-GISEL-NEXT: buffer_store_b32 v0, off, s[0:3], 0 +; GFX11-GISEL-NEXT: s_endpgm + %result = fptrunc afn double %in to float + store float %result, ptr addrspace(1) %out + ret void +} + define amdgpu_kernel void @fptrunc_f64_to_f16(ptr addrspace(1) %out, double %in) { ; SI-LABEL: fptrunc_f64_to_f16: ; SI: ; %bb.0: @@ -203,56 +280,56 @@ define amdgpu_kernel void @fptrunc_f64_to_f16(ptr addrspace(1) %out, double %in) ; VI-SAFE-SDAG-NEXT: buffer_store_short v0, off, s[0:3], 0 ; VI-SAFE-SDAG-NEXT: s_endpgm ; -; VI-SAFE-GISEL-LABEL: fptrunc_f64_to_f16: -; VI-SAFE-GISEL: ; %bb.0: -; VI-SAFE-GISEL-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 -; VI-SAFE-GISEL-NEXT: s_waitcnt lgkmcnt(0) -; VI-SAFE-GISEL-NEXT: s_bfe_u32 s4, s3, 0xb0014 -; VI-SAFE-GISEL-NEXT: s_lshr_b32 s5, s3, 8 -; VI-SAFE-GISEL-NEXT: s_and_b32 s6, s3, 0x1ff -; VI-SAFE-GISEL-NEXT: s_addk_i32 s4, 0xfc10 -; VI-SAFE-GISEL-NEXT: s_and_b32 s5, s5, 0xffe -; VI-SAFE-GISEL-NEXT: s_or_b32 s2, s6, s2 -; VI-SAFE-GISEL-NEXT: s_cmp_lg_u32 s2, 0 -; VI-SAFE-GISEL-NEXT: s_cselect_b32 s2, 1, 0 -; VI-SAFE-GISEL-NEXT: s_or_b32 s2, s5, s2 -; VI-SAFE-GISEL-NEXT: s_cmp_lg_u32 s2, 0 -; VI-SAFE-GISEL-NEXT: s_cselect_b32 s5, 1, 0 -; VI-SAFE-GISEL-NEXT: s_sub_i32 s7, 1, s4 -; VI-SAFE-GISEL-NEXT: s_lshl_b32 s6, s4, 12 -; VI-SAFE-GISEL-NEXT: s_max_i32 s7, s7, 0 -; VI-SAFE-GISEL-NEXT: s_or_b32 s6, s2, s6 -; VI-SAFE-GISEL-NEXT: s_min_i32 s7, s7, 13 -; VI-SAFE-GISEL-NEXT: s_bitset1_b32 s2, 12 -; VI-SAFE-GISEL-NEXT: s_lshl_b32 s5, s5, 9 -; VI-SAFE-GISEL-NEXT: s_lshr_b32 s8, s2, s7 -; VI-SAFE-GISEL-NEXT: s_or_b32 s5, s5, 0x7c00 -; VI-SAFE-GISEL-NEXT: s_lshl_b32 s7, s8, s7 -; VI-SAFE-GISEL-NEXT: s_cmp_lg_u32 s7, s2 -; VI-SAFE-GISEL-NEXT: s_cselect_b32 s2, 1, 0 -; VI-SAFE-GISEL-NEXT: s_or_b32 s2, s8, s2 -; VI-SAFE-GISEL-NEXT: s_cmp_lt_i32 s4, 1 -; VI-SAFE-GISEL-NEXT: s_cselect_b32 s2, s2, s6 -; VI-SAFE-GISEL-NEXT: s_and_b32 s6, s2, 7 -; VI-SAFE-GISEL-NEXT: s_lshr_b32 s2, s2, 2 -; VI-SAFE-GISEL-NEXT: s_cmp_eq_u32 s6, 3 -; VI-SAFE-GISEL-NEXT: s_cselect_b32 s7, 1, 0 -; VI-SAFE-GISEL-NEXT: s_cmp_gt_i32 s6, 5 -; VI-SAFE-GISEL-NEXT: s_cselect_b32 s6, 1, 0 -; VI-SAFE-GISEL-NEXT: s_or_b32 s6, s7, s6 -; VI-SAFE-GISEL-NEXT: s_add_i32 s2, s2, s6 -; VI-SAFE-GISEL-NEXT: s_cmp_gt_i32 s4, 30 -; VI-SAFE-GISEL-NEXT: s_cselect_b32 s2, 0x7c00, s2 -; VI-SAFE-GISEL-NEXT: s_cmpk_eq_i32 s4, 0x40f -; VI-SAFE-GISEL-NEXT: s_cselect_b32 s2, s5, s2 -; VI-SAFE-GISEL-NEXT: s_lshr_b32 s3, s3, 16 -; VI-SAFE-GISEL-NEXT: s_and_b32 s3, s3, 0x8000 -; VI-SAFE-GISEL-NEXT: s_or_b32 s2, s3, s2 -; VI-SAFE-GISEL-NEXT: v_mov_b32_e32 v0, s2 -; VI-SAFE-GISEL-NEXT: s_mov_b32 s2, -1 -; VI-SAFE-GISEL-NEXT: s_mov_b32 s3, 0xf000 -; VI-SAFE-GISEL-NEXT: buffer_store_short v0, off, s[0:3], 0 -; VI-SAFE-GISEL-NEXT: s_endpgm +; VI-GISEL-LABEL: fptrunc_f64_to_f16: +; VI-GISEL: ; %bb.0: +; VI-GISEL-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: s_bfe_u32 s4, s3, 0xb0014 +; VI-GISEL-NEXT: s_lshr_b32 s5, s3, 8 +; VI-GISEL-NEXT: s_and_b32 s6, s3, 0x1ff +; VI-GISEL-NEXT: s_addk_i32 s4, 0xfc10 +; VI-GISEL-NEXT: s_and_b32 s5, s5, 0xffe +; VI-GISEL-NEXT: s_or_b32 s2, s6, s2 +; VI-GISEL-NEXT: s_cmp_lg_u32 s2, 0 +; VI-GISEL-NEXT: s_cselect_b32 s2, 1, 0 +; VI-GISEL-NEXT: s_or_b32 s2, s5, s2 +; VI-GISEL-NEXT: s_cmp_lg_u32 s2, 0 +; VI-GISEL-NEXT: s_cselect_b32 s5, 1, 0 +; VI-GISEL-NEXT: s_sub_i32 s7, 1, s4 +; VI-GISEL-NEXT: s_lshl_b32 s6, s4, 12 +; VI-GISEL-NEXT: s_max_i32 s7, s7, 0 +; VI-GISEL-NEXT: s_or_b32 s6, s2, s6 +; VI-GISEL-NEXT: s_min_i32 s7, s7, 13 +; VI-GISEL-NEXT: s_bitset1_b32 s2, 12 +; VI-GISEL-NEXT: s_lshl_b32 s5, s5, 9 +; VI-GISEL-NEXT: s_lshr_b32 s8, s2, s7 +; VI-GISEL-NEXT: s_or_b32 s5, s5, 0x7c00 +; VI-GISEL-NEXT: s_lshl_b32 s7, s8, s7 +; VI-GISEL-NEXT: s_cmp_lg_u32 s7, s2 +; VI-GISEL-NEXT: s_cselect_b32 s2, 1, 0 +; VI-GISEL-NEXT: s_or_b32 s2, s8, s2 +; VI-GISEL-NEXT: s_cmp_lt_i32 s4, 1 +; VI-GISEL-NEXT: s_cselect_b32 s2, s2, s6 +; VI-GISEL-NEXT: s_and_b32 s6, s2, 7 +; VI-GISEL-NEXT: s_lshr_b32 s2, s2, 2 +; VI-GISEL-NEXT: s_cmp_eq_u32 s6, 3 +; VI-GISEL-NEXT: s_cselect_b32 s7, 1, 0 +; VI-GISEL-NEXT: s_cmp_gt_i32 s6, 5 +; VI-GISEL-NEXT: s_cselect_b32 s6, 1, 0 +; VI-GISEL-NEXT: s_or_b32 s6, s7, s6 +; VI-GISEL-NEXT: s_add_i32 s2, s2, s6 +; VI-GISEL-NEXT: s_cmp_gt_i32 s4, 30 +; VI-GISEL-NEXT: s_cselect_b32 s2, 0x7c00, s2 +; VI-GISEL-NEXT: s_cmpk_eq_i32 s4, 0x40f +; VI-GISEL-NEXT: s_cselect_b32 s2, s5, s2 +; VI-GISEL-NEXT: s_lshr_b32 s3, s3, 16 +; VI-GISEL-NEXT: s_and_b32 s3, s3, 0x8000 +; VI-GISEL-NEXT: s_or_b32 s2, s3, s2 +; VI-GISEL-NEXT: v_mov_b32_e32 v0, s2 +; VI-GISEL-NEXT: s_mov_b32 s2, -1 +; VI-GISEL-NEXT: s_mov_b32 s3, 0xf000 +; VI-GISEL-NEXT: buffer_store_short v0, off, s[0:3], 0 +; VI-GISEL-NEXT: s_endpgm ; ; VI-UNSAFE-SDAG-LABEL: fptrunc_f64_to_f16: ; VI-UNSAFE-SDAG: ; %bb.0: @@ -265,17 +342,6 @@ define amdgpu_kernel void @fptrunc_f64_to_f16(ptr addrspace(1) %out, double %in) ; VI-UNSAFE-SDAG-NEXT: buffer_store_short v0, off, s[0:3], 0 ; VI-UNSAFE-SDAG-NEXT: s_endpgm ; -; VI-UNSAFE-GISEL-LABEL: fptrunc_f64_to_f16: -; VI-UNSAFE-GISEL: ; %bb.0: -; VI-UNSAFE-GISEL-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 -; VI-UNSAFE-GISEL-NEXT: s_waitcnt lgkmcnt(0) -; VI-UNSAFE-GISEL-NEXT: v_cvt_f32_f64_e32 v0, s[2:3] -; VI-UNSAFE-GISEL-NEXT: s_mov_b32 s2, -1 -; VI-UNSAFE-GISEL-NEXT: s_mov_b32 s3, 0xf000 -; VI-UNSAFE-GISEL-NEXT: v_cvt_f16_f32_e32 v0, v0 -; VI-UNSAFE-GISEL-NEXT: buffer_store_short v0, off, s[0:3], 0 -; VI-UNSAFE-GISEL-NEXT: s_endpgm -; ; GFX10-SAFE-SDAG-LABEL: fptrunc_f64_to_f16: ; GFX10-SAFE-SDAG: ; %bb.0: ; GFX10-SAFE-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 @@ -328,56 +394,56 @@ define amdgpu_kernel void @fptrunc_f64_to_f16(ptr addrspace(1) %out, double %in) ; GFX10-SAFE-SDAG-NEXT: buffer_store_short v0, off, s[0:3], 0 ; GFX10-SAFE-SDAG-NEXT: s_endpgm ; -; GFX10-SAFE-GISEL-LABEL: fptrunc_f64_to_f16: -; GFX10-SAFE-GISEL: ; %bb.0: -; GFX10-SAFE-GISEL-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 -; GFX10-SAFE-GISEL-NEXT: s_waitcnt lgkmcnt(0) -; GFX10-SAFE-GISEL-NEXT: s_and_b32 s6, s3, 0x1ff -; GFX10-SAFE-GISEL-NEXT: s_bfe_u32 s4, s3, 0xb0014 -; GFX10-SAFE-GISEL-NEXT: s_lshr_b32 s5, s3, 8 -; GFX10-SAFE-GISEL-NEXT: s_or_b32 s2, s6, s2 -; GFX10-SAFE-GISEL-NEXT: s_addk_i32 s4, 0xfc10 -; GFX10-SAFE-GISEL-NEXT: s_and_b32 s5, s5, 0xffe -; GFX10-SAFE-GISEL-NEXT: s_cmp_lg_u32 s2, 0 -; GFX10-SAFE-GISEL-NEXT: s_cselect_b32 s2, 1, 0 -; GFX10-SAFE-GISEL-NEXT: s_or_b32 s2, s5, s2 -; GFX10-SAFE-GISEL-NEXT: s_cmp_lg_u32 s2, 0 -; GFX10-SAFE-GISEL-NEXT: s_cselect_b32 s5, 1, 0 -; GFX10-SAFE-GISEL-NEXT: s_sub_i32 s6, 1, s4 -; GFX10-SAFE-GISEL-NEXT: s_or_b32 s8, s2, 0x1000 -; GFX10-SAFE-GISEL-NEXT: s_max_i32 s6, s6, 0 -; GFX10-SAFE-GISEL-NEXT: s_lshl_b32 s7, s4, 12 -; GFX10-SAFE-GISEL-NEXT: s_min_i32 s6, s6, 13 -; GFX10-SAFE-GISEL-NEXT: s_lshl_b32 s5, s5, 9 -; GFX10-SAFE-GISEL-NEXT: s_lshr_b32 s9, s8, s6 -; GFX10-SAFE-GISEL-NEXT: s_or_b32 s2, s2, s7 -; GFX10-SAFE-GISEL-NEXT: s_lshl_b32 s6, s9, s6 -; GFX10-SAFE-GISEL-NEXT: s_or_b32 s5, s5, 0x7c00 -; GFX10-SAFE-GISEL-NEXT: s_cmp_lg_u32 s6, s8 -; GFX10-SAFE-GISEL-NEXT: s_cselect_b32 s6, 1, 0 -; GFX10-SAFE-GISEL-NEXT: s_or_b32 s6, s9, s6 -; GFX10-SAFE-GISEL-NEXT: s_cmp_lt_i32 s4, 1 -; GFX10-SAFE-GISEL-NEXT: s_cselect_b32 s2, s6, s2 -; GFX10-SAFE-GISEL-NEXT: s_and_b32 s6, s2, 7 -; GFX10-SAFE-GISEL-NEXT: s_lshr_b32 s2, s2, 2 -; GFX10-SAFE-GISEL-NEXT: s_cmp_eq_u32 s6, 3 -; GFX10-SAFE-GISEL-NEXT: s_cselect_b32 s7, 1, 0 -; GFX10-SAFE-GISEL-NEXT: s_cmp_gt_i32 s6, 5 -; GFX10-SAFE-GISEL-NEXT: s_cselect_b32 s6, 1, 0 -; GFX10-SAFE-GISEL-NEXT: s_or_b32 s6, s7, s6 -; GFX10-SAFE-GISEL-NEXT: s_add_i32 s2, s2, s6 -; GFX10-SAFE-GISEL-NEXT: s_cmp_gt_i32 s4, 30 -; GFX10-SAFE-GISEL-NEXT: s_cselect_b32 s2, 0x7c00, s2 -; GFX10-SAFE-GISEL-NEXT: s_cmpk_eq_i32 s4, 0x40f -; GFX10-SAFE-GISEL-NEXT: s_cselect_b32 s2, s5, s2 -; GFX10-SAFE-GISEL-NEXT: s_lshr_b32 s3, s3, 16 -; GFX10-SAFE-GISEL-NEXT: s_and_b32 s3, s3, 0x8000 -; GFX10-SAFE-GISEL-NEXT: s_or_b32 s2, s3, s2 -; GFX10-SAFE-GISEL-NEXT: s_mov_b32 s3, 0x31016000 -; GFX10-SAFE-GISEL-NEXT: v_mov_b32_e32 v0, s2 -; GFX10-SAFE-GISEL-NEXT: s_mov_b32 s2, -1 -; GFX10-SAFE-GISEL-NEXT: buffer_store_short v0, off, s[0:3], 0 -; GFX10-SAFE-GISEL-NEXT: s_endpgm +; GFX10-GISEL-LABEL: fptrunc_f64_to_f16: +; GFX10-GISEL: ; %bb.0: +; GFX10-GISEL-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 +; GFX10-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX10-GISEL-NEXT: s_and_b32 s6, s3, 0x1ff +; GFX10-GISEL-NEXT: s_bfe_u32 s4, s3, 0xb0014 +; GFX10-GISEL-NEXT: s_lshr_b32 s5, s3, 8 +; GFX10-GISEL-NEXT: s_or_b32 s2, s6, s2 +; GFX10-GISEL-NEXT: s_addk_i32 s4, 0xfc10 +; GFX10-GISEL-NEXT: s_and_b32 s5, s5, 0xffe +; GFX10-GISEL-NEXT: s_cmp_lg_u32 s2, 0 +; GFX10-GISEL-NEXT: s_cselect_b32 s2, 1, 0 +; GFX10-GISEL-NEXT: s_or_b32 s2, s5, s2 +; GFX10-GISEL-NEXT: s_cmp_lg_u32 s2, 0 +; GFX10-GISEL-NEXT: s_cselect_b32 s5, 1, 0 +; GFX10-GISEL-NEXT: s_sub_i32 s6, 1, s4 +; GFX10-GISEL-NEXT: s_or_b32 s8, s2, 0x1000 +; GFX10-GISEL-NEXT: s_max_i32 s6, s6, 0 +; GFX10-GISEL-NEXT: s_lshl_b32 s7, s4, 12 +; GFX10-GISEL-NEXT: s_min_i32 s6, s6, 13 +; GFX10-GISEL-NEXT: s_lshl_b32 s5, s5, 9 +; GFX10-GISEL-NEXT: s_lshr_b32 s9, s8, s6 +; GFX10-GISEL-NEXT: s_or_b32 s2, s2, s7 +; GFX10-GISEL-NEXT: s_lshl_b32 s6, s9, s6 +; GFX10-GISEL-NEXT: s_or_b32 s5, s5, 0x7c00 +; GFX10-GISEL-NEXT: s_cmp_lg_u32 s6, s8 +; GFX10-GISEL-NEXT: s_cselect_b32 s6, 1, 0 +; GFX10-GISEL-NEXT: s_or_b32 s6, s9, s6 +; GFX10-GISEL-NEXT: s_cmp_lt_i32 s4, 1 +; GFX10-GISEL-NEXT: s_cselect_b32 s2, s6, s2 +; GFX10-GISEL-NEXT: s_and_b32 s6, s2, 7 +; GFX10-GISEL-NEXT: s_lshr_b32 s2, s2, 2 +; GFX10-GISEL-NEXT: s_cmp_eq_u32 s6, 3 +; GFX10-GISEL-NEXT: s_cselect_b32 s7, 1, 0 +; GFX10-GISEL-NEXT: s_cmp_gt_i32 s6, 5 +; GFX10-GISEL-NEXT: s_cselect_b32 s6, 1, 0 +; GFX10-GISEL-NEXT: s_or_b32 s6, s7, s6 +; GFX10-GISEL-NEXT: s_add_i32 s2, s2, s6 +; GFX10-GISEL-NEXT: s_cmp_gt_i32 s4, 30 +; GFX10-GISEL-NEXT: s_cselect_b32 s2, 0x7c00, s2 +; GFX10-GISEL-NEXT: s_cmpk_eq_i32 s4, 0x40f +; GFX10-GISEL-NEXT: s_cselect_b32 s2, s5, s2 +; GFX10-GISEL-NEXT: s_lshr_b32 s3, s3, 16 +; GFX10-GISEL-NEXT: s_and_b32 s3, s3, 0x8000 +; GFX10-GISEL-NEXT: s_or_b32 s2, s3, s2 +; GFX10-GISEL-NEXT: s_mov_b32 s3, 0x31016000 +; GFX10-GISEL-NEXT: v_mov_b32_e32 v0, s2 +; GFX10-GISEL-NEXT: s_mov_b32 s2, -1 +; GFX10-GISEL-NEXT: buffer_store_short v0, off, s[0:3], 0 +; GFX10-GISEL-NEXT: s_endpgm ; ; GFX10-UNSAFE-SDAG-LABEL: fptrunc_f64_to_f16: ; GFX10-UNSAFE-SDAG: ; %bb.0: @@ -390,17 +456,6 @@ define amdgpu_kernel void @fptrunc_f64_to_f16(ptr addrspace(1) %out, double %in) ; GFX10-UNSAFE-SDAG-NEXT: buffer_store_short v0, off, s[0:3], 0 ; GFX10-UNSAFE-SDAG-NEXT: s_endpgm ; -; GFX10-UNSAFE-GISEL-LABEL: fptrunc_f64_to_f16: -; GFX10-UNSAFE-GISEL: ; %bb.0: -; GFX10-UNSAFE-GISEL-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 -; GFX10-UNSAFE-GISEL-NEXT: s_waitcnt lgkmcnt(0) -; GFX10-UNSAFE-GISEL-NEXT: v_cvt_f32_f64_e32 v0, s[2:3] -; GFX10-UNSAFE-GISEL-NEXT: s_mov_b32 s2, -1 -; GFX10-UNSAFE-GISEL-NEXT: s_mov_b32 s3, 0x31016000 -; GFX10-UNSAFE-GISEL-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX10-UNSAFE-GISEL-NEXT: buffer_store_short v0, off, s[0:3], 0 -; GFX10-UNSAFE-GISEL-NEXT: s_endpgm -; ; GFX11-SAFE-SDAG-LABEL: fptrunc_f64_to_f16: ; GFX11-SAFE-SDAG: ; %bb.0: ; GFX11-SAFE-SDAG-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 @@ -461,62 +516,368 @@ define amdgpu_kernel void @fptrunc_f64_to_f16(ptr addrspace(1) %out, double %in) ; GFX11-SAFE-SDAG-NEXT: buffer_store_b16 v0, off, s[0:3], 0 ; GFX11-SAFE-SDAG-NEXT: s_endpgm ; -; GFX11-SAFE-GISEL-LABEL: fptrunc_f64_to_f16: +; GFX11-GISEL-LABEL: fptrunc_f64_to_f16: +; GFX11-GISEL: ; %bb.0: +; GFX11-GISEL-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-GISEL-NEXT: s_and_b32 s6, s3, 0x1ff +; GFX11-GISEL-NEXT: s_bfe_u32 s4, s3, 0xb0014 +; GFX11-GISEL-NEXT: s_lshr_b32 s5, s3, 8 +; GFX11-GISEL-NEXT: s_or_b32 s2, s6, s2 +; GFX11-GISEL-NEXT: s_addk_i32 s4, 0xfc10 +; GFX11-GISEL-NEXT: s_and_b32 s5, s5, 0xffe +; GFX11-GISEL-NEXT: s_cmp_lg_u32 s2, 0 +; GFX11-GISEL-NEXT: s_cselect_b32 s2, 1, 0 +; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) +; GFX11-GISEL-NEXT: s_or_b32 s2, s5, s2 +; GFX11-GISEL-NEXT: s_cmp_lg_u32 s2, 0 +; GFX11-GISEL-NEXT: s_cselect_b32 s5, 1, 0 +; GFX11-GISEL-NEXT: s_sub_i32 s6, 1, s4 +; GFX11-GISEL-NEXT: s_or_b32 s8, s2, 0x1000 +; GFX11-GISEL-NEXT: s_max_i32 s6, s6, 0 +; GFX11-GISEL-NEXT: s_lshl_b32 s7, s4, 12 +; GFX11-GISEL-NEXT: s_min_i32 s6, s6, 13 +; GFX11-GISEL-NEXT: s_lshl_b32 s5, s5, 9 +; GFX11-GISEL-NEXT: s_lshr_b32 s9, s8, s6 +; GFX11-GISEL-NEXT: s_or_b32 s2, s2, s7 +; GFX11-GISEL-NEXT: s_lshl_b32 s6, s9, s6 +; GFX11-GISEL-NEXT: s_or_b32 s5, s5, 0x7c00 +; GFX11-GISEL-NEXT: s_cmp_lg_u32 s6, s8 +; GFX11-GISEL-NEXT: s_cselect_b32 s6, 1, 0 +; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1) +; GFX11-GISEL-NEXT: s_or_b32 s6, s9, s6 +; GFX11-GISEL-NEXT: s_cmp_lt_i32 s4, 1 +; GFX11-GISEL-NEXT: s_cselect_b32 s2, s6, s2 +; GFX11-GISEL-NEXT: s_and_b32 s6, s2, 7 +; GFX11-GISEL-NEXT: s_lshr_b32 s2, s2, 2 +; GFX11-GISEL-NEXT: s_cmp_eq_u32 s6, 3 +; GFX11-GISEL-NEXT: s_cselect_b32 s7, 1, 0 +; GFX11-GISEL-NEXT: s_cmp_gt_i32 s6, 5 +; GFX11-GISEL-NEXT: s_cselect_b32 s6, 1, 0 +; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) +; GFX11-GISEL-NEXT: s_or_b32 s6, s7, s6 +; GFX11-GISEL-NEXT: s_add_i32 s2, s2, s6 +; GFX11-GISEL-NEXT: s_cmp_gt_i32 s4, 30 +; GFX11-GISEL-NEXT: s_cselect_b32 s2, 0x7c00, s2 +; GFX11-GISEL-NEXT: s_cmpk_eq_i32 s4, 0x40f +; GFX11-GISEL-NEXT: s_cselect_b32 s2, s5, s2 +; GFX11-GISEL-NEXT: s_lshr_b32 s3, s3, 16 +; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) +; GFX11-GISEL-NEXT: s_and_b32 s3, s3, 0x8000 +; GFX11-GISEL-NEXT: s_or_b32 s2, s3, s2 +; GFX11-GISEL-NEXT: s_mov_b32 s3, 0x31016000 +; GFX11-GISEL-NEXT: v_mov_b32_e32 v0, s2 +; GFX11-GISEL-NEXT: s_mov_b32 s2, -1 +; GFX11-GISEL-NEXT: buffer_store_b16 v0, off, s[0:3], 0 +; GFX11-GISEL-NEXT: s_endpgm +; +; GFX11-UNSAFE-DAG-TRUE16-LABEL: fptrunc_f64_to_f16: +; GFX11-UNSAFE-DAG-TRUE16: ; %bb.0: +; GFX11-UNSAFE-DAG-TRUE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-UNSAFE-DAG-TRUE16-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-UNSAFE-DAG-TRUE16-NEXT: v_cvt_f32_f64_e32 v0, s[2:3] +; GFX11-UNSAFE-DAG-TRUE16-NEXT: s_mov_b32 s3, 0x31016000 +; GFX11-UNSAFE-DAG-TRUE16-NEXT: s_mov_b32 s2, -1 +; GFX11-UNSAFE-DAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-UNSAFE-DAG-TRUE16-NEXT: v_cvt_f16_f32_e32 v0.l, v0 +; GFX11-UNSAFE-DAG-TRUE16-NEXT: buffer_store_b16 v0, off, s[0:3], 0 +; GFX11-UNSAFE-DAG-TRUE16-NEXT: s_endpgm +; +; GFX11-UNSAFE-DAG-FAKE16-LABEL: fptrunc_f64_to_f16: +; GFX11-UNSAFE-DAG-FAKE16: ; %bb.0: +; GFX11-UNSAFE-DAG-FAKE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-UNSAFE-DAG-FAKE16-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-UNSAFE-DAG-FAKE16-NEXT: v_cvt_f32_f64_e32 v0, s[2:3] +; GFX11-UNSAFE-DAG-FAKE16-NEXT: s_mov_b32 s3, 0x31016000 +; GFX11-UNSAFE-DAG-FAKE16-NEXT: s_mov_b32 s2, -1 +; GFX11-UNSAFE-DAG-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-UNSAFE-DAG-FAKE16-NEXT: v_cvt_f16_f32_e32 v0, v0 +; GFX11-UNSAFE-DAG-FAKE16-NEXT: buffer_store_b16 v0, off, s[0:3], 0 +; GFX11-UNSAFE-DAG-FAKE16-NEXT: s_endpgm + %result = fptrunc double %in to half + %result_i16 = bitcast half %result to i16 + store i16 %result_i16, ptr addrspace(1) %out + ret void +} + +define amdgpu_kernel void @fptrunc_f64_to_f16_afn(ptr addrspace(1) %out, double %in) { +; SI-LABEL: fptrunc_f64_to_f16_afn: +; SI: ; %bb.0: +; SI-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x9 +; SI-NEXT: s_mov_b32 s3, 0xf000 +; SI-NEXT: s_movk_i32 s2, 0x7e00 +; SI-NEXT: s_waitcnt lgkmcnt(0) +; SI-NEXT: s_lshr_b32 s0, s7, 8 +; SI-NEXT: s_and_b32 s1, s7, 0x1ff +; SI-NEXT: s_and_b32 s8, s0, 0xffe +; SI-NEXT: s_or_b32 s0, s1, s6 +; SI-NEXT: s_cmp_lg_u32 s0, 0 +; SI-NEXT: s_cselect_b64 s[0:1], -1, 0 +; SI-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; SI-NEXT: s_bfe_u32 s0, s7, 0xb0014 +; SI-NEXT: v_readfirstlane_b32 s1, v0 +; SI-NEXT: s_sub_i32 s6, 0x3f1, s0 +; SI-NEXT: s_or_b32 s1, s8, s1 +; SI-NEXT: v_med3_i32 v0, s6, 0, 13 +; SI-NEXT: s_or_b32 s6, s1, 0x1000 +; SI-NEXT: v_readfirstlane_b32 s8, v0 +; SI-NEXT: s_lshr_b32 s9, s6, s8 +; SI-NEXT: s_lshl_b32 s8, s9, s8 +; SI-NEXT: s_cmp_lg_u32 s8, s6 +; SI-NEXT: s_cselect_b32 s6, 1, 0 +; SI-NEXT: s_addk_i32 s0, 0xfc10 +; SI-NEXT: s_or_b32 s6, s9, s6 +; SI-NEXT: s_lshl_b32 s8, s0, 12 +; SI-NEXT: s_or_b32 s8, s1, s8 +; SI-NEXT: s_cmp_lt_i32 s0, 1 +; SI-NEXT: s_cselect_b32 s6, s6, s8 +; SI-NEXT: s_and_b32 s8, s6, 7 +; SI-NEXT: s_cmp_gt_i32 s8, 5 +; SI-NEXT: s_cselect_b32 s9, 1, 0 +; SI-NEXT: s_cmp_eq_u32 s8, 3 +; SI-NEXT: s_cselect_b32 s8, 1, 0 +; SI-NEXT: s_lshr_b32 s6, s6, 2 +; SI-NEXT: s_or_b32 s8, s8, s9 +; SI-NEXT: s_add_i32 s6, s6, s8 +; SI-NEXT: s_cmp_lt_i32 s0, 31 +; SI-NEXT: s_cselect_b32 s6, s6, 0x7c00 +; SI-NEXT: s_cmp_lg_u32 s1, 0 +; SI-NEXT: s_cselect_b32 s1, s2, 0x7c00 +; SI-NEXT: s_cmpk_eq_i32 s0, 0x40f +; SI-NEXT: s_cselect_b32 s0, s1, s6 +; SI-NEXT: s_lshr_b32 s1, s7, 16 +; SI-NEXT: s_and_b32 s1, s1, 0x8000 +; SI-NEXT: s_or_b32 s6, s1, s0 +; SI-NEXT: s_mov_b32 s2, -1 +; SI-NEXT: s_mov_b32 s0, s4 +; SI-NEXT: s_mov_b32 s1, s5 +; SI-NEXT: v_mov_b32_e32 v0, s6 +; SI-NEXT: buffer_store_short v0, off, s[0:3], 0 +; SI-NEXT: s_endpgm +; +; VI-SAFE-SDAG-LABEL: fptrunc_f64_to_f16_afn: +; VI-SAFE-SDAG: ; %bb.0: +; VI-SAFE-SDAG-NEXT: s_load_dwordx4 s[4:7], s[4:5], 0x24 +; VI-SAFE-SDAG-NEXT: s_mov_b32 s3, 0xf000 +; VI-SAFE-SDAG-NEXT: s_mov_b32 s2, -1 +; VI-SAFE-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; VI-SAFE-SDAG-NEXT: s_mov_b32 s0, s4 +; VI-SAFE-SDAG-NEXT: s_lshr_b32 s4, s7, 8 +; VI-SAFE-SDAG-NEXT: s_and_b32 s8, s4, 0xffe +; VI-SAFE-SDAG-NEXT: s_and_b32 s4, s7, 0x1ff +; VI-SAFE-SDAG-NEXT: s_or_b32 s4, s4, s6 +; VI-SAFE-SDAG-NEXT: s_cmp_lg_u32 s4, 0 +; VI-SAFE-SDAG-NEXT: s_mov_b32 s1, s5 +; VI-SAFE-SDAG-NEXT: s_cselect_b64 s[4:5], -1, 0 +; VI-SAFE-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[4:5] +; VI-SAFE-SDAG-NEXT: v_readfirstlane_b32 s4, v0 +; VI-SAFE-SDAG-NEXT: s_bfe_u32 s6, s7, 0xb0014 +; VI-SAFE-SDAG-NEXT: s_or_b32 s4, s8, s4 +; VI-SAFE-SDAG-NEXT: s_sub_i32 s8, 0x3f1, s6 +; VI-SAFE-SDAG-NEXT: v_med3_i32 v0, s8, 0, 13 +; VI-SAFE-SDAG-NEXT: s_or_b32 s5, s4, 0x1000 +; VI-SAFE-SDAG-NEXT: v_readfirstlane_b32 s8, v0 +; VI-SAFE-SDAG-NEXT: s_lshr_b32 s9, s5, s8 +; VI-SAFE-SDAG-NEXT: s_lshl_b32 s8, s9, s8 +; VI-SAFE-SDAG-NEXT: s_cmp_lg_u32 s8, s5 +; VI-SAFE-SDAG-NEXT: s_cselect_b32 s5, 1, 0 +; VI-SAFE-SDAG-NEXT: s_addk_i32 s6, 0xfc10 +; VI-SAFE-SDAG-NEXT: s_lshl_b32 s8, s6, 12 +; VI-SAFE-SDAG-NEXT: s_or_b32 s5, s9, s5 +; VI-SAFE-SDAG-NEXT: s_or_b32 s8, s4, s8 +; VI-SAFE-SDAG-NEXT: s_cmp_lt_i32 s6, 1 +; VI-SAFE-SDAG-NEXT: s_cselect_b32 s5, s5, s8 +; VI-SAFE-SDAG-NEXT: s_and_b32 s8, s5, 7 +; VI-SAFE-SDAG-NEXT: s_cmp_gt_i32 s8, 5 +; VI-SAFE-SDAG-NEXT: s_cselect_b32 s9, 1, 0 +; VI-SAFE-SDAG-NEXT: s_cmp_eq_u32 s8, 3 +; VI-SAFE-SDAG-NEXT: s_cselect_b32 s8, 1, 0 +; VI-SAFE-SDAG-NEXT: s_or_b32 s8, s8, s9 +; VI-SAFE-SDAG-NEXT: s_lshr_b32 s5, s5, 2 +; VI-SAFE-SDAG-NEXT: s_add_i32 s5, s5, s8 +; VI-SAFE-SDAG-NEXT: s_cmp_lt_i32 s6, 31 +; VI-SAFE-SDAG-NEXT: s_cselect_b32 s5, s5, 0x7c00 +; VI-SAFE-SDAG-NEXT: s_cmp_lg_u32 s4, 0 +; VI-SAFE-SDAG-NEXT: s_movk_i32 s4, 0x7e00 +; VI-SAFE-SDAG-NEXT: s_cselect_b32 s4, s4, 0x7c00 +; VI-SAFE-SDAG-NEXT: s_cmpk_eq_i32 s6, 0x40f +; VI-SAFE-SDAG-NEXT: s_cselect_b32 s4, s4, s5 +; VI-SAFE-SDAG-NEXT: s_lshr_b32 s5, s7, 16 +; VI-SAFE-SDAG-NEXT: s_and_b32 s5, s5, 0x8000 +; VI-SAFE-SDAG-NEXT: s_or_b32 s4, s5, s4 +; VI-SAFE-SDAG-NEXT: v_mov_b32_e32 v0, s4 +; VI-SAFE-SDAG-NEXT: buffer_store_short v0, off, s[0:3], 0 +; VI-SAFE-SDAG-NEXT: s_endpgm +; +; VI-GISEL-LABEL: fptrunc_f64_to_f16_afn: +; VI-GISEL: ; %bb.0: +; VI-GISEL-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v0, s[2:3] +; VI-GISEL-NEXT: s_mov_b32 s2, -1 +; VI-GISEL-NEXT: s_mov_b32 s3, 0xf000 +; VI-GISEL-NEXT: v_cvt_f16_f32_e32 v0, v0 +; VI-GISEL-NEXT: buffer_store_short v0, off, s[0:3], 0 +; VI-GISEL-NEXT: s_endpgm +; +; VI-UNSAFE-SDAG-LABEL: fptrunc_f64_to_f16_afn: +; VI-UNSAFE-SDAG: ; %bb.0: +; VI-UNSAFE-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 +; VI-UNSAFE-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; VI-UNSAFE-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[2:3] +; VI-UNSAFE-SDAG-NEXT: s_mov_b32 s3, 0xf000 +; VI-UNSAFE-SDAG-NEXT: s_mov_b32 s2, -1 +; VI-UNSAFE-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; VI-UNSAFE-SDAG-NEXT: buffer_store_short v0, off, s[0:3], 0 +; VI-UNSAFE-SDAG-NEXT: s_endpgm +; +; GFX10-SAFE-SDAG-LABEL: fptrunc_f64_to_f16_afn: +; GFX10-SAFE-SDAG: ; %bb.0: +; GFX10-SAFE-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 +; GFX10-SAFE-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX10-SAFE-SDAG-NEXT: s_and_b32 s4, s3, 0x1ff +; GFX10-SAFE-SDAG-NEXT: s_lshr_b32 s5, s3, 8 +; GFX10-SAFE-SDAG-NEXT: s_or_b32 s2, s4, s2 +; GFX10-SAFE-SDAG-NEXT: s_and_b32 s4, s5, 0xffe +; GFX10-SAFE-SDAG-NEXT: s_cmp_lg_u32 s2, 0 +; GFX10-SAFE-SDAG-NEXT: s_cselect_b32 s2, -1, 0 +; GFX10-SAFE-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, s2 +; GFX10-SAFE-SDAG-NEXT: s_bfe_u32 s2, s3, 0xb0014 +; GFX10-SAFE-SDAG-NEXT: s_sub_i32 s5, 0x3f1, s2 +; GFX10-SAFE-SDAG-NEXT: v_med3_i32 v1, s5, 0, 13 +; GFX10-SAFE-SDAG-NEXT: v_readfirstlane_b32 s5, v0 +; GFX10-SAFE-SDAG-NEXT: v_readfirstlane_b32 s6, v1 +; GFX10-SAFE-SDAG-NEXT: s_or_b32 s4, s4, s5 +; GFX10-SAFE-SDAG-NEXT: s_or_b32 s5, s4, 0x1000 +; GFX10-SAFE-SDAG-NEXT: s_lshr_b32 s7, s5, s6 +; GFX10-SAFE-SDAG-NEXT: s_lshl_b32 s6, s7, s6 +; GFX10-SAFE-SDAG-NEXT: s_cmp_lg_u32 s6, s5 +; GFX10-SAFE-SDAG-NEXT: s_cselect_b32 s5, 1, 0 +; GFX10-SAFE-SDAG-NEXT: s_addk_i32 s2, 0xfc10 +; GFX10-SAFE-SDAG-NEXT: s_or_b32 s5, s7, s5 +; GFX10-SAFE-SDAG-NEXT: s_lshl_b32 s6, s2, 12 +; GFX10-SAFE-SDAG-NEXT: s_or_b32 s6, s4, s6 +; GFX10-SAFE-SDAG-NEXT: s_cmp_lt_i32 s2, 1 +; GFX10-SAFE-SDAG-NEXT: s_cselect_b32 s5, s5, s6 +; GFX10-SAFE-SDAG-NEXT: s_and_b32 s6, s5, 7 +; GFX10-SAFE-SDAG-NEXT: s_cmp_gt_i32 s6, 5 +; GFX10-SAFE-SDAG-NEXT: s_cselect_b32 s7, 1, 0 +; GFX10-SAFE-SDAG-NEXT: s_cmp_eq_u32 s6, 3 +; GFX10-SAFE-SDAG-NEXT: s_cselect_b32 s6, 1, 0 +; GFX10-SAFE-SDAG-NEXT: s_lshr_b32 s5, s5, 2 +; GFX10-SAFE-SDAG-NEXT: s_or_b32 s6, s6, s7 +; GFX10-SAFE-SDAG-NEXT: s_add_i32 s5, s5, s6 +; GFX10-SAFE-SDAG-NEXT: s_cmp_lt_i32 s2, 31 +; GFX10-SAFE-SDAG-NEXT: s_movk_i32 s6, 0x7e00 +; GFX10-SAFE-SDAG-NEXT: s_cselect_b32 s5, s5, 0x7c00 +; GFX10-SAFE-SDAG-NEXT: s_cmp_lg_u32 s4, 0 +; GFX10-SAFE-SDAG-NEXT: s_cselect_b32 s4, s6, 0x7c00 +; GFX10-SAFE-SDAG-NEXT: s_cmpk_eq_i32 s2, 0x40f +; GFX10-SAFE-SDAG-NEXT: s_cselect_b32 s2, s4, s5 +; GFX10-SAFE-SDAG-NEXT: s_lshr_b32 s3, s3, 16 +; GFX10-SAFE-SDAG-NEXT: s_and_b32 s3, s3, 0x8000 +; GFX10-SAFE-SDAG-NEXT: s_or_b32 s2, s3, s2 +; GFX10-SAFE-SDAG-NEXT: s_mov_b32 s3, 0x31016000 +; GFX10-SAFE-SDAG-NEXT: v_mov_b32_e32 v0, s2 +; GFX10-SAFE-SDAG-NEXT: s_mov_b32 s2, -1 +; GFX10-SAFE-SDAG-NEXT: buffer_store_short v0, off, s[0:3], 0 +; GFX10-SAFE-SDAG-NEXT: s_endpgm +; +; GFX10-GISEL-LABEL: fptrunc_f64_to_f16_afn: +; GFX10-GISEL: ; %bb.0: +; GFX10-GISEL-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 +; GFX10-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v0, s[2:3] +; GFX10-GISEL-NEXT: s_mov_b32 s2, -1 +; GFX10-GISEL-NEXT: s_mov_b32 s3, 0x31016000 +; GFX10-GISEL-NEXT: v_cvt_f16_f32_e32 v0, v0 +; GFX10-GISEL-NEXT: buffer_store_short v0, off, s[0:3], 0 +; GFX10-GISEL-NEXT: s_endpgm +; +; GFX10-UNSAFE-SDAG-LABEL: fptrunc_f64_to_f16_afn: +; GFX10-UNSAFE-SDAG: ; %bb.0: +; GFX10-UNSAFE-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 +; GFX10-UNSAFE-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX10-UNSAFE-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[2:3] +; GFX10-UNSAFE-SDAG-NEXT: s_mov_b32 s3, 0x31016000 +; GFX10-UNSAFE-SDAG-NEXT: s_mov_b32 s2, -1 +; GFX10-UNSAFE-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; GFX10-UNSAFE-SDAG-NEXT: buffer_store_short v0, off, s[0:3], 0 +; GFX10-UNSAFE-SDAG-NEXT: s_endpgm +; +; GFX11-SAFE-SDAG-LABEL: fptrunc_f64_to_f16_afn: +; GFX11-SAFE-SDAG: ; %bb.0: +; GFX11-SAFE-SDAG-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-SAFE-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-SAFE-SDAG-NEXT: s_and_b32 s4, s3, 0x1ff +; GFX11-SAFE-SDAG-NEXT: s_lshr_b32 s5, s3, 8 +; GFX11-SAFE-SDAG-NEXT: s_or_b32 s2, s4, s2 +; GFX11-SAFE-SDAG-NEXT: s_and_b32 s4, s5, 0xffe +; GFX11-SAFE-SDAG-NEXT: s_cmp_lg_u32 s2, 0 +; GFX11-SAFE-SDAG-NEXT: s_cselect_b32 s2, -1, 0 +; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) +; GFX11-SAFE-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, s2 +; GFX11-SAFE-SDAG-NEXT: s_bfe_u32 s2, s3, 0xb0014 +; GFX11-SAFE-SDAG-NEXT: s_sub_i32 s5, 0x3f1, s2 +; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-SAFE-SDAG-NEXT: v_med3_i32 v1, s5, 0, 13 +; GFX11-SAFE-SDAG-NEXT: v_readfirstlane_b32 s5, v0 +; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) +; GFX11-SAFE-SDAG-NEXT: v_readfirstlane_b32 s6, v1 +; GFX11-SAFE-SDAG-NEXT: s_or_b32 s4, s4, s5 +; GFX11-SAFE-SDAG-NEXT: s_or_b32 s5, s4, 0x1000 +; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) +; GFX11-SAFE-SDAG-NEXT: s_lshr_b32 s7, s5, s6 +; GFX11-SAFE-SDAG-NEXT: s_lshl_b32 s6, s7, s6 +; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1) +; GFX11-SAFE-SDAG-NEXT: s_cmp_lg_u32 s6, s5 +; GFX11-SAFE-SDAG-NEXT: s_cselect_b32 s5, 1, 0 +; GFX11-SAFE-SDAG-NEXT: s_addk_i32 s2, 0xfc10 +; GFX11-SAFE-SDAG-NEXT: s_or_b32 s5, s7, s5 +; GFX11-SAFE-SDAG-NEXT: s_lshl_b32 s6, s2, 12 +; GFX11-SAFE-SDAG-NEXT: s_or_b32 s6, s4, s6 +; GFX11-SAFE-SDAG-NEXT: s_cmp_lt_i32 s2, 1 +; GFX11-SAFE-SDAG-NEXT: s_cselect_b32 s5, s5, s6 +; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) +; GFX11-SAFE-SDAG-NEXT: s_and_b32 s6, s5, 7 +; GFX11-SAFE-SDAG-NEXT: s_cmp_gt_i32 s6, 5 +; GFX11-SAFE-SDAG-NEXT: s_cselect_b32 s7, 1, 0 +; GFX11-SAFE-SDAG-NEXT: s_cmp_eq_u32 s6, 3 +; GFX11-SAFE-SDAG-NEXT: s_cselect_b32 s6, 1, 0 +; GFX11-SAFE-SDAG-NEXT: s_lshr_b32 s5, s5, 2 +; GFX11-SAFE-SDAG-NEXT: s_or_b32 s6, s6, s7 +; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) +; GFX11-SAFE-SDAG-NEXT: s_add_i32 s5, s5, s6 +; GFX11-SAFE-SDAG-NEXT: s_cmp_lt_i32 s2, 31 +; GFX11-SAFE-SDAG-NEXT: s_movk_i32 s6, 0x7e00 +; GFX11-SAFE-SDAG-NEXT: s_cselect_b32 s5, s5, 0x7c00 +; GFX11-SAFE-SDAG-NEXT: s_cmp_lg_u32 s4, 0 +; GFX11-SAFE-SDAG-NEXT: s_cselect_b32 s4, s6, 0x7c00 +; GFX11-SAFE-SDAG-NEXT: s_cmpk_eq_i32 s2, 0x40f +; GFX11-SAFE-SDAG-NEXT: s_cselect_b32 s2, s4, s5 +; GFX11-SAFE-SDAG-NEXT: s_lshr_b32 s3, s3, 16 +; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) +; GFX11-SAFE-SDAG-NEXT: s_and_b32 s3, s3, 0x8000 +; GFX11-SAFE-SDAG-NEXT: s_or_b32 s2, s3, s2 +; GFX11-SAFE-SDAG-NEXT: s_mov_b32 s3, 0x31016000 +; GFX11-SAFE-SDAG-NEXT: v_mov_b32_e32 v0, s2 +; GFX11-SAFE-SDAG-NEXT: s_mov_b32 s2, -1 +; GFX11-SAFE-SDAG-NEXT: buffer_store_b16 v0, off, s[0:3], 0 +; GFX11-SAFE-SDAG-NEXT: s_endpgm +; +; GFX11-SAFE-GISEL-LABEL: fptrunc_f64_to_f16_afn: ; GFX11-SAFE-GISEL: ; %bb.0: ; GFX11-SAFE-GISEL-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 ; GFX11-SAFE-GISEL-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-SAFE-GISEL-NEXT: s_and_b32 s6, s3, 0x1ff -; GFX11-SAFE-GISEL-NEXT: s_bfe_u32 s4, s3, 0xb0014 -; GFX11-SAFE-GISEL-NEXT: s_lshr_b32 s5, s3, 8 -; GFX11-SAFE-GISEL-NEXT: s_or_b32 s2, s6, s2 -; GFX11-SAFE-GISEL-NEXT: s_addk_i32 s4, 0xfc10 -; GFX11-SAFE-GISEL-NEXT: s_and_b32 s5, s5, 0xffe -; GFX11-SAFE-GISEL-NEXT: s_cmp_lg_u32 s2, 0 -; GFX11-SAFE-GISEL-NEXT: s_cselect_b32 s2, 1, 0 -; GFX11-SAFE-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) -; GFX11-SAFE-GISEL-NEXT: s_or_b32 s2, s5, s2 -; GFX11-SAFE-GISEL-NEXT: s_cmp_lg_u32 s2, 0 -; GFX11-SAFE-GISEL-NEXT: s_cselect_b32 s5, 1, 0 -; GFX11-SAFE-GISEL-NEXT: s_sub_i32 s6, 1, s4 -; GFX11-SAFE-GISEL-NEXT: s_or_b32 s8, s2, 0x1000 -; GFX11-SAFE-GISEL-NEXT: s_max_i32 s6, s6, 0 -; GFX11-SAFE-GISEL-NEXT: s_lshl_b32 s7, s4, 12 -; GFX11-SAFE-GISEL-NEXT: s_min_i32 s6, s6, 13 -; GFX11-SAFE-GISEL-NEXT: s_lshl_b32 s5, s5, 9 -; GFX11-SAFE-GISEL-NEXT: s_lshr_b32 s9, s8, s6 -; GFX11-SAFE-GISEL-NEXT: s_or_b32 s2, s2, s7 -; GFX11-SAFE-GISEL-NEXT: s_lshl_b32 s6, s9, s6 -; GFX11-SAFE-GISEL-NEXT: s_or_b32 s5, s5, 0x7c00 -; GFX11-SAFE-GISEL-NEXT: s_cmp_lg_u32 s6, s8 -; GFX11-SAFE-GISEL-NEXT: s_cselect_b32 s6, 1, 0 -; GFX11-SAFE-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1) -; GFX11-SAFE-GISEL-NEXT: s_or_b32 s6, s9, s6 -; GFX11-SAFE-GISEL-NEXT: s_cmp_lt_i32 s4, 1 -; GFX11-SAFE-GISEL-NEXT: s_cselect_b32 s2, s6, s2 -; GFX11-SAFE-GISEL-NEXT: s_and_b32 s6, s2, 7 -; GFX11-SAFE-GISEL-NEXT: s_lshr_b32 s2, s2, 2 -; GFX11-SAFE-GISEL-NEXT: s_cmp_eq_u32 s6, 3 -; GFX11-SAFE-GISEL-NEXT: s_cselect_b32 s7, 1, 0 -; GFX11-SAFE-GISEL-NEXT: s_cmp_gt_i32 s6, 5 -; GFX11-SAFE-GISEL-NEXT: s_cselect_b32 s6, 1, 0 -; GFX11-SAFE-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) -; GFX11-SAFE-GISEL-NEXT: s_or_b32 s6, s7, s6 -; GFX11-SAFE-GISEL-NEXT: s_add_i32 s2, s2, s6 -; GFX11-SAFE-GISEL-NEXT: s_cmp_gt_i32 s4, 30 -; GFX11-SAFE-GISEL-NEXT: s_cselect_b32 s2, 0x7c00, s2 -; GFX11-SAFE-GISEL-NEXT: s_cmpk_eq_i32 s4, 0x40f -; GFX11-SAFE-GISEL-NEXT: s_cselect_b32 s2, s5, s2 -; GFX11-SAFE-GISEL-NEXT: s_lshr_b32 s3, s3, 16 -; GFX11-SAFE-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) -; GFX11-SAFE-GISEL-NEXT: s_and_b32 s3, s3, 0x8000 -; GFX11-SAFE-GISEL-NEXT: s_or_b32 s2, s3, s2 -; GFX11-SAFE-GISEL-NEXT: s_mov_b32 s3, 0x31016000 -; GFX11-SAFE-GISEL-NEXT: v_mov_b32_e32 v0, s2 +; GFX11-SAFE-GISEL-NEXT: v_cvt_f32_f64_e32 v0, s[2:3] ; GFX11-SAFE-GISEL-NEXT: s_mov_b32 s2, -1 +; GFX11-SAFE-GISEL-NEXT: s_mov_b32 s3, 0x31016000 +; GFX11-SAFE-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-SAFE-GISEL-NEXT: v_cvt_f16_f32_e32 v0.l, v0 ; GFX11-SAFE-GISEL-NEXT: buffer_store_b16 v0, off, s[0:3], 0 ; GFX11-SAFE-GISEL-NEXT: s_endpgm ; -; GFX11-UNSAFE-DAG-TRUE16-LABEL: fptrunc_f64_to_f16: +; GFX11-UNSAFE-DAG-TRUE16-LABEL: fptrunc_f64_to_f16_afn: ; GFX11-UNSAFE-DAG-TRUE16: ; %bb.0: ; GFX11-UNSAFE-DAG-TRUE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 ; GFX11-UNSAFE-DAG-TRUE16-NEXT: s_waitcnt lgkmcnt(0) @@ -528,7 +889,7 @@ define amdgpu_kernel void @fptrunc_f64_to_f16(ptr addrspace(1) %out, double %in) ; GFX11-UNSAFE-DAG-TRUE16-NEXT: buffer_store_b16 v0, off, s[0:3], 0 ; GFX11-UNSAFE-DAG-TRUE16-NEXT: s_endpgm ; -; GFX11-UNSAFE-DAG-FAKE16-LABEL: fptrunc_f64_to_f16: +; GFX11-UNSAFE-DAG-FAKE16-LABEL: fptrunc_f64_to_f16_afn: ; GFX11-UNSAFE-DAG-FAKE16: ; %bb.0: ; GFX11-UNSAFE-DAG-FAKE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 ; GFX11-UNSAFE-DAG-FAKE16-NEXT: s_waitcnt lgkmcnt(0) @@ -540,7 +901,7 @@ define amdgpu_kernel void @fptrunc_f64_to_f16(ptr addrspace(1) %out, double %in) ; GFX11-UNSAFE-DAG-FAKE16-NEXT: buffer_store_b16 v0, off, s[0:3], 0 ; GFX11-UNSAFE-DAG-FAKE16-NEXT: s_endpgm ; -; GFX11-UNSAFE-GISEL-TRUE16-LABEL: fptrunc_f64_to_f16: +; GFX11-UNSAFE-GISEL-TRUE16-LABEL: fptrunc_f64_to_f16_afn: ; GFX11-UNSAFE-GISEL-TRUE16: ; %bb.0: ; GFX11-UNSAFE-GISEL-TRUE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 ; GFX11-UNSAFE-GISEL-TRUE16-NEXT: s_waitcnt lgkmcnt(0) @@ -552,7 +913,7 @@ define amdgpu_kernel void @fptrunc_f64_to_f16(ptr addrspace(1) %out, double %in) ; GFX11-UNSAFE-GISEL-TRUE16-NEXT: buffer_store_b16 v0, off, s[0:3], 0 ; GFX11-UNSAFE-GISEL-TRUE16-NEXT: s_endpgm ; -; GFX11-UNSAFE-GISEL-FAKE16-LABEL: fptrunc_f64_to_f16: +; GFX11-UNSAFE-GISEL-FAKE16-LABEL: fptrunc_f64_to_f16_afn: ; GFX11-UNSAFE-GISEL-FAKE16: ; %bb.0: ; GFX11-UNSAFE-GISEL-FAKE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 ; GFX11-UNSAFE-GISEL-FAKE16-NEXT: s_waitcnt lgkmcnt(0) @@ -563,7 +924,7 @@ define amdgpu_kernel void @fptrunc_f64_to_f16(ptr addrspace(1) %out, double %in) ; GFX11-UNSAFE-GISEL-FAKE16-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX11-UNSAFE-GISEL-FAKE16-NEXT: buffer_store_b16 v0, off, s[0:3], 0 ; GFX11-UNSAFE-GISEL-FAKE16-NEXT: s_endpgm - %result = fptrunc double %in to half + %result = fptrunc afn double %in to half %result_i16 = bitcast half %result to i16 store i16 %result_i16, ptr addrspace(1) %out ret void @@ -662,6 +1023,99 @@ define amdgpu_kernel void @fptrunc_v2f64_to_v2f32(ptr addrspace(1) %out, <2 x do ret void } +define amdgpu_kernel void @fptrunc_v2f64_to_v2f32_afn(ptr addrspace(1) %out, <2 x double> %in) { +; SI-LABEL: fptrunc_v2f64_to_v2f32_afn: +; SI: ; %bb.0: +; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0xd +; SI-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x9 +; SI-NEXT: s_mov_b32 s7, 0xf000 +; SI-NEXT: s_mov_b32 s6, -1 +; SI-NEXT: s_waitcnt lgkmcnt(0) +; SI-NEXT: v_cvt_f32_f64_e32 v1, s[2:3] +; SI-NEXT: v_cvt_f32_f64_e32 v0, s[0:1] +; SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 +; SI-NEXT: s_endpgm +; +; VI-SDAG-LABEL: fptrunc_v2f64_to_v2f32_afn: +; VI-SDAG: ; %bb.0: +; VI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x34 +; VI-SDAG-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x24 +; VI-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; VI-SDAG-NEXT: s_mov_b32 s6, -1 +; VI-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v1, s[2:3] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[0:1] +; VI-SDAG-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 +; VI-SDAG-NEXT: s_endpgm +; +; VI-GISEL-LABEL: fptrunc_v2f64_to_v2f32_afn: +; VI-GISEL: ; %bb.0: +; VI-GISEL-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x34 +; VI-GISEL-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x24 +; VI-GISEL-NEXT: s_mov_b32 s6, -1 +; VI-GISEL-NEXT: s_mov_b32 s7, 0xf000 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v0, s[0:1] +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v1, s[2:3] +; VI-GISEL-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 +; VI-GISEL-NEXT: s_endpgm +; +; GFX10-SDAG-LABEL: fptrunc_v2f64_to_v2f32_afn: +; GFX10-SDAG: ; %bb.0: +; GFX10-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x34 +; GFX10-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v1, s[2:3] +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[0:1] +; GFX10-SDAG-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24 +; GFX10-SDAG-NEXT: s_mov_b32 s3, 0x31016000 +; GFX10-SDAG-NEXT: s_mov_b32 s2, -1 +; GFX10-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX10-SDAG-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 +; GFX10-SDAG-NEXT: s_endpgm +; +; GFX10-GISEL-LABEL: fptrunc_v2f64_to_v2f32_afn: +; GFX10-GISEL: ; %bb.0: +; GFX10-GISEL-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x34 +; GFX10-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v0, s[0:1] +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v1, s[2:3] +; GFX10-GISEL-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24 +; GFX10-GISEL-NEXT: s_mov_b32 s2, -1 +; GFX10-GISEL-NEXT: s_mov_b32 s3, 0x31016000 +; GFX10-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX10-GISEL-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 +; GFX10-GISEL-NEXT: s_endpgm +; +; GFX11-SDAG-LABEL: fptrunc_v2f64_to_v2f32_afn: +; GFX11-SDAG: ; %bb.0: +; GFX11-SDAG-NEXT: s_load_b128 s[0:3], s[4:5], 0x34 +; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v1, s[2:3] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[0:1] +; GFX11-SDAG-NEXT: s_load_b64 s[0:1], s[4:5], 0x24 +; GFX11-SDAG-NEXT: s_mov_b32 s3, 0x31016000 +; GFX11-SDAG-NEXT: s_mov_b32 s2, -1 +; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-SDAG-NEXT: buffer_store_b64 v[0:1], off, s[0:3], 0 +; GFX11-SDAG-NEXT: s_endpgm +; +; GFX11-GISEL-LABEL: fptrunc_v2f64_to_v2f32_afn: +; GFX11-GISEL: ; %bb.0: +; GFX11-GISEL-NEXT: s_load_b128 s[0:3], s[4:5], 0x34 +; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v0, s[0:1] +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v1, s[2:3] +; GFX11-GISEL-NEXT: s_load_b64 s[0:1], s[4:5], 0x24 +; GFX11-GISEL-NEXT: s_mov_b32 s2, -1 +; GFX11-GISEL-NEXT: s_mov_b32 s3, 0x31016000 +; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-GISEL-NEXT: buffer_store_b64 v[0:1], off, s[0:3], 0 +; GFX11-GISEL-NEXT: s_endpgm + %result = fptrunc afn <2 x double> %in to <2 x float> + store <2 x float> %result, ptr addrspace(1) %out + ret void +} + define amdgpu_kernel void @fptrunc_v3f64_to_v3f32(ptr addrspace(1) %out, <3 x double> %in) { ; SI-LABEL: fptrunc_v3f64_to_v3f32: ; SI: ; %bb.0: @@ -769,6 +1223,113 @@ define amdgpu_kernel void @fptrunc_v3f64_to_v3f32(ptr addrspace(1) %out, <3 x do ret void } +define amdgpu_kernel void @fptrunc_v3f64_to_v3f32_afn(ptr addrspace(1) %out, <3 x double> %in) { +; SI-LABEL: fptrunc_v3f64_to_v3f32_afn: +; SI: ; %bb.0: +; SI-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9 +; SI-NEXT: s_load_dwordx4 s[8:11], s[4:5], 0x11 +; SI-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x15 +; SI-NEXT: s_mov_b32 s3, 0xf000 +; SI-NEXT: s_mov_b32 s2, -1 +; SI-NEXT: s_waitcnt lgkmcnt(0) +; SI-NEXT: v_cvt_f32_f64_e32 v1, s[10:11] +; SI-NEXT: v_cvt_f32_f64_e32 v0, s[8:9] +; SI-NEXT: v_cvt_f32_f64_e32 v2, s[4:5] +; SI-NEXT: buffer_store_dword v2, off, s[0:3], 0 offset:8 +; SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 +; SI-NEXT: s_endpgm +; +; VI-SDAG-LABEL: fptrunc_v3f64_to_v3f32_afn: +; VI-SDAG: ; %bb.0: +; VI-SDAG-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x54 +; VI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x44 +; VI-SDAG-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x24 +; VI-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v2, s[6:7] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v1, s[2:3] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[0:1] +; VI-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; VI-SDAG-NEXT: s_mov_b32 s6, -1 +; VI-SDAG-NEXT: buffer_store_dwordx3 v[0:2], off, s[4:7], 0 +; VI-SDAG-NEXT: s_endpgm +; +; VI-GISEL-LABEL: fptrunc_v3f64_to_v3f32_afn: +; VI-GISEL: ; %bb.0: +; VI-GISEL-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x44 +; VI-GISEL-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24 +; VI-GISEL-NEXT: s_mov_b32 s2, -1 +; VI-GISEL-NEXT: s_mov_b32 s3, 0xf000 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v0, s[8:9] +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v1, s[10:11] +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v2, s[12:13] +; VI-GISEL-NEXT: buffer_store_dwordx3 v[0:2], off, s[0:3], 0 +; VI-GISEL-NEXT: s_endpgm +; +; GFX10-SDAG-LABEL: fptrunc_v3f64_to_v3f32_afn: +; GFX10-SDAG: ; %bb.0: +; GFX10-SDAG-NEXT: s_clause 0x1 +; GFX10-SDAG-NEXT: s_load_dwordx2 s[6:7], s[4:5], 0x54 +; GFX10-SDAG-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x44 +; GFX10-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v2, s[6:7] +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v1, s[2:3] +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[0:1] +; GFX10-SDAG-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24 +; GFX10-SDAG-NEXT: s_mov_b32 s3, 0x31016000 +; GFX10-SDAG-NEXT: s_mov_b32 s2, -1 +; GFX10-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX10-SDAG-NEXT: buffer_store_dwordx3 v[0:2], off, s[0:3], 0 +; GFX10-SDAG-NEXT: s_endpgm +; +; GFX10-GISEL-LABEL: fptrunc_v3f64_to_v3f32_afn: +; GFX10-GISEL: ; %bb.0: +; GFX10-GISEL-NEXT: s_clause 0x1 +; GFX10-GISEL-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x44 +; GFX10-GISEL-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24 +; GFX10-GISEL-NEXT: s_mov_b32 s2, -1 +; GFX10-GISEL-NEXT: s_mov_b32 s3, 0x31016000 +; GFX10-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v0, s[8:9] +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v1, s[10:11] +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v2, s[12:13] +; GFX10-GISEL-NEXT: buffer_store_dwordx3 v[0:2], off, s[0:3], 0 +; GFX10-GISEL-NEXT: s_endpgm +; +; GFX11-SDAG-LABEL: fptrunc_v3f64_to_v3f32_afn: +; GFX11-SDAG: ; %bb.0: +; GFX11-SDAG-NEXT: s_clause 0x1 +; GFX11-SDAG-NEXT: s_load_b64 s[6:7], s[4:5], 0x54 +; GFX11-SDAG-NEXT: s_load_b128 s[0:3], s[4:5], 0x44 +; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v2, s[6:7] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v1, s[2:3] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[0:1] +; GFX11-SDAG-NEXT: s_load_b64 s[0:1], s[4:5], 0x24 +; GFX11-SDAG-NEXT: s_mov_b32 s3, 0x31016000 +; GFX11-SDAG-NEXT: s_mov_b32 s2, -1 +; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-SDAG-NEXT: buffer_store_b96 v[0:2], off, s[0:3], 0 +; GFX11-SDAG-NEXT: s_endpgm +; +; GFX11-GISEL-LABEL: fptrunc_v3f64_to_v3f32_afn: +; GFX11-GISEL: ; %bb.0: +; GFX11-GISEL-NEXT: s_clause 0x1 +; GFX11-GISEL-NEXT: s_load_b256 s[8:15], s[4:5], 0x44 +; GFX11-GISEL-NEXT: s_load_b64 s[0:1], s[4:5], 0x24 +; GFX11-GISEL-NEXT: s_mov_b32 s2, -1 +; GFX11-GISEL-NEXT: s_mov_b32 s3, 0x31016000 +; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v0, s[8:9] +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v1, s[10:11] +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v2, s[12:13] +; GFX11-GISEL-NEXT: buffer_store_b96 v[0:2], off, s[0:3], 0 +; GFX11-GISEL-NEXT: s_endpgm + %result = fptrunc afn <3 x double> %in to <3 x float> + store <3 x float> %result, ptr addrspace(1) %out + ret void +} + define amdgpu_kernel void @fptrunc_v4f64_to_v4f32(ptr addrspace(1) %out, <4 x double> %in) { ; SI-LABEL: fptrunc_v4f64_to_v4f32: ; SI: ; %bb.0: @@ -876,6 +1437,113 @@ define amdgpu_kernel void @fptrunc_v4f64_to_v4f32(ptr addrspace(1) %out, <4 x do ret void } +define amdgpu_kernel void @fptrunc_v4f64_to_v4f32_afn(ptr addrspace(1) %out, <4 x double> %in) { +; SI-LABEL: fptrunc_v4f64_to_v4f32_afn: +; SI: ; %bb.0: +; SI-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x11 +; SI-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9 +; SI-NEXT: s_mov_b32 s3, 0xf000 +; SI-NEXT: s_mov_b32 s2, -1 +; SI-NEXT: s_waitcnt lgkmcnt(0) +; SI-NEXT: v_cvt_f32_f64_e32 v3, s[14:15] +; SI-NEXT: v_cvt_f32_f64_e32 v2, s[12:13] +; SI-NEXT: v_cvt_f32_f64_e32 v1, s[10:11] +; SI-NEXT: v_cvt_f32_f64_e32 v0, s[8:9] +; SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 +; SI-NEXT: s_endpgm +; +; VI-SDAG-LABEL: fptrunc_v4f64_to_v4f32_afn: +; VI-SDAG: ; %bb.0: +; VI-SDAG-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x44 +; VI-SDAG-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24 +; VI-SDAG-NEXT: s_mov_b32 s3, 0xf000 +; VI-SDAG-NEXT: s_mov_b32 s2, -1 +; VI-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v3, s[14:15] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v2, s[12:13] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v1, s[10:11] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[8:9] +; VI-SDAG-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 +; VI-SDAG-NEXT: s_endpgm +; +; VI-GISEL-LABEL: fptrunc_v4f64_to_v4f32_afn: +; VI-GISEL: ; %bb.0: +; VI-GISEL-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x44 +; VI-GISEL-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24 +; VI-GISEL-NEXT: s_mov_b32 s2, -1 +; VI-GISEL-NEXT: s_mov_b32 s3, 0xf000 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v0, s[8:9] +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v1, s[10:11] +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v2, s[12:13] +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v3, s[14:15] +; VI-GISEL-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 +; VI-GISEL-NEXT: s_endpgm +; +; GFX10-SDAG-LABEL: fptrunc_v4f64_to_v4f32_afn: +; GFX10-SDAG: ; %bb.0: +; GFX10-SDAG-NEXT: s_clause 0x1 +; GFX10-SDAG-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x44 +; GFX10-SDAG-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24 +; GFX10-SDAG-NEXT: s_mov_b32 s3, 0x31016000 +; GFX10-SDAG-NEXT: s_mov_b32 s2, -1 +; GFX10-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v3, s[14:15] +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v2, s[12:13] +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v1, s[10:11] +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[8:9] +; GFX10-SDAG-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 +; GFX10-SDAG-NEXT: s_endpgm +; +; GFX10-GISEL-LABEL: fptrunc_v4f64_to_v4f32_afn: +; GFX10-GISEL: ; %bb.0: +; GFX10-GISEL-NEXT: s_clause 0x1 +; GFX10-GISEL-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x44 +; GFX10-GISEL-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24 +; GFX10-GISEL-NEXT: s_mov_b32 s2, -1 +; GFX10-GISEL-NEXT: s_mov_b32 s3, 0x31016000 +; GFX10-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v0, s[8:9] +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v1, s[10:11] +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v2, s[12:13] +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v3, s[14:15] +; GFX10-GISEL-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 +; GFX10-GISEL-NEXT: s_endpgm +; +; GFX11-SDAG-LABEL: fptrunc_v4f64_to_v4f32_afn: +; GFX11-SDAG: ; %bb.0: +; GFX11-SDAG-NEXT: s_clause 0x1 +; GFX11-SDAG-NEXT: s_load_b256 s[8:15], s[4:5], 0x44 +; GFX11-SDAG-NEXT: s_load_b64 s[0:1], s[4:5], 0x24 +; GFX11-SDAG-NEXT: s_mov_b32 s3, 0x31016000 +; GFX11-SDAG-NEXT: s_mov_b32 s2, -1 +; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v3, s[14:15] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v2, s[12:13] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v1, s[10:11] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[8:9] +; GFX11-SDAG-NEXT: buffer_store_b128 v[0:3], off, s[0:3], 0 +; GFX11-SDAG-NEXT: s_endpgm +; +; GFX11-GISEL-LABEL: fptrunc_v4f64_to_v4f32_afn: +; GFX11-GISEL: ; %bb.0: +; GFX11-GISEL-NEXT: s_clause 0x1 +; GFX11-GISEL-NEXT: s_load_b256 s[8:15], s[4:5], 0x44 +; GFX11-GISEL-NEXT: s_load_b64 s[0:1], s[4:5], 0x24 +; GFX11-GISEL-NEXT: s_mov_b32 s2, -1 +; GFX11-GISEL-NEXT: s_mov_b32 s3, 0x31016000 +; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v0, s[8:9] +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v1, s[10:11] +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v2, s[12:13] +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v3, s[14:15] +; GFX11-GISEL-NEXT: buffer_store_b128 v[0:3], off, s[0:3], 0 +; GFX11-GISEL-NEXT: s_endpgm + %result = fptrunc afn <4 x double> %in to <4 x float> + store <4 x float> %result, ptr addrspace(1) %out + ret void +} + define amdgpu_kernel void @fptrunc_v8f64_to_v8f32(ptr addrspace(1) %out, <8 x double> %in) { ; SI-LABEL: fptrunc_v8f64_to_v8f32: ; SI: ; %bb.0: @@ -1019,3 +1687,150 @@ define amdgpu_kernel void @fptrunc_v8f64_to_v8f32(ptr addrspace(1) %out, <8 x do store <8 x float> %result, ptr addrspace(1) %out ret void } + +define amdgpu_kernel void @fptrunc_v8f64_to_v8f32_afn(ptr addrspace(1) %out, <8 x double> %in) { +; SI-LABEL: fptrunc_v8f64_to_v8f32_afn: +; SI: ; %bb.0: +; SI-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x19 +; SI-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9 +; SI-NEXT: s_mov_b32 s3, 0xf000 +; SI-NEXT: s_mov_b32 s2, -1 +; SI-NEXT: s_waitcnt lgkmcnt(0) +; SI-NEXT: v_cvt_f32_f64_e32 v3, s[14:15] +; SI-NEXT: v_cvt_f32_f64_e32 v2, s[12:13] +; SI-NEXT: v_cvt_f32_f64_e32 v1, s[10:11] +; SI-NEXT: v_cvt_f32_f64_e32 v0, s[8:9] +; SI-NEXT: v_cvt_f32_f64_e32 v7, s[22:23] +; SI-NEXT: v_cvt_f32_f64_e32 v6, s[20:21] +; SI-NEXT: v_cvt_f32_f64_e32 v5, s[18:19] +; SI-NEXT: v_cvt_f32_f64_e32 v4, s[16:17] +; SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16 +; SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 +; SI-NEXT: s_endpgm +; +; VI-SDAG-LABEL: fptrunc_v8f64_to_v8f32_afn: +; VI-SDAG: ; %bb.0: +; VI-SDAG-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x64 +; VI-SDAG-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24 +; VI-SDAG-NEXT: s_mov_b32 s3, 0xf000 +; VI-SDAG-NEXT: s_mov_b32 s2, -1 +; VI-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v7, s[22:23] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v6, s[20:21] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v5, s[18:19] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v4, s[16:17] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v3, s[14:15] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v2, s[12:13] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v1, s[10:11] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[8:9] +; VI-SDAG-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16 +; VI-SDAG-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 +; VI-SDAG-NEXT: s_endpgm +; +; VI-GISEL-LABEL: fptrunc_v8f64_to_v8f32_afn: +; VI-GISEL: ; %bb.0: +; VI-GISEL-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x64 +; VI-GISEL-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24 +; VI-GISEL-NEXT: s_mov_b32 s2, -1 +; VI-GISEL-NEXT: s_mov_b32 s3, 0xf000 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v0, s[8:9] +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v1, s[10:11] +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v2, s[12:13] +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v3, s[14:15] +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v4, s[16:17] +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v5, s[18:19] +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v6, s[20:21] +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v7, s[22:23] +; VI-GISEL-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 +; VI-GISEL-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16 +; VI-GISEL-NEXT: s_endpgm +; +; GFX10-SDAG-LABEL: fptrunc_v8f64_to_v8f32_afn: +; GFX10-SDAG: ; %bb.0: +; GFX10-SDAG-NEXT: s_clause 0x1 +; GFX10-SDAG-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x64 +; GFX10-SDAG-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24 +; GFX10-SDAG-NEXT: s_mov_b32 s3, 0x31016000 +; GFX10-SDAG-NEXT: s_mov_b32 s2, -1 +; GFX10-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v7, s[22:23] +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v6, s[20:21] +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v5, s[18:19] +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v4, s[16:17] +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v3, s[14:15] +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v2, s[12:13] +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v1, s[10:11] +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[8:9] +; GFX10-SDAG-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16 +; GFX10-SDAG-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 +; GFX10-SDAG-NEXT: s_endpgm +; +; GFX10-GISEL-LABEL: fptrunc_v8f64_to_v8f32_afn: +; GFX10-GISEL: ; %bb.0: +; GFX10-GISEL-NEXT: s_clause 0x1 +; GFX10-GISEL-NEXT: s_load_dwordx16 s[8:23], s[4:5], 0x64 +; GFX10-GISEL-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24 +; GFX10-GISEL-NEXT: s_mov_b32 s2, -1 +; GFX10-GISEL-NEXT: s_mov_b32 s3, 0x31016000 +; GFX10-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v0, s[8:9] +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v1, s[10:11] +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v2, s[12:13] +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v3, s[14:15] +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v4, s[16:17] +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v5, s[18:19] +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v6, s[20:21] +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v7, s[22:23] +; GFX10-GISEL-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 +; GFX10-GISEL-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16 +; GFX10-GISEL-NEXT: s_endpgm +; +; GFX11-SDAG-LABEL: fptrunc_v8f64_to_v8f32_afn: +; GFX11-SDAG: ; %bb.0: +; GFX11-SDAG-NEXT: s_clause 0x1 +; GFX11-SDAG-NEXT: s_load_b512 s[8:23], s[4:5], 0x64 +; GFX11-SDAG-NEXT: s_load_b64 s[0:1], s[4:5], 0x24 +; GFX11-SDAG-NEXT: s_mov_b32 s3, 0x31016000 +; GFX11-SDAG-NEXT: s_mov_b32 s2, -1 +; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v7, s[22:23] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v6, s[20:21] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v5, s[18:19] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v4, s[16:17] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v3, s[14:15] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v2, s[12:13] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v1, s[10:11] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[8:9] +; GFX11-SDAG-NEXT: s_clause 0x1 +; GFX11-SDAG-NEXT: buffer_store_b128 v[4:7], off, s[0:3], 0 offset:16 +; GFX11-SDAG-NEXT: buffer_store_b128 v[0:3], off, s[0:3], 0 +; GFX11-SDAG-NEXT: s_endpgm +; +; GFX11-GISEL-LABEL: fptrunc_v8f64_to_v8f32_afn: +; GFX11-GISEL: ; %bb.0: +; GFX11-GISEL-NEXT: s_clause 0x1 +; GFX11-GISEL-NEXT: s_load_b512 s[8:23], s[4:5], 0x64 +; GFX11-GISEL-NEXT: s_load_b64 s[0:1], s[4:5], 0x24 +; GFX11-GISEL-NEXT: s_mov_b32 s2, -1 +; GFX11-GISEL-NEXT: s_mov_b32 s3, 0x31016000 +; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v0, s[8:9] +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v1, s[10:11] +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v2, s[12:13] +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v3, s[14:15] +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v4, s[16:17] +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v5, s[18:19] +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v6, s[20:21] +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v7, s[22:23] +; GFX11-GISEL-NEXT: s_clause 0x1 +; GFX11-GISEL-NEXT: buffer_store_b128 v[0:3], off, s[0:3], 0 +; GFX11-GISEL-NEXT: buffer_store_b128 v[4:7], off, s[0:3], 0 offset:16 +; GFX11-GISEL-NEXT: s_endpgm + %result = fptrunc <8 x double> %in to <8 x float> + store <8 x float> %result, ptr addrspace(1) %out + ret void +} +;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line: +; GFX10-SAFE-GISEL: {{.*}} +; VI-SAFE-GISEL: {{.*}} |