diff --git a/llvm/test/CodeGen/AMDGPU/fptrunc.f16.ll b/llvm/test/CodeGen/AMDGPU/fptrunc.f16.ll --- a/llvm/test/CodeGen/AMDGPU/fptrunc.f16.ll +++ b/llvm/test/CodeGen/AMDGPU/fptrunc.f16.ll @@ -1,82 +1,124 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2 -; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -verify-machineinstrs -enable-unsafe-fp-math < %s | FileCheck -enable-var-scope -check-prefixes=SI %s -; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=fiji -mattr=-flat-for-global -verify-machineinstrs -enable-unsafe-fp-math < %s | FileCheck -enable-var-scope -check-prefixes=VI %s -; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx900 -mattr=-flat-for-global -denormal-fp-math=preserve-sign -verify-machineinstrs -enable-unsafe-fp-math < %s | FileCheck -enable-var-scope -check-prefixes=GFX9 %s -; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx1100 -mattr=-flat-for-global -denormal-fp-math=preserve-sign -verify-machineinstrs -enable-unsafe-fp-math < %s | FileCheck -enable-var-scope -check-prefixes=GFX11 %s +; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -global-isel=0 -verify-machineinstrs -enable-unsafe-fp-math < %s | FileCheck -enable-var-scope -check-prefixes=SI-SDAG %s +; TODO: Crashes on selecting G_STORE. +; RUN: not --crash llc -amdgpu-scalarize-global-loads=false -march=amdgcn -global-isel=1 -verify-machineinstrs -enable-unsafe-fp-math < %s +; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=fiji -global-isel=0 -mattr=-flat-for-global -verify-machineinstrs -enable-unsafe-fp-math < %s | FileCheck -enable-var-scope -check-prefixes=VI-SDAG %s +; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=fiji -global-isel=1 -mattr=-flat-for-global -verify-machineinstrs -enable-unsafe-fp-math < %s | FileCheck -enable-var-scope -check-prefixes=VI-GISEL %s +; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx900 -global-isel=0 -mattr=-flat-for-global -denormal-fp-math=preserve-sign -verify-machineinstrs -enable-unsafe-fp-math < %s | FileCheck -enable-var-scope -check-prefixes=GFX9-SDAG %s +; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx900 -global-isel=1 -mattr=-flat-for-global -denormal-fp-math=preserve-sign -verify-machineinstrs -enable-unsafe-fp-math < %s | FileCheck -enable-var-scope -check-prefixes=GFX9-GISEL %s +; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx1100 -global-isel=0 -mattr=-flat-for-global -denormal-fp-math=preserve-sign -verify-machineinstrs -enable-unsafe-fp-math < %s | FileCheck -enable-var-scope -check-prefixes=GFX11-SDAG %s +; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx1100 -global-isel=1 -mattr=-flat-for-global -denormal-fp-math=preserve-sign -verify-machineinstrs -enable-unsafe-fp-math < %s | FileCheck -enable-var-scope -check-prefixes=GFX11-GISEL %s define amdgpu_kernel void @fptrunc_f32_to_f16( -; SI-LABEL: fptrunc_f32_to_f16: -; SI: ; %bb.0: ; %entry -; SI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 -; SI-NEXT: s_mov_b32 s7, 0xf000 -; SI-NEXT: s_mov_b32 s6, -1 -; SI-NEXT: s_mov_b32 s10, s6 -; SI-NEXT: s_mov_b32 s11, s7 -; SI-NEXT: s_waitcnt lgkmcnt(0) -; SI-NEXT: s_mov_b32 s8, s2 -; SI-NEXT: s_mov_b32 s9, s3 -; SI-NEXT: buffer_load_dword v0, off, s[8:11], 0 -; SI-NEXT: s_mov_b32 s4, s0 -; SI-NEXT: s_mov_b32 s5, s1 -; SI-NEXT: s_waitcnt vmcnt(0) -; SI-NEXT: v_cvt_f16_f32_e32 v0, v0 -; SI-NEXT: buffer_store_short v0, off, s[4:7], 0 -; SI-NEXT: s_endpgm -; -; VI-LABEL: fptrunc_f32_to_f16: -; VI: ; %bb.0: ; %entry -; VI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 -; VI-NEXT: s_mov_b32 s7, 0xf000 -; VI-NEXT: s_mov_b32 s6, -1 -; VI-NEXT: s_mov_b32 s10, s6 -; VI-NEXT: s_mov_b32 s11, s7 -; VI-NEXT: s_waitcnt lgkmcnt(0) -; VI-NEXT: s_mov_b32 s8, s2 -; VI-NEXT: s_mov_b32 s9, s3 -; VI-NEXT: buffer_load_dword v0, off, s[8:11], 0 -; VI-NEXT: s_mov_b32 s4, s0 -; VI-NEXT: s_mov_b32 s5, s1 -; VI-NEXT: s_waitcnt vmcnt(0) -; VI-NEXT: v_cvt_f16_f32_e32 v0, v0 -; VI-NEXT: buffer_store_short v0, off, s[4:7], 0 -; VI-NEXT: s_endpgm -; -; GFX9-LABEL: fptrunc_f32_to_f16: -; GFX9: ; %bb.0: ; %entry -; GFX9-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 -; GFX9-NEXT: s_mov_b32 s7, 0xf000 -; GFX9-NEXT: s_mov_b32 s6, -1 -; GFX9-NEXT: s_mov_b32 s10, s6 -; GFX9-NEXT: s_mov_b32 s11, s7 -; GFX9-NEXT: s_waitcnt lgkmcnt(0) -; GFX9-NEXT: s_mov_b32 s8, s2 -; GFX9-NEXT: s_mov_b32 s9, s3 -; GFX9-NEXT: buffer_load_dword v0, off, s[8:11], 0 -; GFX9-NEXT: s_mov_b32 s4, s0 -; GFX9-NEXT: s_mov_b32 s5, s1 -; GFX9-NEXT: s_waitcnt vmcnt(0) -; GFX9-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX9-NEXT: buffer_store_short v0, off, s[4:7], 0 -; GFX9-NEXT: s_endpgm -; -; GFX11-LABEL: fptrunc_f32_to_f16: -; GFX11: ; %bb.0: ; %entry -; GFX11-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 -; GFX11-NEXT: s_mov_b32 s6, -1 -; GFX11-NEXT: s_mov_b32 s7, 0x31016000 -; GFX11-NEXT: s_mov_b32 s10, s6 -; GFX11-NEXT: s_mov_b32 s11, s7 -; GFX11-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NEXT: s_mov_b32 s8, s2 -; GFX11-NEXT: s_mov_b32 s9, s3 -; GFX11-NEXT: s_mov_b32 s4, s0 -; GFX11-NEXT: buffer_load_b32 v0, off, s[8:11], 0 -; GFX11-NEXT: s_mov_b32 s5, s1 -; GFX11-NEXT: s_waitcnt vmcnt(0) -; GFX11-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX11-NEXT: buffer_store_b16 v0, off, s[4:7], 0 -; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) -; GFX11-NEXT: s_endpgm +; SI-SDAG-LABEL: fptrunc_f32_to_f16: +; SI-SDAG: ; %bb.0: ; %entry +; SI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 +; SI-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; SI-SDAG-NEXT: s_mov_b32 s6, -1 +; SI-SDAG-NEXT: s_mov_b32 s10, s6 +; SI-SDAG-NEXT: s_mov_b32 s11, s7 +; SI-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; SI-SDAG-NEXT: s_mov_b32 s8, s2 +; SI-SDAG-NEXT: s_mov_b32 s9, s3 +; SI-SDAG-NEXT: buffer_load_dword v0, off, s[8:11], 0 +; SI-SDAG-NEXT: s_mov_b32 s4, s0 +; SI-SDAG-NEXT: s_mov_b32 s5, s1 +; SI-SDAG-NEXT: s_waitcnt vmcnt(0) +; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; SI-SDAG-NEXT: buffer_store_short v0, off, s[4:7], 0 +; SI-SDAG-NEXT: s_endpgm +; +; VI-SDAG-LABEL: fptrunc_f32_to_f16: +; VI-SDAG: ; %bb.0: ; %entry +; VI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; VI-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; VI-SDAG-NEXT: s_mov_b32 s6, -1 +; VI-SDAG-NEXT: s_mov_b32 s10, s6 +; VI-SDAG-NEXT: s_mov_b32 s11, s7 +; VI-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; VI-SDAG-NEXT: s_mov_b32 s8, s2 +; VI-SDAG-NEXT: s_mov_b32 s9, s3 +; VI-SDAG-NEXT: buffer_load_dword v0, off, s[8:11], 0 +; VI-SDAG-NEXT: s_mov_b32 s4, s0 +; VI-SDAG-NEXT: s_mov_b32 s5, s1 +; VI-SDAG-NEXT: s_waitcnt vmcnt(0) +; VI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; VI-SDAG-NEXT: buffer_store_short v0, off, s[4:7], 0 +; VI-SDAG-NEXT: s_endpgm +; +; VI-GISEL-LABEL: fptrunc_f32_to_f16: +; VI-GISEL: ; %bb.0: ; %entry +; VI-GISEL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: s_load_dword s2, s[2:3], 0x0 +; VI-GISEL-NEXT: s_mov_b32 s3, 0xf000 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: v_cvt_f16_f32_e32 v0, s2 +; VI-GISEL-NEXT: s_mov_b32 s2, -1 +; VI-GISEL-NEXT: buffer_store_short v0, off, s[0:3], 0 +; VI-GISEL-NEXT: s_endpgm +; +; GFX9-SDAG-LABEL: fptrunc_f32_to_f16: +; GFX9-SDAG: ; %bb.0: ; %entry +; GFX9-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; GFX9-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; GFX9-SDAG-NEXT: s_mov_b32 s6, -1 +; GFX9-SDAG-NEXT: s_mov_b32 s10, s6 +; GFX9-SDAG-NEXT: s_mov_b32 s11, s7 +; GFX9-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-SDAG-NEXT: s_mov_b32 s8, s2 +; GFX9-SDAG-NEXT: s_mov_b32 s9, s3 +; GFX9-SDAG-NEXT: buffer_load_dword v0, off, s[8:11], 0 +; GFX9-SDAG-NEXT: s_mov_b32 s4, s0 +; GFX9-SDAG-NEXT: s_mov_b32 s5, s1 +; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) +; GFX9-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; GFX9-SDAG-NEXT: buffer_store_short v0, off, s[4:7], 0 +; GFX9-SDAG-NEXT: s_endpgm +; +; GFX9-GISEL-LABEL: fptrunc_f32_to_f16: +; GFX9-GISEL: ; %bb.0: ; %entry +; GFX9-GISEL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-GISEL-NEXT: s_load_dword s2, s[2:3], 0x0 +; GFX9-GISEL-NEXT: s_mov_b32 s3, 0xf000 +; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-GISEL-NEXT: v_cvt_f16_f32_e32 v0, s2 +; GFX9-GISEL-NEXT: s_mov_b32 s2, -1 +; GFX9-GISEL-NEXT: buffer_store_short v0, off, s[0:3], 0 +; GFX9-GISEL-NEXT: s_endpgm +; +; GFX11-SDAG-LABEL: fptrunc_f32_to_f16: +; GFX11-SDAG: ; %bb.0: ; %entry +; GFX11-SDAG-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 +; GFX11-SDAG-NEXT: s_mov_b32 s6, -1 +; GFX11-SDAG-NEXT: s_mov_b32 s7, 0x31016000 +; GFX11-SDAG-NEXT: s_mov_b32 s10, s6 +; GFX11-SDAG-NEXT: s_mov_b32 s11, s7 +; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-SDAG-NEXT: s_mov_b32 s8, s2 +; GFX11-SDAG-NEXT: s_mov_b32 s9, s3 +; GFX11-SDAG-NEXT: s_mov_b32 s4, s0 +; GFX11-SDAG-NEXT: buffer_load_b32 v0, off, s[8:11], 0 +; GFX11-SDAG-NEXT: s_mov_b32 s5, s1 +; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) +; GFX11-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; GFX11-SDAG-NEXT: buffer_store_b16 v0, off, s[4:7], 0 +; GFX11-SDAG-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-SDAG-NEXT: s_endpgm +; +; GFX11-GISEL-LABEL: fptrunc_f32_to_f16: +; GFX11-GISEL: ; %bb.0: ; %entry +; GFX11-GISEL-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 +; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-GISEL-NEXT: s_load_b32 s2, s[2:3], 0x0 +; GFX11-GISEL-NEXT: s_mov_b32 s3, 0x31016000 +; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-GISEL-NEXT: v_cvt_f16_f32_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_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-GISEL-NEXT: s_endpgm ptr addrspace(1) %r, ptr addrspace(1) %a) { entry: @@ -87,83 +129,249 @@ } define amdgpu_kernel void @fptrunc_f64_to_f16( -; SI-LABEL: fptrunc_f64_to_f16: -; SI: ; %bb.0: ; %entry -; SI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 -; SI-NEXT: s_mov_b32 s7, 0xf000 -; SI-NEXT: s_mov_b32 s6, -1 -; SI-NEXT: s_mov_b32 s10, s6 -; SI-NEXT: s_mov_b32 s11, s7 -; SI-NEXT: s_waitcnt lgkmcnt(0) -; SI-NEXT: s_mov_b32 s8, s2 -; SI-NEXT: s_mov_b32 s9, s3 -; SI-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0 -; SI-NEXT: s_mov_b32 s4, s0 -; SI-NEXT: s_mov_b32 s5, s1 -; SI-NEXT: s_waitcnt vmcnt(0) -; SI-NEXT: v_cvt_f32_f64_e32 v0, v[0:1] -; SI-NEXT: v_cvt_f16_f32_e32 v0, v0 -; SI-NEXT: buffer_store_short v0, off, s[4:7], 0 -; SI-NEXT: s_endpgm -; -; VI-LABEL: fptrunc_f64_to_f16: -; VI: ; %bb.0: ; %entry -; VI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 -; VI-NEXT: s_mov_b32 s7, 0xf000 -; VI-NEXT: s_mov_b32 s6, -1 -; VI-NEXT: s_mov_b32 s10, s6 -; VI-NEXT: s_mov_b32 s11, s7 -; VI-NEXT: s_waitcnt lgkmcnt(0) -; VI-NEXT: s_mov_b32 s8, s2 -; VI-NEXT: s_mov_b32 s9, s3 -; VI-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0 -; VI-NEXT: s_mov_b32 s4, s0 -; VI-NEXT: s_mov_b32 s5, s1 -; VI-NEXT: s_waitcnt vmcnt(0) -; VI-NEXT: v_cvt_f32_f64_e32 v0, v[0:1] -; VI-NEXT: v_cvt_f16_f32_e32 v0, v0 -; VI-NEXT: buffer_store_short v0, off, s[4:7], 0 -; VI-NEXT: s_endpgm -; -; GFX9-LABEL: fptrunc_f64_to_f16: -; GFX9: ; %bb.0: ; %entry -; GFX9-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 -; GFX9-NEXT: s_mov_b32 s7, 0xf000 -; GFX9-NEXT: s_mov_b32 s6, -1 -; GFX9-NEXT: s_mov_b32 s10, s6 -; GFX9-NEXT: s_mov_b32 s11, s7 -; GFX9-NEXT: s_waitcnt lgkmcnt(0) -; GFX9-NEXT: s_mov_b32 s8, s2 -; GFX9-NEXT: s_mov_b32 s9, s3 -; GFX9-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0 -; GFX9-NEXT: s_mov_b32 s4, s0 -; GFX9-NEXT: s_mov_b32 s5, s1 -; GFX9-NEXT: s_waitcnt vmcnt(0) -; GFX9-NEXT: v_cvt_f32_f64_e32 v0, v[0:1] -; GFX9-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX9-NEXT: buffer_store_short v0, off, s[4:7], 0 -; GFX9-NEXT: s_endpgm -; -; GFX11-LABEL: fptrunc_f64_to_f16: -; GFX11: ; %bb.0: ; %entry -; GFX11-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 -; GFX11-NEXT: s_mov_b32 s6, -1 -; GFX11-NEXT: s_mov_b32 s7, 0x31016000 -; GFX11-NEXT: s_mov_b32 s10, s6 -; GFX11-NEXT: s_mov_b32 s11, s7 -; GFX11-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NEXT: s_mov_b32 s8, s2 -; GFX11-NEXT: s_mov_b32 s9, s3 -; GFX11-NEXT: s_mov_b32 s4, s0 -; GFX11-NEXT: buffer_load_b64 v[0:1], off, s[8:11], 0 -; GFX11-NEXT: s_mov_b32 s5, s1 -; GFX11-NEXT: s_waitcnt vmcnt(0) -; GFX11-NEXT: v_cvt_f32_f64_e32 v0, v[0:1] -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX11-NEXT: buffer_store_b16 v0, off, s[4:7], 0 -; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) -; GFX11-NEXT: s_endpgm +; SI-SDAG-LABEL: fptrunc_f64_to_f16: +; SI-SDAG: ; %bb.0: ; %entry +; SI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 +; SI-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; SI-SDAG-NEXT: s_mov_b32 s6, -1 +; SI-SDAG-NEXT: s_mov_b32 s10, s6 +; SI-SDAG-NEXT: s_mov_b32 s11, s7 +; SI-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; SI-SDAG-NEXT: s_mov_b32 s8, s2 +; SI-SDAG-NEXT: s_mov_b32 s9, s3 +; SI-SDAG-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0 +; SI-SDAG-NEXT: s_mov_b32 s4, s0 +; SI-SDAG-NEXT: s_mov_b32 s5, s1 +; SI-SDAG-NEXT: s_waitcnt vmcnt(0) +; SI-SDAG-NEXT: v_cvt_f32_f64_e32 v0, v[0:1] +; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; SI-SDAG-NEXT: buffer_store_short v0, off, s[4:7], 0 +; SI-SDAG-NEXT: s_endpgm +; +; VI-SDAG-LABEL: fptrunc_f64_to_f16: +; VI-SDAG: ; %bb.0: ; %entry +; VI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; VI-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; VI-SDAG-NEXT: s_mov_b32 s6, -1 +; VI-SDAG-NEXT: s_mov_b32 s10, s6 +; VI-SDAG-NEXT: s_mov_b32 s11, s7 +; VI-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; VI-SDAG-NEXT: s_mov_b32 s8, s2 +; VI-SDAG-NEXT: s_mov_b32 s9, s3 +; VI-SDAG-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0 +; VI-SDAG-NEXT: s_mov_b32 s4, s0 +; VI-SDAG-NEXT: s_mov_b32 s5, s1 +; VI-SDAG-NEXT: s_waitcnt vmcnt(0) +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v0, v[0:1] +; VI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; VI-SDAG-NEXT: buffer_store_short v0, off, s[4:7], 0 +; VI-SDAG-NEXT: s_endpgm +; +; VI-GISEL-LABEL: fptrunc_f64_to_f16: +; VI-GISEL: ; %bb.0: ; %entry +; VI-GISEL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; VI-GISEL-NEXT: s_movk_i32 s4, 0x7e00 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x0 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: s_bfe_u32 s5, s3, 0xb0014 +; VI-GISEL-NEXT: s_lshr_b32 s6, s3, 8 +; VI-GISEL-NEXT: s_and_b32 s7, s3, 0x1ff +; VI-GISEL-NEXT: s_addk_i32 s5, 0xfc10 +; VI-GISEL-NEXT: s_and_b32 s6, s6, 0xffe +; VI-GISEL-NEXT: s_or_b32 s2, s7, 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, s6, s2 +; VI-GISEL-NEXT: s_cmp_lg_u32 s2, 0 +; VI-GISEL-NEXT: s_cselect_b32 s4, s4, 0x7c00 +; VI-GISEL-NEXT: s_sub_i32 s7, 1, s5 +; VI-GISEL-NEXT: s_lshl_b32 s6, s5, 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_lshr_b32 s8, s2, s7 +; 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 s5, 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_and_b32 s6, s6, 1 +; VI-GISEL-NEXT: s_add_i32 s2, s2, s6 +; VI-GISEL-NEXT: s_cmp_gt_i32 s5, 30 +; VI-GISEL-NEXT: s_cselect_b32 s2, 0x7c00, s2 +; VI-GISEL-NEXT: s_cmpk_eq_i32 s5, 0x40f +; VI-GISEL-NEXT: s_cselect_b32 s2, s4, 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 +; +; GFX9-SDAG-LABEL: fptrunc_f64_to_f16: +; GFX9-SDAG: ; %bb.0: ; %entry +; GFX9-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; GFX9-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; GFX9-SDAG-NEXT: s_mov_b32 s6, -1 +; GFX9-SDAG-NEXT: s_mov_b32 s10, s6 +; GFX9-SDAG-NEXT: s_mov_b32 s11, s7 +; GFX9-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-SDAG-NEXT: s_mov_b32 s8, s2 +; GFX9-SDAG-NEXT: s_mov_b32 s9, s3 +; GFX9-SDAG-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0 +; GFX9-SDAG-NEXT: s_mov_b32 s4, s0 +; GFX9-SDAG-NEXT: s_mov_b32 s5, s1 +; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) +; GFX9-SDAG-NEXT: v_cvt_f32_f64_e32 v0, v[0:1] +; GFX9-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; GFX9-SDAG-NEXT: buffer_store_short v0, off, s[4:7], 0 +; GFX9-SDAG-NEXT: s_endpgm +; +; GFX9-GISEL-LABEL: fptrunc_f64_to_f16: +; GFX9-GISEL: ; %bb.0: ; %entry +; GFX9-GISEL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; GFX9-GISEL-NEXT: s_movk_i32 s4, 0x7e00 +; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-GISEL-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x0 +; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-GISEL-NEXT: s_bfe_u32 s5, s3, 0xb0014 +; GFX9-GISEL-NEXT: s_lshr_b32 s6, s3, 8 +; GFX9-GISEL-NEXT: s_and_b32 s7, s3, 0x1ff +; GFX9-GISEL-NEXT: s_addk_i32 s5, 0xfc10 +; GFX9-GISEL-NEXT: s_and_b32 s6, s6, 0xffe +; GFX9-GISEL-NEXT: s_or_b32 s2, s7, s2 +; GFX9-GISEL-NEXT: s_cmp_lg_u32 s2, 0 +; GFX9-GISEL-NEXT: s_cselect_b32 s2, 1, 0 +; GFX9-GISEL-NEXT: s_or_b32 s2, s6, s2 +; GFX9-GISEL-NEXT: s_cmp_lg_u32 s2, 0 +; GFX9-GISEL-NEXT: s_cselect_b32 s4, s4, 0x7c00 +; GFX9-GISEL-NEXT: s_sub_i32 s7, 1, s5 +; GFX9-GISEL-NEXT: s_lshl_b32 s6, s5, 12 +; GFX9-GISEL-NEXT: s_max_i32 s7, s7, 0 +; GFX9-GISEL-NEXT: s_or_b32 s6, s2, s6 +; GFX9-GISEL-NEXT: s_min_i32 s7, s7, 13 +; GFX9-GISEL-NEXT: s_bitset1_b32 s2, 12 +; GFX9-GISEL-NEXT: s_lshr_b32 s8, s2, s7 +; GFX9-GISEL-NEXT: s_lshl_b32 s7, s8, s7 +; GFX9-GISEL-NEXT: s_cmp_lg_u32 s7, s2 +; GFX9-GISEL-NEXT: s_cselect_b32 s2, 1, 0 +; GFX9-GISEL-NEXT: s_or_b32 s2, s8, s2 +; GFX9-GISEL-NEXT: s_cmp_lt_i32 s5, 1 +; GFX9-GISEL-NEXT: s_cselect_b32 s2, s2, s6 +; GFX9-GISEL-NEXT: s_and_b32 s6, s2, 7 +; GFX9-GISEL-NEXT: s_lshr_b32 s2, s2, 2 +; GFX9-GISEL-NEXT: s_cmp_eq_u32 s6, 3 +; GFX9-GISEL-NEXT: s_cselect_b32 s7, 1, 0 +; GFX9-GISEL-NEXT: s_cmp_gt_i32 s6, 5 +; GFX9-GISEL-NEXT: s_cselect_b32 s6, 1, 0 +; GFX9-GISEL-NEXT: s_or_b32 s6, s7, s6 +; GFX9-GISEL-NEXT: s_and_b32 s6, s6, 1 +; GFX9-GISEL-NEXT: s_add_i32 s2, s2, s6 +; GFX9-GISEL-NEXT: s_cmp_gt_i32 s5, 30 +; GFX9-GISEL-NEXT: s_cselect_b32 s2, 0x7c00, s2 +; GFX9-GISEL-NEXT: s_cmpk_eq_i32 s5, 0x40f +; GFX9-GISEL-NEXT: s_cselect_b32 s2, s4, s2 +; GFX9-GISEL-NEXT: s_lshr_b32 s3, s3, 16 +; GFX9-GISEL-NEXT: s_and_b32 s3, s3, 0x8000 +; GFX9-GISEL-NEXT: s_or_b32 s2, s3, s2 +; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, s2 +; GFX9-GISEL-NEXT: s_mov_b32 s2, -1 +; GFX9-GISEL-NEXT: s_mov_b32 s3, 0xf000 +; GFX9-GISEL-NEXT: buffer_store_short v0, off, s[0:3], 0 +; GFX9-GISEL-NEXT: s_endpgm +; +; GFX11-SDAG-LABEL: fptrunc_f64_to_f16: +; GFX11-SDAG: ; %bb.0: ; %entry +; GFX11-SDAG-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 +; GFX11-SDAG-NEXT: s_mov_b32 s6, -1 +; GFX11-SDAG-NEXT: s_mov_b32 s7, 0x31016000 +; GFX11-SDAG-NEXT: s_mov_b32 s10, s6 +; GFX11-SDAG-NEXT: s_mov_b32 s11, s7 +; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-SDAG-NEXT: s_mov_b32 s8, s2 +; GFX11-SDAG-NEXT: s_mov_b32 s9, s3 +; GFX11-SDAG-NEXT: s_mov_b32 s4, s0 +; GFX11-SDAG-NEXT: buffer_load_b64 v[0:1], off, s[8:11], 0 +; GFX11-SDAG-NEXT: s_mov_b32 s5, s1 +; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v0, v[0:1] +; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; GFX11-SDAG-NEXT: buffer_store_b16 v0, off, s[4:7], 0 +; GFX11-SDAG-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-SDAG-NEXT: s_endpgm +; +; GFX11-GISEL-LABEL: fptrunc_f64_to_f16: +; GFX11-GISEL: ; %bb.0: ; %entry +; GFX11-GISEL-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 +; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-GISEL-NEXT: s_load_b64 s[2:3], s[2:3], 0x0 +; 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) +; GFX11-GISEL-NEXT: s_or_b32 s2, s5, s2 +; GFX11-GISEL-NEXT: s_movk_i32 s5, 0x7e00 +; GFX11-GISEL-NEXT: s_cmp_lg_u32 s2, 0 +; GFX11-GISEL-NEXT: s_cselect_b32 s5, s5, 0x7c00 +; GFX11-GISEL-NEXT: s_sub_i32 s6, 1, s4 +; GFX11-GISEL-NEXT: s_or_b32 s7, s2, 0x1000 +; GFX11-GISEL-NEXT: s_max_i32 s6, s6, 0 +; GFX11-GISEL-NEXT: s_lshl_b32 s9, s4, 12 +; GFX11-GISEL-NEXT: s_min_i32 s6, s6, 13 +; GFX11-GISEL-NEXT: s_or_b32 s2, s2, s9 +; GFX11-GISEL-NEXT: s_lshr_b32 s8, s7, s6 +; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) +; GFX11-GISEL-NEXT: s_lshl_b32 s6, s8, s6 +; GFX11-GISEL-NEXT: s_cmp_lg_u32 s6, s7 +; 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, s8, 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_and_b32 s6, s6, 1 +; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) +; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-GISEL-NEXT: s_endpgm ptr addrspace(1) %r, ptr addrspace(1) %a) { entry: @@ -174,88 +382,131 @@ } define amdgpu_kernel void @fptrunc_v2f32_to_v2f16( -; SI-LABEL: fptrunc_v2f32_to_v2f16: -; SI: ; %bb.0: ; %entry -; SI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 -; SI-NEXT: s_mov_b32 s7, 0xf000 -; SI-NEXT: s_mov_b32 s6, -1 -; SI-NEXT: s_mov_b32 s10, s6 -; SI-NEXT: s_mov_b32 s11, s7 -; SI-NEXT: s_waitcnt lgkmcnt(0) -; SI-NEXT: s_mov_b32 s8, s2 -; SI-NEXT: s_mov_b32 s9, s3 -; SI-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0 -; SI-NEXT: s_mov_b32 s4, s0 -; SI-NEXT: s_mov_b32 s5, s1 -; SI-NEXT: s_waitcnt vmcnt(0) -; SI-NEXT: v_cvt_f16_f32_e32 v1, v1 -; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; SI-NEXT: v_cvt_f16_f32_e32 v0, v0 -; SI-NEXT: v_or_b32_e32 v0, v0, v1 -; SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 -; SI-NEXT: s_endpgm -; -; VI-LABEL: fptrunc_v2f32_to_v2f16: -; VI: ; %bb.0: ; %entry -; VI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 -; VI-NEXT: s_mov_b32 s7, 0xf000 -; VI-NEXT: s_mov_b32 s6, -1 -; VI-NEXT: s_mov_b32 s10, s6 -; VI-NEXT: s_mov_b32 s11, s7 -; VI-NEXT: s_waitcnt lgkmcnt(0) -; VI-NEXT: s_mov_b32 s8, s2 -; VI-NEXT: s_mov_b32 s9, s3 -; VI-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0 -; VI-NEXT: s_mov_b32 s4, s0 -; VI-NEXT: s_mov_b32 s5, s1 -; VI-NEXT: s_waitcnt vmcnt(0) -; VI-NEXT: v_cvt_f16_f32_sdwa v1, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD -; VI-NEXT: v_cvt_f16_f32_e32 v0, v0 -; VI-NEXT: v_or_b32_e32 v0, v0, v1 -; VI-NEXT: buffer_store_dword v0, off, s[4:7], 0 -; VI-NEXT: s_endpgm -; -; GFX9-LABEL: fptrunc_v2f32_to_v2f16: -; GFX9: ; %bb.0: ; %entry -; GFX9-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 -; GFX9-NEXT: s_mov_b32 s7, 0xf000 -; GFX9-NEXT: s_mov_b32 s6, -1 -; GFX9-NEXT: s_mov_b32 s10, s6 -; GFX9-NEXT: s_mov_b32 s11, s7 -; GFX9-NEXT: s_waitcnt lgkmcnt(0) -; GFX9-NEXT: s_mov_b32 s8, s2 -; GFX9-NEXT: s_mov_b32 s9, s3 -; GFX9-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0 -; GFX9-NEXT: s_mov_b32 s4, s0 -; GFX9-NEXT: s_mov_b32 s5, s1 -; GFX9-NEXT: s_waitcnt vmcnt(0) -; GFX9-NEXT: v_cvt_f16_f32_e32 v1, v1 -; GFX9-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX9-NEXT: v_pack_b32_f16 v0, v0, v1 -; GFX9-NEXT: buffer_store_dword v0, off, s[4:7], 0 -; GFX9-NEXT: s_endpgm -; -; GFX11-LABEL: fptrunc_v2f32_to_v2f16: -; GFX11: ; %bb.0: ; %entry -; GFX11-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 -; GFX11-NEXT: s_mov_b32 s6, -1 -; GFX11-NEXT: s_mov_b32 s7, 0x31016000 -; GFX11-NEXT: s_mov_b32 s10, s6 -; GFX11-NEXT: s_mov_b32 s11, s7 -; GFX11-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NEXT: s_mov_b32 s8, s2 -; GFX11-NEXT: s_mov_b32 s9, s3 -; GFX11-NEXT: s_mov_b32 s4, s0 -; GFX11-NEXT: buffer_load_b64 v[0:1], off, s[8:11], 0 -; GFX11-NEXT: s_mov_b32 s5, s1 -; GFX11-NEXT: s_waitcnt vmcnt(0) -; GFX11-NEXT: v_cvt_f16_f32_e32 v1, v1 -; GFX11-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NEXT: v_pack_b32_f16 v0, v0, v1 -; GFX11-NEXT: buffer_store_b32 v0, off, s[4:7], 0 -; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) -; GFX11-NEXT: s_endpgm +; SI-SDAG-LABEL: fptrunc_v2f32_to_v2f16: +; SI-SDAG: ; %bb.0: ; %entry +; SI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 +; SI-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; SI-SDAG-NEXT: s_mov_b32 s6, -1 +; SI-SDAG-NEXT: s_mov_b32 s10, s6 +; SI-SDAG-NEXT: s_mov_b32 s11, s7 +; SI-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; SI-SDAG-NEXT: s_mov_b32 s8, s2 +; SI-SDAG-NEXT: s_mov_b32 s9, s3 +; SI-SDAG-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0 +; SI-SDAG-NEXT: s_mov_b32 s4, s0 +; SI-SDAG-NEXT: s_mov_b32 s5, s1 +; SI-SDAG-NEXT: s_waitcnt vmcnt(0) +; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v1, v1 +; SI-SDAG-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; SI-SDAG-NEXT: v_or_b32_e32 v0, v0, v1 +; SI-SDAG-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; SI-SDAG-NEXT: s_endpgm +; +; VI-SDAG-LABEL: fptrunc_v2f32_to_v2f16: +; VI-SDAG: ; %bb.0: ; %entry +; VI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; VI-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; VI-SDAG-NEXT: s_mov_b32 s6, -1 +; VI-SDAG-NEXT: s_mov_b32 s10, s6 +; VI-SDAG-NEXT: s_mov_b32 s11, s7 +; VI-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; VI-SDAG-NEXT: s_mov_b32 s8, s2 +; VI-SDAG-NEXT: s_mov_b32 s9, s3 +; VI-SDAG-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0 +; VI-SDAG-NEXT: s_mov_b32 s4, s0 +; VI-SDAG-NEXT: s_mov_b32 s5, s1 +; VI-SDAG-NEXT: s_waitcnt vmcnt(0) +; VI-SDAG-NEXT: v_cvt_f16_f32_sdwa v1, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD +; VI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; VI-SDAG-NEXT: v_or_b32_e32 v0, v0, v1 +; VI-SDAG-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; VI-SDAG-NEXT: s_endpgm +; +; VI-GISEL-LABEL: fptrunc_v2f32_to_v2f16: +; VI-GISEL: ; %bb.0: ; %entry +; VI-GISEL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x0 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: v_mov_b32_e32 v1, s3 +; VI-GISEL-NEXT: v_cvt_f16_f32_e32 v0, s2 +; VI-GISEL-NEXT: v_cvt_f16_f32_sdwa v1, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD +; VI-GISEL-NEXT: v_or_b32_e32 v2, v0, v1 +; VI-GISEL-NEXT: v_mov_b32_e32 v0, s0 +; VI-GISEL-NEXT: v_mov_b32_e32 v1, s1 +; VI-GISEL-NEXT: flat_store_dword v[0:1], v2 +; VI-GISEL-NEXT: s_endpgm +; +; GFX9-SDAG-LABEL: fptrunc_v2f32_to_v2f16: +; GFX9-SDAG: ; %bb.0: ; %entry +; GFX9-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; GFX9-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; GFX9-SDAG-NEXT: s_mov_b32 s6, -1 +; GFX9-SDAG-NEXT: s_mov_b32 s10, s6 +; GFX9-SDAG-NEXT: s_mov_b32 s11, s7 +; GFX9-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-SDAG-NEXT: s_mov_b32 s8, s2 +; GFX9-SDAG-NEXT: s_mov_b32 s9, s3 +; GFX9-SDAG-NEXT: buffer_load_dwordx2 v[0:1], off, s[8:11], 0 +; GFX9-SDAG-NEXT: s_mov_b32 s4, s0 +; GFX9-SDAG-NEXT: s_mov_b32 s5, s1 +; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) +; GFX9-SDAG-NEXT: v_cvt_f16_f32_e32 v1, v1 +; GFX9-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; GFX9-SDAG-NEXT: v_pack_b32_f16 v0, v0, v1 +; GFX9-SDAG-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; GFX9-SDAG-NEXT: s_endpgm +; +; GFX9-GISEL-LABEL: fptrunc_v2f32_to_v2f16: +; GFX9-GISEL: ; %bb.0: ; %entry +; GFX9-GISEL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-GISEL-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x0 +; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-GISEL-NEXT: v_cvt_f16_f32_e32 v0, s2 +; GFX9-GISEL-NEXT: v_cvt_f16_f32_e32 v1, s3 +; GFX9-GISEL-NEXT: v_pack_b32_f16 v0, v0, v1 +; GFX9-GISEL-NEXT: v_mov_b32_e32 v1, 0 +; GFX9-GISEL-NEXT: global_store_dword v1, v0, s[0:1] +; GFX9-GISEL-NEXT: s_endpgm +; +; GFX11-SDAG-LABEL: fptrunc_v2f32_to_v2f16: +; GFX11-SDAG: ; %bb.0: ; %entry +; GFX11-SDAG-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 +; GFX11-SDAG-NEXT: s_mov_b32 s6, -1 +; GFX11-SDAG-NEXT: s_mov_b32 s7, 0x31016000 +; GFX11-SDAG-NEXT: s_mov_b32 s10, s6 +; GFX11-SDAG-NEXT: s_mov_b32 s11, s7 +; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-SDAG-NEXT: s_mov_b32 s8, s2 +; GFX11-SDAG-NEXT: s_mov_b32 s9, s3 +; GFX11-SDAG-NEXT: s_mov_b32 s4, s0 +; GFX11-SDAG-NEXT: buffer_load_b64 v[0:1], off, s[8:11], 0 +; GFX11-SDAG-NEXT: s_mov_b32 s5, s1 +; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) +; GFX11-SDAG-NEXT: v_cvt_f16_f32_e32 v1, v1 +; GFX11-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-SDAG-NEXT: v_pack_b32_f16 v0, v0, v1 +; GFX11-SDAG-NEXT: buffer_store_b32 v0, off, s[4:7], 0 +; GFX11-SDAG-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-SDAG-NEXT: s_endpgm +; +; GFX11-GISEL-LABEL: fptrunc_v2f32_to_v2f16: +; GFX11-GISEL: ; %bb.0: ; %entry +; GFX11-GISEL-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 +; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-GISEL-NEXT: s_load_b64 s[2:3], s[2:3], 0x0 +; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-GISEL-NEXT: v_cvt_f16_f32_e32 v0, s2 +; GFX11-GISEL-NEXT: v_cvt_f16_f32_e32 v1, s3 +; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-GISEL-NEXT: v_pack_b32_f16 v0, v0, v1 +; GFX11-GISEL-NEXT: v_mov_b32_e32 v1, 0 +; GFX11-GISEL-NEXT: global_store_b32 v1, v0, s[0:1] +; GFX11-GISEL-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-GISEL-NEXT: s_endpgm ptr addrspace(1) %r, ptr addrspace(1) %a) { entry: @@ -266,98 +517,394 @@ } define amdgpu_kernel void @fptrunc_v2f64_to_v2f16( -; SI-LABEL: fptrunc_v2f64_to_v2f16: -; SI: ; %bb.0: ; %entry -; SI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 -; SI-NEXT: s_mov_b32 s7, 0xf000 -; SI-NEXT: s_mov_b32 s6, -1 -; SI-NEXT: s_mov_b32 s10, s6 -; SI-NEXT: s_mov_b32 s11, s7 -; SI-NEXT: s_waitcnt lgkmcnt(0) -; SI-NEXT: s_mov_b32 s8, s2 -; SI-NEXT: s_mov_b32 s9, s3 -; SI-NEXT: buffer_load_dwordx4 v[0:3], off, s[8:11], 0 -; SI-NEXT: s_mov_b32 s4, s0 -; SI-NEXT: s_mov_b32 s5, s1 -; SI-NEXT: s_waitcnt vmcnt(0) -; SI-NEXT: v_cvt_f32_f64_e32 v2, v[2:3] -; SI-NEXT: v_cvt_f32_f64_e32 v0, v[0:1] -; SI-NEXT: v_cvt_f16_f32_e32 v1, v2 -; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; SI-NEXT: v_cvt_f16_f32_e32 v0, v0 -; SI-NEXT: v_or_b32_e32 v0, v0, v1 -; SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 -; SI-NEXT: s_endpgm -; -; VI-LABEL: fptrunc_v2f64_to_v2f16: -; VI: ; %bb.0: ; %entry -; VI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 -; VI-NEXT: s_mov_b32 s7, 0xf000 -; VI-NEXT: s_mov_b32 s6, -1 -; VI-NEXT: s_mov_b32 s10, s6 -; VI-NEXT: s_mov_b32 s11, s7 -; VI-NEXT: s_waitcnt lgkmcnt(0) -; VI-NEXT: s_mov_b32 s8, s2 -; VI-NEXT: s_mov_b32 s9, s3 -; VI-NEXT: buffer_load_dwordx4 v[0:3], off, s[8:11], 0 -; VI-NEXT: s_mov_b32 s4, s0 -; VI-NEXT: s_mov_b32 s5, s1 -; VI-NEXT: s_waitcnt vmcnt(0) -; VI-NEXT: v_cvt_f32_f64_e32 v2, v[2:3] -; VI-NEXT: v_cvt_f32_f64_e32 v0, v[0:1] -; VI-NEXT: v_cvt_f16_f32_sdwa v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD -; VI-NEXT: v_cvt_f16_f32_e32 v0, v0 -; VI-NEXT: v_or_b32_e32 v0, v0, v1 -; VI-NEXT: buffer_store_dword v0, off, s[4:7], 0 -; VI-NEXT: s_endpgm -; -; GFX9-LABEL: fptrunc_v2f64_to_v2f16: -; GFX9: ; %bb.0: ; %entry -; GFX9-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 -; GFX9-NEXT: s_mov_b32 s7, 0xf000 -; GFX9-NEXT: s_mov_b32 s6, -1 -; GFX9-NEXT: s_mov_b32 s10, s6 -; GFX9-NEXT: s_mov_b32 s11, s7 -; GFX9-NEXT: s_waitcnt lgkmcnt(0) -; GFX9-NEXT: s_mov_b32 s8, s2 -; GFX9-NEXT: s_mov_b32 s9, s3 -; GFX9-NEXT: buffer_load_dwordx4 v[0:3], off, s[8:11], 0 -; GFX9-NEXT: s_mov_b32 s4, s0 -; GFX9-NEXT: s_mov_b32 s5, s1 -; GFX9-NEXT: s_waitcnt vmcnt(0) -; GFX9-NEXT: v_cvt_f32_f64_e32 v2, v[2:3] -; GFX9-NEXT: v_cvt_f32_f64_e32 v0, v[0:1] -; GFX9-NEXT: v_cvt_f16_f32_e32 v1, v2 -; GFX9-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX9-NEXT: v_lshl_or_b32 v0, v1, 16, v0 -; GFX9-NEXT: buffer_store_dword v0, off, s[4:7], 0 -; GFX9-NEXT: s_endpgm -; -; GFX11-LABEL: fptrunc_v2f64_to_v2f16: -; GFX11: ; %bb.0: ; %entry -; GFX11-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 -; GFX11-NEXT: s_mov_b32 s6, -1 -; GFX11-NEXT: s_mov_b32 s7, 0x31016000 -; GFX11-NEXT: s_mov_b32 s10, s6 -; GFX11-NEXT: s_mov_b32 s11, s7 -; GFX11-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NEXT: s_mov_b32 s8, s2 -; GFX11-NEXT: s_mov_b32 s9, s3 -; GFX11-NEXT: s_mov_b32 s4, s0 -; GFX11-NEXT: buffer_load_b128 v[0:3], off, s[8:11], 0 -; GFX11-NEXT: s_mov_b32 s5, s1 -; GFX11-NEXT: s_waitcnt vmcnt(0) -; GFX11-NEXT: v_cvt_f32_f64_e32 v0, v[0:1] -; GFX11-NEXT: v_cvt_f32_f64_e32 v1, v[2:3] -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX11-NEXT: v_cvt_f16_f32_e32 v1, v1 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-NEXT: v_lshl_or_b32 v0, v1, 16, v0 -; GFX11-NEXT: buffer_store_b32 v0, off, s[4:7], 0 -; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) -; GFX11-NEXT: s_endpgm +; SI-SDAG-LABEL: fptrunc_v2f64_to_v2f16: +; SI-SDAG: ; %bb.0: ; %entry +; SI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 +; SI-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; SI-SDAG-NEXT: s_mov_b32 s6, -1 +; SI-SDAG-NEXT: s_mov_b32 s10, s6 +; SI-SDAG-NEXT: s_mov_b32 s11, s7 +; SI-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; SI-SDAG-NEXT: s_mov_b32 s8, s2 +; SI-SDAG-NEXT: s_mov_b32 s9, s3 +; SI-SDAG-NEXT: buffer_load_dwordx4 v[0:3], off, s[8:11], 0 +; SI-SDAG-NEXT: s_mov_b32 s4, s0 +; SI-SDAG-NEXT: s_mov_b32 s5, s1 +; SI-SDAG-NEXT: s_waitcnt vmcnt(0) +; SI-SDAG-NEXT: v_cvt_f32_f64_e32 v2, v[2:3] +; SI-SDAG-NEXT: v_cvt_f32_f64_e32 v0, v[0:1] +; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v1, v2 +; SI-SDAG-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; SI-SDAG-NEXT: v_or_b32_e32 v0, v0, v1 +; SI-SDAG-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; SI-SDAG-NEXT: s_endpgm +; +; VI-SDAG-LABEL: fptrunc_v2f64_to_v2f16: +; VI-SDAG: ; %bb.0: ; %entry +; VI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; VI-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; VI-SDAG-NEXT: s_mov_b32 s6, -1 +; VI-SDAG-NEXT: s_mov_b32 s10, s6 +; VI-SDAG-NEXT: s_mov_b32 s11, s7 +; VI-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; VI-SDAG-NEXT: s_mov_b32 s8, s2 +; VI-SDAG-NEXT: s_mov_b32 s9, s3 +; VI-SDAG-NEXT: buffer_load_dwordx4 v[0:3], off, s[8:11], 0 +; VI-SDAG-NEXT: s_mov_b32 s4, s0 +; VI-SDAG-NEXT: s_mov_b32 s5, s1 +; VI-SDAG-NEXT: s_waitcnt vmcnt(0) +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v2, v[2:3] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v0, v[0:1] +; VI-SDAG-NEXT: v_cvt_f16_f32_sdwa v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD +; VI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; VI-SDAG-NEXT: v_or_b32_e32 v0, v0, v1 +; VI-SDAG-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; VI-SDAG-NEXT: s_endpgm +; +; VI-GISEL-LABEL: fptrunc_v2f64_to_v2f16: +; VI-GISEL: ; %bb.0: ; %entry +; VI-GISEL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: s_load_dwordx4 s[4:7], s[2:3], 0x0 +; VI-GISEL-NEXT: s_movk_i32 s2, 0x7e00 +; VI-GISEL-NEXT: v_mov_b32_e32 v0, s0 +; VI-GISEL-NEXT: v_mov_b32_e32 v1, s1 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: s_bfe_u32 s3, s5, 0xb0014 +; VI-GISEL-NEXT: s_lshr_b32 s8, s5, 8 +; VI-GISEL-NEXT: s_and_b32 s9, s5, 0x1ff +; VI-GISEL-NEXT: s_addk_i32 s3, 0xfc10 +; VI-GISEL-NEXT: s_and_b32 s8, s8, 0xffe +; VI-GISEL-NEXT: s_or_b32 s4, s9, s4 +; VI-GISEL-NEXT: s_cmp_lg_u32 s4, 0 +; VI-GISEL-NEXT: s_cselect_b32 s4, 1, 0 +; VI-GISEL-NEXT: s_or_b32 s4, s8, s4 +; VI-GISEL-NEXT: s_cmp_lg_u32 s4, 0 +; VI-GISEL-NEXT: s_cselect_b32 s8, s2, 0x7c00 +; VI-GISEL-NEXT: s_sub_i32 s10, 1, s3 +; VI-GISEL-NEXT: s_lshl_b32 s9, s3, 12 +; VI-GISEL-NEXT: s_max_i32 s10, s10, 0 +; VI-GISEL-NEXT: s_or_b32 s9, s4, s9 +; VI-GISEL-NEXT: s_min_i32 s10, s10, 13 +; VI-GISEL-NEXT: s_bitset1_b32 s4, 12 +; VI-GISEL-NEXT: s_lshr_b32 s11, s4, s10 +; VI-GISEL-NEXT: s_lshl_b32 s10, s11, s10 +; VI-GISEL-NEXT: s_cmp_lg_u32 s10, s4 +; VI-GISEL-NEXT: s_cselect_b32 s4, 1, 0 +; VI-GISEL-NEXT: s_or_b32 s4, s11, s4 +; VI-GISEL-NEXT: s_cmp_lt_i32 s3, 1 +; VI-GISEL-NEXT: s_cselect_b32 s4, s4, s9 +; VI-GISEL-NEXT: s_and_b32 s9, s4, 7 +; VI-GISEL-NEXT: s_lshr_b32 s4, s4, 2 +; VI-GISEL-NEXT: s_cmp_eq_u32 s9, 3 +; VI-GISEL-NEXT: s_cselect_b32 s10, 1, 0 +; VI-GISEL-NEXT: s_cmp_gt_i32 s9, 5 +; VI-GISEL-NEXT: s_cselect_b32 s9, 1, 0 +; VI-GISEL-NEXT: s_or_b32 s9, s10, s9 +; VI-GISEL-NEXT: s_and_b32 s9, s9, 1 +; VI-GISEL-NEXT: s_add_i32 s4, s4, s9 +; VI-GISEL-NEXT: s_cmp_gt_i32 s3, 30 +; VI-GISEL-NEXT: s_cselect_b32 s4, 0x7c00, s4 +; VI-GISEL-NEXT: s_cmpk_eq_i32 s3, 0x40f +; VI-GISEL-NEXT: s_cselect_b32 s3, s8, s4 +; VI-GISEL-NEXT: s_lshr_b32 s4, s5, 16 +; VI-GISEL-NEXT: s_and_b32 s4, s4, 0x8000 +; VI-GISEL-NEXT: s_or_b32 s3, s4, s3 +; VI-GISEL-NEXT: s_bfe_u32 s4, s7, 0xb0014 +; VI-GISEL-NEXT: s_lshr_b32 s5, s7, 8 +; VI-GISEL-NEXT: s_and_b32 s8, s7, 0x1ff +; VI-GISEL-NEXT: s_addk_i32 s4, 0xfc10 +; VI-GISEL-NEXT: s_and_b32 s5, s5, 0xffe +; VI-GISEL-NEXT: s_or_b32 s6, s8, s6 +; VI-GISEL-NEXT: s_cmp_lg_u32 s6, 0 +; VI-GISEL-NEXT: s_cselect_b32 s6, 1, 0 +; VI-GISEL-NEXT: s_or_b32 s5, s5, s6 +; VI-GISEL-NEXT: s_cmp_lg_u32 s5, 0 +; VI-GISEL-NEXT: s_cselect_b32 s2, s2, 0x7c00 +; VI-GISEL-NEXT: s_sub_i32 s8, 1, s4 +; VI-GISEL-NEXT: s_lshl_b32 s6, s4, 12 +; VI-GISEL-NEXT: s_max_i32 s8, s8, 0 +; VI-GISEL-NEXT: s_or_b32 s6, s5, s6 +; VI-GISEL-NEXT: s_min_i32 s8, s8, 13 +; VI-GISEL-NEXT: s_bitset1_b32 s5, 12 +; VI-GISEL-NEXT: s_lshr_b32 s9, s5, s8 +; VI-GISEL-NEXT: s_lshl_b32 s8, s9, s8 +; VI-GISEL-NEXT: s_cmp_lg_u32 s8, s5 +; VI-GISEL-NEXT: s_cselect_b32 s5, 1, 0 +; VI-GISEL-NEXT: s_or_b32 s5, s9, s5 +; VI-GISEL-NEXT: s_cmp_lt_i32 s4, 1 +; VI-GISEL-NEXT: s_cselect_b32 s5, s5, s6 +; VI-GISEL-NEXT: s_and_b32 s6, s5, 7 +; VI-GISEL-NEXT: s_lshr_b32 s5, s5, 2 +; VI-GISEL-NEXT: s_cmp_eq_u32 s6, 3 +; VI-GISEL-NEXT: s_cselect_b32 s8, 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, s8, s6 +; VI-GISEL-NEXT: s_and_b32 s6, s6, 1 +; VI-GISEL-NEXT: s_add_i32 s5, s5, s6 +; VI-GISEL-NEXT: s_cmp_gt_i32 s4, 30 +; VI-GISEL-NEXT: s_cselect_b32 s5, 0x7c00, s5 +; VI-GISEL-NEXT: s_cmpk_eq_i32 s4, 0x40f +; VI-GISEL-NEXT: s_cselect_b32 s2, s2, s5 +; VI-GISEL-NEXT: s_lshr_b32 s4, s7, 16 +; VI-GISEL-NEXT: s_and_b32 s4, s4, 0x8000 +; VI-GISEL-NEXT: s_or_b32 s2, s4, s2 +; VI-GISEL-NEXT: s_and_b32 s2, s2, 0xffff +; VI-GISEL-NEXT: s_and_b32 s3, s3, 0xffff +; VI-GISEL-NEXT: s_lshl_b32 s2, s2, 16 +; VI-GISEL-NEXT: s_or_b32 s2, s3, s2 +; VI-GISEL-NEXT: v_mov_b32_e32 v2, s2 +; VI-GISEL-NEXT: flat_store_dword v[0:1], v2 +; VI-GISEL-NEXT: s_endpgm +; +; GFX9-SDAG-LABEL: fptrunc_v2f64_to_v2f16: +; GFX9-SDAG: ; %bb.0: ; %entry +; GFX9-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; GFX9-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; GFX9-SDAG-NEXT: s_mov_b32 s6, -1 +; GFX9-SDAG-NEXT: s_mov_b32 s10, s6 +; GFX9-SDAG-NEXT: s_mov_b32 s11, s7 +; GFX9-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-SDAG-NEXT: s_mov_b32 s8, s2 +; GFX9-SDAG-NEXT: s_mov_b32 s9, s3 +; GFX9-SDAG-NEXT: buffer_load_dwordx4 v[0:3], off, s[8:11], 0 +; GFX9-SDAG-NEXT: s_mov_b32 s4, s0 +; GFX9-SDAG-NEXT: s_mov_b32 s5, s1 +; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) +; GFX9-SDAG-NEXT: v_cvt_f32_f64_e32 v2, v[2:3] +; GFX9-SDAG-NEXT: v_cvt_f32_f64_e32 v0, v[0:1] +; GFX9-SDAG-NEXT: v_cvt_f16_f32_e32 v1, v2 +; GFX9-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; GFX9-SDAG-NEXT: v_lshl_or_b32 v0, v1, 16, v0 +; GFX9-SDAG-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; GFX9-SDAG-NEXT: s_endpgm +; +; GFX9-GISEL-LABEL: fptrunc_v2f64_to_v2f16: +; GFX9-GISEL: ; %bb.0: ; %entry +; GFX9-GISEL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; GFX9-GISEL-NEXT: v_mov_b32_e32 v1, 0 +; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-GISEL-NEXT: s_load_dwordx4 s[4:7], s[2:3], 0x0 +; GFX9-GISEL-NEXT: s_movk_i32 s2, 0x7e00 +; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-GISEL-NEXT: s_bfe_u32 s3, s5, 0xb0014 +; GFX9-GISEL-NEXT: s_lshr_b32 s8, s5, 8 +; GFX9-GISEL-NEXT: s_and_b32 s9, s5, 0x1ff +; GFX9-GISEL-NEXT: s_addk_i32 s3, 0xfc10 +; GFX9-GISEL-NEXT: s_and_b32 s8, s8, 0xffe +; GFX9-GISEL-NEXT: s_or_b32 s4, s9, s4 +; GFX9-GISEL-NEXT: s_cmp_lg_u32 s4, 0 +; GFX9-GISEL-NEXT: s_cselect_b32 s4, 1, 0 +; GFX9-GISEL-NEXT: s_or_b32 s4, s8, s4 +; GFX9-GISEL-NEXT: s_cmp_lg_u32 s4, 0 +; GFX9-GISEL-NEXT: s_cselect_b32 s8, s2, 0x7c00 +; GFX9-GISEL-NEXT: s_sub_i32 s10, 1, s3 +; GFX9-GISEL-NEXT: s_lshl_b32 s9, s3, 12 +; GFX9-GISEL-NEXT: s_max_i32 s10, s10, 0 +; GFX9-GISEL-NEXT: s_or_b32 s9, s4, s9 +; GFX9-GISEL-NEXT: s_min_i32 s10, s10, 13 +; GFX9-GISEL-NEXT: s_bitset1_b32 s4, 12 +; GFX9-GISEL-NEXT: s_lshr_b32 s11, s4, s10 +; GFX9-GISEL-NEXT: s_lshl_b32 s10, s11, s10 +; GFX9-GISEL-NEXT: s_cmp_lg_u32 s10, s4 +; GFX9-GISEL-NEXT: s_cselect_b32 s4, 1, 0 +; GFX9-GISEL-NEXT: s_or_b32 s4, s11, s4 +; GFX9-GISEL-NEXT: s_cmp_lt_i32 s3, 1 +; GFX9-GISEL-NEXT: s_cselect_b32 s4, s4, s9 +; GFX9-GISEL-NEXT: s_and_b32 s9, s4, 7 +; GFX9-GISEL-NEXT: s_lshr_b32 s4, s4, 2 +; GFX9-GISEL-NEXT: s_cmp_eq_u32 s9, 3 +; GFX9-GISEL-NEXT: s_cselect_b32 s10, 1, 0 +; GFX9-GISEL-NEXT: s_cmp_gt_i32 s9, 5 +; GFX9-GISEL-NEXT: s_cselect_b32 s9, 1, 0 +; GFX9-GISEL-NEXT: s_or_b32 s9, s10, s9 +; GFX9-GISEL-NEXT: s_and_b32 s9, s9, 1 +; GFX9-GISEL-NEXT: s_add_i32 s4, s4, s9 +; GFX9-GISEL-NEXT: s_cmp_gt_i32 s3, 30 +; GFX9-GISEL-NEXT: s_cselect_b32 s4, 0x7c00, s4 +; GFX9-GISEL-NEXT: s_cmpk_eq_i32 s3, 0x40f +; GFX9-GISEL-NEXT: s_cselect_b32 s3, s8, s4 +; GFX9-GISEL-NEXT: s_lshr_b32 s4, s5, 16 +; GFX9-GISEL-NEXT: s_and_b32 s4, s4, 0x8000 +; GFX9-GISEL-NEXT: s_or_b32 s3, s4, s3 +; GFX9-GISEL-NEXT: s_bfe_u32 s4, s7, 0xb0014 +; GFX9-GISEL-NEXT: s_lshr_b32 s5, s7, 8 +; GFX9-GISEL-NEXT: s_and_b32 s8, s7, 0x1ff +; GFX9-GISEL-NEXT: s_addk_i32 s4, 0xfc10 +; GFX9-GISEL-NEXT: s_and_b32 s5, s5, 0xffe +; GFX9-GISEL-NEXT: s_or_b32 s6, s8, s6 +; GFX9-GISEL-NEXT: s_cmp_lg_u32 s6, 0 +; GFX9-GISEL-NEXT: s_cselect_b32 s6, 1, 0 +; GFX9-GISEL-NEXT: s_or_b32 s5, s5, s6 +; GFX9-GISEL-NEXT: s_cmp_lg_u32 s5, 0 +; GFX9-GISEL-NEXT: s_cselect_b32 s2, s2, 0x7c00 +; GFX9-GISEL-NEXT: s_sub_i32 s8, 1, s4 +; GFX9-GISEL-NEXT: s_lshl_b32 s6, s4, 12 +; GFX9-GISEL-NEXT: s_max_i32 s8, s8, 0 +; GFX9-GISEL-NEXT: s_or_b32 s6, s5, s6 +; GFX9-GISEL-NEXT: s_min_i32 s8, s8, 13 +; GFX9-GISEL-NEXT: s_bitset1_b32 s5, 12 +; GFX9-GISEL-NEXT: s_lshr_b32 s9, s5, s8 +; GFX9-GISEL-NEXT: s_lshl_b32 s8, s9, s8 +; GFX9-GISEL-NEXT: s_cmp_lg_u32 s8, s5 +; GFX9-GISEL-NEXT: s_cselect_b32 s5, 1, 0 +; GFX9-GISEL-NEXT: s_or_b32 s5, s9, s5 +; GFX9-GISEL-NEXT: s_cmp_lt_i32 s4, 1 +; GFX9-GISEL-NEXT: s_cselect_b32 s5, s5, s6 +; GFX9-GISEL-NEXT: s_and_b32 s6, s5, 7 +; GFX9-GISEL-NEXT: s_lshr_b32 s5, s5, 2 +; GFX9-GISEL-NEXT: s_cmp_eq_u32 s6, 3 +; GFX9-GISEL-NEXT: s_cselect_b32 s8, 1, 0 +; GFX9-GISEL-NEXT: s_cmp_gt_i32 s6, 5 +; GFX9-GISEL-NEXT: s_cselect_b32 s6, 1, 0 +; GFX9-GISEL-NEXT: s_or_b32 s6, s8, s6 +; GFX9-GISEL-NEXT: s_and_b32 s6, s6, 1 +; GFX9-GISEL-NEXT: s_add_i32 s5, s5, s6 +; GFX9-GISEL-NEXT: s_cmp_gt_i32 s4, 30 +; GFX9-GISEL-NEXT: s_cselect_b32 s5, 0x7c00, s5 +; GFX9-GISEL-NEXT: s_cmpk_eq_i32 s4, 0x40f +; GFX9-GISEL-NEXT: s_cselect_b32 s2, s2, s5 +; GFX9-GISEL-NEXT: s_lshr_b32 s4, s7, 16 +; GFX9-GISEL-NEXT: s_and_b32 s4, s4, 0x8000 +; GFX9-GISEL-NEXT: s_or_b32 s2, s4, s2 +; GFX9-GISEL-NEXT: s_pack_ll_b32_b16 s2, s3, s2 +; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, s2 +; GFX9-GISEL-NEXT: global_store_dword v1, v0, s[0:1] +; GFX9-GISEL-NEXT: s_endpgm +; +; GFX11-SDAG-LABEL: fptrunc_v2f64_to_v2f16: +; GFX11-SDAG: ; %bb.0: ; %entry +; GFX11-SDAG-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 +; GFX11-SDAG-NEXT: s_mov_b32 s6, -1 +; GFX11-SDAG-NEXT: s_mov_b32 s7, 0x31016000 +; GFX11-SDAG-NEXT: s_mov_b32 s10, s6 +; GFX11-SDAG-NEXT: s_mov_b32 s11, s7 +; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-SDAG-NEXT: s_mov_b32 s8, s2 +; GFX11-SDAG-NEXT: s_mov_b32 s9, s3 +; GFX11-SDAG-NEXT: s_mov_b32 s4, s0 +; GFX11-SDAG-NEXT: buffer_load_b128 v[0:3], off, s[8:11], 0 +; GFX11-SDAG-NEXT: s_mov_b32 s5, s1 +; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v0, v[0:1] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v1, v[2:3] +; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; GFX11-SDAG-NEXT: v_cvt_f16_f32_e32 v1, v1 +; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0 +; GFX11-SDAG-NEXT: v_lshl_or_b32 v0, v1, 16, v0 +; GFX11-SDAG-NEXT: buffer_store_b32 v0, off, s[4:7], 0 +; GFX11-SDAG-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-SDAG-NEXT: s_endpgm +; +; GFX11-GISEL-LABEL: fptrunc_v2f64_to_v2f16: +; GFX11-GISEL: ; %bb.0: ; %entry +; GFX11-GISEL-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 +; GFX11-GISEL-NEXT: v_mov_b32_e32 v1, 0 +; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-GISEL-NEXT: s_load_b128 s[4:7], s[2:3], 0x0 +; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-GISEL-NEXT: s_and_b32 s8, s5, 0x1ff +; GFX11-GISEL-NEXT: s_bfe_u32 s2, s5, 0xb0014 +; GFX11-GISEL-NEXT: s_lshr_b32 s3, s5, 8 +; GFX11-GISEL-NEXT: s_or_b32 s4, s8, s4 +; GFX11-GISEL-NEXT: s_addk_i32 s2, 0xfc10 +; GFX11-GISEL-NEXT: s_and_b32 s3, s3, 0xffe +; GFX11-GISEL-NEXT: s_cmp_lg_u32 s4, 0 +; GFX11-GISEL-NEXT: s_cselect_b32 s4, 1, 0 +; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) +; GFX11-GISEL-NEXT: s_or_b32 s3, s3, s4 +; GFX11-GISEL-NEXT: s_movk_i32 s4, 0x7e00 +; GFX11-GISEL-NEXT: s_cmp_lg_u32 s3, 0 +; GFX11-GISEL-NEXT: s_cselect_b32 s8, s4, 0x7c00 +; GFX11-GISEL-NEXT: s_sub_i32 s9, 1, s2 +; GFX11-GISEL-NEXT: s_or_b32 s10, s3, 0x1000 +; GFX11-GISEL-NEXT: s_max_i32 s9, s9, 0 +; GFX11-GISEL-NEXT: s_lshl_b32 s12, s2, 12 +; GFX11-GISEL-NEXT: s_min_i32 s9, s9, 13 +; GFX11-GISEL-NEXT: s_or_b32 s3, s3, s12 +; GFX11-GISEL-NEXT: s_lshr_b32 s11, s10, s9 +; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) +; GFX11-GISEL-NEXT: s_lshl_b32 s9, s11, s9 +; GFX11-GISEL-NEXT: s_cmp_lg_u32 s9, s10 +; GFX11-GISEL-NEXT: s_cselect_b32 s9, 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 s9, s11, s9 +; GFX11-GISEL-NEXT: s_cmp_lt_i32 s2, 1 +; GFX11-GISEL-NEXT: s_cselect_b32 s3, s9, s3 +; GFX11-GISEL-NEXT: s_and_b32 s9, s3, 7 +; GFX11-GISEL-NEXT: s_lshr_b32 s3, s3, 2 +; GFX11-GISEL-NEXT: s_cmp_eq_u32 s9, 3 +; GFX11-GISEL-NEXT: s_cselect_b32 s10, 1, 0 +; GFX11-GISEL-NEXT: s_cmp_gt_i32 s9, 5 +; GFX11-GISEL-NEXT: s_cselect_b32 s9, 1, 0 +; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) +; GFX11-GISEL-NEXT: s_or_b32 s9, s10, s9 +; GFX11-GISEL-NEXT: s_and_b32 s9, s9, 1 +; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) +; GFX11-GISEL-NEXT: s_add_i32 s3, s3, s9 +; GFX11-GISEL-NEXT: s_cmp_gt_i32 s2, 30 +; GFX11-GISEL-NEXT: s_cselect_b32 s3, 0x7c00, s3 +; GFX11-GISEL-NEXT: s_cmpk_eq_i32 s2, 0x40f +; GFX11-GISEL-NEXT: s_cselect_b32 s2, s8, s3 +; GFX11-GISEL-NEXT: s_lshr_b32 s3, s5, 16 +; GFX11-GISEL-NEXT: s_and_b32 s9, s7, 0x1ff +; GFX11-GISEL-NEXT: s_bfe_u32 s5, s7, 0xb0014 +; GFX11-GISEL-NEXT: s_lshr_b32 s8, s7, 8 +; GFX11-GISEL-NEXT: s_and_b32 s3, s3, 0x8000 +; GFX11-GISEL-NEXT: s_or_b32 s6, s9, s6 +; GFX11-GISEL-NEXT: s_addk_i32 s5, 0xfc10 +; GFX11-GISEL-NEXT: s_and_b32 s8, s8, 0xffe +; GFX11-GISEL-NEXT: s_or_b32 s2, s3, s2 +; GFX11-GISEL-NEXT: s_cmp_lg_u32 s6, 0 +; GFX11-GISEL-NEXT: s_cselect_b32 s3, 1, 0 +; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) +; GFX11-GISEL-NEXT: s_or_b32 s3, s8, s3 +; GFX11-GISEL-NEXT: s_cmp_lg_u32 s3, 0 +; GFX11-GISEL-NEXT: s_cselect_b32 s4, s4, 0x7c00 +; GFX11-GISEL-NEXT: s_sub_i32 s6, 1, s5 +; GFX11-GISEL-NEXT: s_or_b32 s8, s3, 0x1000 +; GFX11-GISEL-NEXT: s_max_i32 s6, s6, 0 +; GFX11-GISEL-NEXT: s_lshl_b32 s10, s5, 12 +; GFX11-GISEL-NEXT: s_min_i32 s6, s6, 13 +; GFX11-GISEL-NEXT: s_or_b32 s3, s3, s10 +; GFX11-GISEL-NEXT: s_lshr_b32 s9, s8, s6 +; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) +; GFX11-GISEL-NEXT: s_lshl_b32 s6, s9, s6 +; 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 s5, 1 +; GFX11-GISEL-NEXT: s_cselect_b32 s3, s6, s3 +; GFX11-GISEL-NEXT: s_and_b32 s6, s3, 7 +; GFX11-GISEL-NEXT: s_lshr_b32 s3, s3, 2 +; GFX11-GISEL-NEXT: s_cmp_eq_u32 s6, 3 +; GFX11-GISEL-NEXT: s_cselect_b32 s8, 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, s8, s6 +; GFX11-GISEL-NEXT: s_and_b32 s6, s6, 1 +; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) +; GFX11-GISEL-NEXT: s_add_i32 s3, s3, s6 +; GFX11-GISEL-NEXT: s_cmp_gt_i32 s5, 30 +; GFX11-GISEL-NEXT: s_cselect_b32 s3, 0x7c00, s3 +; GFX11-GISEL-NEXT: s_cmpk_eq_i32 s5, 0x40f +; GFX11-GISEL-NEXT: s_cselect_b32 s3, s4, s3 +; GFX11-GISEL-NEXT: s_lshr_b32 s4, s7, 16 +; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) +; GFX11-GISEL-NEXT: s_and_b32 s4, s4, 0x8000 +; GFX11-GISEL-NEXT: s_or_b32 s3, s4, s3 +; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) +; GFX11-GISEL-NEXT: s_pack_ll_b32_b16 s2, s2, s3 +; GFX11-GISEL-NEXT: v_mov_b32_e32 v0, s2 +; GFX11-GISEL-NEXT: global_store_b32 v1, v0, s[0:1] +; GFX11-GISEL-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-GISEL-NEXT: s_endpgm ptr addrspace(1) %r, ptr addrspace(1) %a) { entry: @@ -368,78 +915,115 @@ } define amdgpu_kernel void @fneg_fptrunc_f32_to_f16( -; SI-LABEL: fneg_fptrunc_f32_to_f16: -; SI: ; %bb.0: ; %entry -; SI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 -; SI-NEXT: s_mov_b32 s7, 0xf000 -; SI-NEXT: s_mov_b32 s6, -1 -; SI-NEXT: s_mov_b32 s10, s6 -; SI-NEXT: s_mov_b32 s11, s7 -; SI-NEXT: s_waitcnt lgkmcnt(0) -; SI-NEXT: s_mov_b32 s8, s2 -; SI-NEXT: s_mov_b32 s9, s3 -; SI-NEXT: buffer_load_dword v0, off, s[8:11], 0 -; SI-NEXT: s_mov_b32 s4, s0 -; SI-NEXT: s_mov_b32 s5, s1 -; SI-NEXT: s_waitcnt vmcnt(0) -; SI-NEXT: v_cvt_f16_f32_e64 v0, -v0 -; SI-NEXT: buffer_store_short v0, off, s[4:7], 0 -; SI-NEXT: s_endpgm -; -; VI-LABEL: fneg_fptrunc_f32_to_f16: -; VI: ; %bb.0: ; %entry -; VI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 -; VI-NEXT: s_mov_b32 s7, 0xf000 -; VI-NEXT: s_mov_b32 s6, -1 -; VI-NEXT: s_mov_b32 s10, s6 -; VI-NEXT: s_mov_b32 s11, s7 -; VI-NEXT: s_waitcnt lgkmcnt(0) -; VI-NEXT: s_mov_b32 s8, s2 -; VI-NEXT: s_mov_b32 s9, s3 -; VI-NEXT: buffer_load_dword v0, off, s[8:11], 0 -; VI-NEXT: s_mov_b32 s4, s0 -; VI-NEXT: s_mov_b32 s5, s1 -; VI-NEXT: s_waitcnt vmcnt(0) -; VI-NEXT: v_cvt_f16_f32_e64 v0, -v0 -; VI-NEXT: buffer_store_short v0, off, s[4:7], 0 -; VI-NEXT: s_endpgm -; -; GFX9-LABEL: fneg_fptrunc_f32_to_f16: -; GFX9: ; %bb.0: ; %entry -; GFX9-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 -; GFX9-NEXT: s_mov_b32 s7, 0xf000 -; GFX9-NEXT: s_mov_b32 s6, -1 -; GFX9-NEXT: s_mov_b32 s10, s6 -; GFX9-NEXT: s_mov_b32 s11, s7 -; GFX9-NEXT: s_waitcnt lgkmcnt(0) -; GFX9-NEXT: s_mov_b32 s8, s2 -; GFX9-NEXT: s_mov_b32 s9, s3 -; GFX9-NEXT: buffer_load_dword v0, off, s[8:11], 0 -; GFX9-NEXT: s_mov_b32 s4, s0 -; GFX9-NEXT: s_mov_b32 s5, s1 -; GFX9-NEXT: s_waitcnt vmcnt(0) -; GFX9-NEXT: v_cvt_f16_f32_e64 v0, -v0 -; GFX9-NEXT: buffer_store_short v0, off, s[4:7], 0 -; GFX9-NEXT: s_endpgm -; -; GFX11-LABEL: fneg_fptrunc_f32_to_f16: -; GFX11: ; %bb.0: ; %entry -; GFX11-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 -; GFX11-NEXT: s_mov_b32 s6, -1 -; GFX11-NEXT: s_mov_b32 s7, 0x31016000 -; GFX11-NEXT: s_mov_b32 s10, s6 -; GFX11-NEXT: s_mov_b32 s11, s7 -; GFX11-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NEXT: s_mov_b32 s8, s2 -; GFX11-NEXT: s_mov_b32 s9, s3 -; GFX11-NEXT: s_mov_b32 s4, s0 -; GFX11-NEXT: buffer_load_b32 v0, off, s[8:11], 0 -; GFX11-NEXT: s_mov_b32 s5, s1 -; GFX11-NEXT: s_waitcnt vmcnt(0) -; GFX11-NEXT: v_cvt_f16_f32_e64 v0, -v0 -; GFX11-NEXT: buffer_store_b16 v0, off, s[4:7], 0 -; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) -; GFX11-NEXT: s_endpgm +; SI-SDAG-LABEL: fneg_fptrunc_f32_to_f16: +; SI-SDAG: ; %bb.0: ; %entry +; SI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 +; SI-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; SI-SDAG-NEXT: s_mov_b32 s6, -1 +; SI-SDAG-NEXT: s_mov_b32 s10, s6 +; SI-SDAG-NEXT: s_mov_b32 s11, s7 +; SI-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; SI-SDAG-NEXT: s_mov_b32 s8, s2 +; SI-SDAG-NEXT: s_mov_b32 s9, s3 +; SI-SDAG-NEXT: buffer_load_dword v0, off, s[8:11], 0 +; SI-SDAG-NEXT: s_mov_b32 s4, s0 +; SI-SDAG-NEXT: s_mov_b32 s5, s1 +; SI-SDAG-NEXT: s_waitcnt vmcnt(0) +; SI-SDAG-NEXT: v_cvt_f16_f32_e64 v0, -v0 +; SI-SDAG-NEXT: buffer_store_short v0, off, s[4:7], 0 +; SI-SDAG-NEXT: s_endpgm +; +; VI-SDAG-LABEL: fneg_fptrunc_f32_to_f16: +; VI-SDAG: ; %bb.0: ; %entry +; VI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; VI-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; VI-SDAG-NEXT: s_mov_b32 s6, -1 +; VI-SDAG-NEXT: s_mov_b32 s10, s6 +; VI-SDAG-NEXT: s_mov_b32 s11, s7 +; VI-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; VI-SDAG-NEXT: s_mov_b32 s8, s2 +; VI-SDAG-NEXT: s_mov_b32 s9, s3 +; VI-SDAG-NEXT: buffer_load_dword v0, off, s[8:11], 0 +; VI-SDAG-NEXT: s_mov_b32 s4, s0 +; VI-SDAG-NEXT: s_mov_b32 s5, s1 +; VI-SDAG-NEXT: s_waitcnt vmcnt(0) +; VI-SDAG-NEXT: v_cvt_f16_f32_e64 v0, -v0 +; VI-SDAG-NEXT: buffer_store_short v0, off, s[4:7], 0 +; VI-SDAG-NEXT: s_endpgm +; +; VI-GISEL-LABEL: fneg_fptrunc_f32_to_f16: +; VI-GISEL: ; %bb.0: ; %entry +; VI-GISEL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: s_load_dword s2, s[2:3], 0x0 +; VI-GISEL-NEXT: s_mov_b32 s3, 0xf000 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: v_cvt_f16_f32_e64 v0, -s2 +; VI-GISEL-NEXT: s_mov_b32 s2, -1 +; VI-GISEL-NEXT: buffer_store_short v0, off, s[0:3], 0 +; VI-GISEL-NEXT: s_endpgm +; +; GFX9-SDAG-LABEL: fneg_fptrunc_f32_to_f16: +; GFX9-SDAG: ; %bb.0: ; %entry +; GFX9-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; GFX9-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; GFX9-SDAG-NEXT: s_mov_b32 s6, -1 +; GFX9-SDAG-NEXT: s_mov_b32 s10, s6 +; GFX9-SDAG-NEXT: s_mov_b32 s11, s7 +; GFX9-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-SDAG-NEXT: s_mov_b32 s8, s2 +; GFX9-SDAG-NEXT: s_mov_b32 s9, s3 +; GFX9-SDAG-NEXT: buffer_load_dword v0, off, s[8:11], 0 +; GFX9-SDAG-NEXT: s_mov_b32 s4, s0 +; GFX9-SDAG-NEXT: s_mov_b32 s5, s1 +; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) +; GFX9-SDAG-NEXT: v_cvt_f16_f32_e64 v0, -v0 +; GFX9-SDAG-NEXT: buffer_store_short v0, off, s[4:7], 0 +; GFX9-SDAG-NEXT: s_endpgm +; +; GFX9-GISEL-LABEL: fneg_fptrunc_f32_to_f16: +; GFX9-GISEL: ; %bb.0: ; %entry +; GFX9-GISEL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-GISEL-NEXT: s_load_dword s2, s[2:3], 0x0 +; GFX9-GISEL-NEXT: s_mov_b32 s3, 0xf000 +; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-GISEL-NEXT: v_cvt_f16_f32_e64 v0, -s2 +; GFX9-GISEL-NEXT: s_mov_b32 s2, -1 +; GFX9-GISEL-NEXT: buffer_store_short v0, off, s[0:3], 0 +; GFX9-GISEL-NEXT: s_endpgm +; +; GFX11-SDAG-LABEL: fneg_fptrunc_f32_to_f16: +; GFX11-SDAG: ; %bb.0: ; %entry +; GFX11-SDAG-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 +; GFX11-SDAG-NEXT: s_mov_b32 s6, -1 +; GFX11-SDAG-NEXT: s_mov_b32 s7, 0x31016000 +; GFX11-SDAG-NEXT: s_mov_b32 s10, s6 +; GFX11-SDAG-NEXT: s_mov_b32 s11, s7 +; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-SDAG-NEXT: s_mov_b32 s8, s2 +; GFX11-SDAG-NEXT: s_mov_b32 s9, s3 +; GFX11-SDAG-NEXT: s_mov_b32 s4, s0 +; GFX11-SDAG-NEXT: buffer_load_b32 v0, off, s[8:11], 0 +; GFX11-SDAG-NEXT: s_mov_b32 s5, s1 +; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) +; GFX11-SDAG-NEXT: v_cvt_f16_f32_e64 v0, -v0 +; GFX11-SDAG-NEXT: buffer_store_b16 v0, off, s[4:7], 0 +; GFX11-SDAG-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-SDAG-NEXT: s_endpgm +; +; GFX11-GISEL-LABEL: fneg_fptrunc_f32_to_f16: +; GFX11-GISEL: ; %bb.0: ; %entry +; GFX11-GISEL-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 +; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-GISEL-NEXT: s_load_b32 s2, s[2:3], 0x0 +; GFX11-GISEL-NEXT: s_mov_b32 s3, 0x31016000 +; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-GISEL-NEXT: v_cvt_f16_f32_e64 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-GISEL-NEXT: s_endpgm ptr addrspace(1) %r, ptr addrspace(1) %a) { entry: @@ -451,78 +1035,115 @@ } define amdgpu_kernel void @fabs_fptrunc_f32_to_f16( -; SI-LABEL: fabs_fptrunc_f32_to_f16: -; SI: ; %bb.0: ; %entry -; SI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 -; SI-NEXT: s_mov_b32 s7, 0xf000 -; SI-NEXT: s_mov_b32 s6, -1 -; SI-NEXT: s_mov_b32 s10, s6 -; SI-NEXT: s_mov_b32 s11, s7 -; SI-NEXT: s_waitcnt lgkmcnt(0) -; SI-NEXT: s_mov_b32 s8, s2 -; SI-NEXT: s_mov_b32 s9, s3 -; SI-NEXT: buffer_load_dword v0, off, s[8:11], 0 -; SI-NEXT: s_mov_b32 s4, s0 -; SI-NEXT: s_mov_b32 s5, s1 -; SI-NEXT: s_waitcnt vmcnt(0) -; SI-NEXT: v_cvt_f16_f32_e64 v0, |v0| -; SI-NEXT: buffer_store_short v0, off, s[4:7], 0 -; SI-NEXT: s_endpgm -; -; VI-LABEL: fabs_fptrunc_f32_to_f16: -; VI: ; %bb.0: ; %entry -; VI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 -; VI-NEXT: s_mov_b32 s7, 0xf000 -; VI-NEXT: s_mov_b32 s6, -1 -; VI-NEXT: s_mov_b32 s10, s6 -; VI-NEXT: s_mov_b32 s11, s7 -; VI-NEXT: s_waitcnt lgkmcnt(0) -; VI-NEXT: s_mov_b32 s8, s2 -; VI-NEXT: s_mov_b32 s9, s3 -; VI-NEXT: buffer_load_dword v0, off, s[8:11], 0 -; VI-NEXT: s_mov_b32 s4, s0 -; VI-NEXT: s_mov_b32 s5, s1 -; VI-NEXT: s_waitcnt vmcnt(0) -; VI-NEXT: v_cvt_f16_f32_e64 v0, |v0| -; VI-NEXT: buffer_store_short v0, off, s[4:7], 0 -; VI-NEXT: s_endpgm -; -; GFX9-LABEL: fabs_fptrunc_f32_to_f16: -; GFX9: ; %bb.0: ; %entry -; GFX9-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 -; GFX9-NEXT: s_mov_b32 s7, 0xf000 -; GFX9-NEXT: s_mov_b32 s6, -1 -; GFX9-NEXT: s_mov_b32 s10, s6 -; GFX9-NEXT: s_mov_b32 s11, s7 -; GFX9-NEXT: s_waitcnt lgkmcnt(0) -; GFX9-NEXT: s_mov_b32 s8, s2 -; GFX9-NEXT: s_mov_b32 s9, s3 -; GFX9-NEXT: buffer_load_dword v0, off, s[8:11], 0 -; GFX9-NEXT: s_mov_b32 s4, s0 -; GFX9-NEXT: s_mov_b32 s5, s1 -; GFX9-NEXT: s_waitcnt vmcnt(0) -; GFX9-NEXT: v_cvt_f16_f32_e64 v0, |v0| -; GFX9-NEXT: buffer_store_short v0, off, s[4:7], 0 -; GFX9-NEXT: s_endpgm -; -; GFX11-LABEL: fabs_fptrunc_f32_to_f16: -; GFX11: ; %bb.0: ; %entry -; GFX11-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 -; GFX11-NEXT: s_mov_b32 s6, -1 -; GFX11-NEXT: s_mov_b32 s7, 0x31016000 -; GFX11-NEXT: s_mov_b32 s10, s6 -; GFX11-NEXT: s_mov_b32 s11, s7 -; GFX11-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NEXT: s_mov_b32 s8, s2 -; GFX11-NEXT: s_mov_b32 s9, s3 -; GFX11-NEXT: s_mov_b32 s4, s0 -; GFX11-NEXT: buffer_load_b32 v0, off, s[8:11], 0 -; GFX11-NEXT: s_mov_b32 s5, s1 -; GFX11-NEXT: s_waitcnt vmcnt(0) -; GFX11-NEXT: v_cvt_f16_f32_e64 v0, |v0| -; GFX11-NEXT: buffer_store_b16 v0, off, s[4:7], 0 -; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) -; GFX11-NEXT: s_endpgm +; SI-SDAG-LABEL: fabs_fptrunc_f32_to_f16: +; SI-SDAG: ; %bb.0: ; %entry +; SI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 +; SI-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; SI-SDAG-NEXT: s_mov_b32 s6, -1 +; SI-SDAG-NEXT: s_mov_b32 s10, s6 +; SI-SDAG-NEXT: s_mov_b32 s11, s7 +; SI-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; SI-SDAG-NEXT: s_mov_b32 s8, s2 +; SI-SDAG-NEXT: s_mov_b32 s9, s3 +; SI-SDAG-NEXT: buffer_load_dword v0, off, s[8:11], 0 +; SI-SDAG-NEXT: s_mov_b32 s4, s0 +; SI-SDAG-NEXT: s_mov_b32 s5, s1 +; SI-SDAG-NEXT: s_waitcnt vmcnt(0) +; SI-SDAG-NEXT: v_cvt_f16_f32_e64 v0, |v0| +; SI-SDAG-NEXT: buffer_store_short v0, off, s[4:7], 0 +; SI-SDAG-NEXT: s_endpgm +; +; VI-SDAG-LABEL: fabs_fptrunc_f32_to_f16: +; VI-SDAG: ; %bb.0: ; %entry +; VI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; VI-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; VI-SDAG-NEXT: s_mov_b32 s6, -1 +; VI-SDAG-NEXT: s_mov_b32 s10, s6 +; VI-SDAG-NEXT: s_mov_b32 s11, s7 +; VI-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; VI-SDAG-NEXT: s_mov_b32 s8, s2 +; VI-SDAG-NEXT: s_mov_b32 s9, s3 +; VI-SDAG-NEXT: buffer_load_dword v0, off, s[8:11], 0 +; VI-SDAG-NEXT: s_mov_b32 s4, s0 +; VI-SDAG-NEXT: s_mov_b32 s5, s1 +; VI-SDAG-NEXT: s_waitcnt vmcnt(0) +; VI-SDAG-NEXT: v_cvt_f16_f32_e64 v0, |v0| +; VI-SDAG-NEXT: buffer_store_short v0, off, s[4:7], 0 +; VI-SDAG-NEXT: s_endpgm +; +; VI-GISEL-LABEL: fabs_fptrunc_f32_to_f16: +; VI-GISEL: ; %bb.0: ; %entry +; VI-GISEL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: s_load_dword s2, s[2:3], 0x0 +; VI-GISEL-NEXT: s_mov_b32 s3, 0xf000 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: v_cvt_f16_f32_e64 v0, |s2| +; VI-GISEL-NEXT: s_mov_b32 s2, -1 +; VI-GISEL-NEXT: buffer_store_short v0, off, s[0:3], 0 +; VI-GISEL-NEXT: s_endpgm +; +; GFX9-SDAG-LABEL: fabs_fptrunc_f32_to_f16: +; GFX9-SDAG: ; %bb.0: ; %entry +; GFX9-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; GFX9-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; GFX9-SDAG-NEXT: s_mov_b32 s6, -1 +; GFX9-SDAG-NEXT: s_mov_b32 s10, s6 +; GFX9-SDAG-NEXT: s_mov_b32 s11, s7 +; GFX9-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-SDAG-NEXT: s_mov_b32 s8, s2 +; GFX9-SDAG-NEXT: s_mov_b32 s9, s3 +; GFX9-SDAG-NEXT: buffer_load_dword v0, off, s[8:11], 0 +; GFX9-SDAG-NEXT: s_mov_b32 s4, s0 +; GFX9-SDAG-NEXT: s_mov_b32 s5, s1 +; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) +; GFX9-SDAG-NEXT: v_cvt_f16_f32_e64 v0, |v0| +; GFX9-SDAG-NEXT: buffer_store_short v0, off, s[4:7], 0 +; GFX9-SDAG-NEXT: s_endpgm +; +; GFX9-GISEL-LABEL: fabs_fptrunc_f32_to_f16: +; GFX9-GISEL: ; %bb.0: ; %entry +; GFX9-GISEL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-GISEL-NEXT: s_load_dword s2, s[2:3], 0x0 +; GFX9-GISEL-NEXT: s_mov_b32 s3, 0xf000 +; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-GISEL-NEXT: v_cvt_f16_f32_e64 v0, |s2| +; GFX9-GISEL-NEXT: s_mov_b32 s2, -1 +; GFX9-GISEL-NEXT: buffer_store_short v0, off, s[0:3], 0 +; GFX9-GISEL-NEXT: s_endpgm +; +; GFX11-SDAG-LABEL: fabs_fptrunc_f32_to_f16: +; GFX11-SDAG: ; %bb.0: ; %entry +; GFX11-SDAG-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 +; GFX11-SDAG-NEXT: s_mov_b32 s6, -1 +; GFX11-SDAG-NEXT: s_mov_b32 s7, 0x31016000 +; GFX11-SDAG-NEXT: s_mov_b32 s10, s6 +; GFX11-SDAG-NEXT: s_mov_b32 s11, s7 +; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-SDAG-NEXT: s_mov_b32 s8, s2 +; GFX11-SDAG-NEXT: s_mov_b32 s9, s3 +; GFX11-SDAG-NEXT: s_mov_b32 s4, s0 +; GFX11-SDAG-NEXT: buffer_load_b32 v0, off, s[8:11], 0 +; GFX11-SDAG-NEXT: s_mov_b32 s5, s1 +; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) +; GFX11-SDAG-NEXT: v_cvt_f16_f32_e64 v0, |v0| +; GFX11-SDAG-NEXT: buffer_store_b16 v0, off, s[4:7], 0 +; GFX11-SDAG-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-SDAG-NEXT: s_endpgm +; +; GFX11-GISEL-LABEL: fabs_fptrunc_f32_to_f16: +; GFX11-GISEL: ; %bb.0: ; %entry +; GFX11-GISEL-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 +; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-GISEL-NEXT: s_load_b32 s2, s[2:3], 0x0 +; GFX11-GISEL-NEXT: s_mov_b32 s3, 0x31016000 +; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-GISEL-NEXT: v_cvt_f16_f32_e64 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-GISEL-NEXT: s_endpgm ptr addrspace(1) %r, ptr addrspace(1) %a) { entry: @@ -534,78 +1155,115 @@ } define amdgpu_kernel void @fneg_fabs_fptrunc_f32_to_f16( -; SI-LABEL: fneg_fabs_fptrunc_f32_to_f16: -; SI: ; %bb.0: ; %entry -; SI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 -; SI-NEXT: s_mov_b32 s7, 0xf000 -; SI-NEXT: s_mov_b32 s6, -1 -; SI-NEXT: s_mov_b32 s10, s6 -; SI-NEXT: s_mov_b32 s11, s7 -; SI-NEXT: s_waitcnt lgkmcnt(0) -; SI-NEXT: s_mov_b32 s8, s2 -; SI-NEXT: s_mov_b32 s9, s3 -; SI-NEXT: buffer_load_dword v0, off, s[8:11], 0 -; SI-NEXT: s_mov_b32 s4, s0 -; SI-NEXT: s_mov_b32 s5, s1 -; SI-NEXT: s_waitcnt vmcnt(0) -; SI-NEXT: v_cvt_f16_f32_e64 v0, -|v0| -; SI-NEXT: buffer_store_short v0, off, s[4:7], 0 -; SI-NEXT: s_endpgm -; -; VI-LABEL: fneg_fabs_fptrunc_f32_to_f16: -; VI: ; %bb.0: ; %entry -; VI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 -; VI-NEXT: s_mov_b32 s7, 0xf000 -; VI-NEXT: s_mov_b32 s6, -1 -; VI-NEXT: s_mov_b32 s10, s6 -; VI-NEXT: s_mov_b32 s11, s7 -; VI-NEXT: s_waitcnt lgkmcnt(0) -; VI-NEXT: s_mov_b32 s8, s2 -; VI-NEXT: s_mov_b32 s9, s3 -; VI-NEXT: buffer_load_dword v0, off, s[8:11], 0 -; VI-NEXT: s_mov_b32 s4, s0 -; VI-NEXT: s_mov_b32 s5, s1 -; VI-NEXT: s_waitcnt vmcnt(0) -; VI-NEXT: v_cvt_f16_f32_e64 v0, -|v0| -; VI-NEXT: buffer_store_short v0, off, s[4:7], 0 -; VI-NEXT: s_endpgm -; -; GFX9-LABEL: fneg_fabs_fptrunc_f32_to_f16: -; GFX9: ; %bb.0: ; %entry -; GFX9-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 -; GFX9-NEXT: s_mov_b32 s7, 0xf000 -; GFX9-NEXT: s_mov_b32 s6, -1 -; GFX9-NEXT: s_mov_b32 s10, s6 -; GFX9-NEXT: s_mov_b32 s11, s7 -; GFX9-NEXT: s_waitcnt lgkmcnt(0) -; GFX9-NEXT: s_mov_b32 s8, s2 -; GFX9-NEXT: s_mov_b32 s9, s3 -; GFX9-NEXT: buffer_load_dword v0, off, s[8:11], 0 -; GFX9-NEXT: s_mov_b32 s4, s0 -; GFX9-NEXT: s_mov_b32 s5, s1 -; GFX9-NEXT: s_waitcnt vmcnt(0) -; GFX9-NEXT: v_cvt_f16_f32_e64 v0, -|v0| -; GFX9-NEXT: buffer_store_short v0, off, s[4:7], 0 -; GFX9-NEXT: s_endpgm -; -; GFX11-LABEL: fneg_fabs_fptrunc_f32_to_f16: -; GFX11: ; %bb.0: ; %entry -; GFX11-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 -; GFX11-NEXT: s_mov_b32 s6, -1 -; GFX11-NEXT: s_mov_b32 s7, 0x31016000 -; GFX11-NEXT: s_mov_b32 s10, s6 -; GFX11-NEXT: s_mov_b32 s11, s7 -; GFX11-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NEXT: s_mov_b32 s8, s2 -; GFX11-NEXT: s_mov_b32 s9, s3 -; GFX11-NEXT: s_mov_b32 s4, s0 -; GFX11-NEXT: buffer_load_b32 v0, off, s[8:11], 0 -; GFX11-NEXT: s_mov_b32 s5, s1 -; GFX11-NEXT: s_waitcnt vmcnt(0) -; GFX11-NEXT: v_cvt_f16_f32_e64 v0, -|v0| -; GFX11-NEXT: buffer_store_b16 v0, off, s[4:7], 0 -; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) -; GFX11-NEXT: s_endpgm +; SI-SDAG-LABEL: fneg_fabs_fptrunc_f32_to_f16: +; SI-SDAG: ; %bb.0: ; %entry +; SI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 +; SI-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; SI-SDAG-NEXT: s_mov_b32 s6, -1 +; SI-SDAG-NEXT: s_mov_b32 s10, s6 +; SI-SDAG-NEXT: s_mov_b32 s11, s7 +; SI-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; SI-SDAG-NEXT: s_mov_b32 s8, s2 +; SI-SDAG-NEXT: s_mov_b32 s9, s3 +; SI-SDAG-NEXT: buffer_load_dword v0, off, s[8:11], 0 +; SI-SDAG-NEXT: s_mov_b32 s4, s0 +; SI-SDAG-NEXT: s_mov_b32 s5, s1 +; SI-SDAG-NEXT: s_waitcnt vmcnt(0) +; SI-SDAG-NEXT: v_cvt_f16_f32_e64 v0, -|v0| +; SI-SDAG-NEXT: buffer_store_short v0, off, s[4:7], 0 +; SI-SDAG-NEXT: s_endpgm +; +; VI-SDAG-LABEL: fneg_fabs_fptrunc_f32_to_f16: +; VI-SDAG: ; %bb.0: ; %entry +; VI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; VI-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; VI-SDAG-NEXT: s_mov_b32 s6, -1 +; VI-SDAG-NEXT: s_mov_b32 s10, s6 +; VI-SDAG-NEXT: s_mov_b32 s11, s7 +; VI-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; VI-SDAG-NEXT: s_mov_b32 s8, s2 +; VI-SDAG-NEXT: s_mov_b32 s9, s3 +; VI-SDAG-NEXT: buffer_load_dword v0, off, s[8:11], 0 +; VI-SDAG-NEXT: s_mov_b32 s4, s0 +; VI-SDAG-NEXT: s_mov_b32 s5, s1 +; VI-SDAG-NEXT: s_waitcnt vmcnt(0) +; VI-SDAG-NEXT: v_cvt_f16_f32_e64 v0, -|v0| +; VI-SDAG-NEXT: buffer_store_short v0, off, s[4:7], 0 +; VI-SDAG-NEXT: s_endpgm +; +; VI-GISEL-LABEL: fneg_fabs_fptrunc_f32_to_f16: +; VI-GISEL: ; %bb.0: ; %entry +; VI-GISEL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: s_load_dword s2, s[2:3], 0x0 +; VI-GISEL-NEXT: s_mov_b32 s3, 0xf000 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: v_cvt_f16_f32_e64 v0, -|s2| +; VI-GISEL-NEXT: s_mov_b32 s2, -1 +; VI-GISEL-NEXT: buffer_store_short v0, off, s[0:3], 0 +; VI-GISEL-NEXT: s_endpgm +; +; GFX9-SDAG-LABEL: fneg_fabs_fptrunc_f32_to_f16: +; GFX9-SDAG: ; %bb.0: ; %entry +; GFX9-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; GFX9-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; GFX9-SDAG-NEXT: s_mov_b32 s6, -1 +; GFX9-SDAG-NEXT: s_mov_b32 s10, s6 +; GFX9-SDAG-NEXT: s_mov_b32 s11, s7 +; GFX9-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-SDAG-NEXT: s_mov_b32 s8, s2 +; GFX9-SDAG-NEXT: s_mov_b32 s9, s3 +; GFX9-SDAG-NEXT: buffer_load_dword v0, off, s[8:11], 0 +; GFX9-SDAG-NEXT: s_mov_b32 s4, s0 +; GFX9-SDAG-NEXT: s_mov_b32 s5, s1 +; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) +; GFX9-SDAG-NEXT: v_cvt_f16_f32_e64 v0, -|v0| +; GFX9-SDAG-NEXT: buffer_store_short v0, off, s[4:7], 0 +; GFX9-SDAG-NEXT: s_endpgm +; +; GFX9-GISEL-LABEL: fneg_fabs_fptrunc_f32_to_f16: +; GFX9-GISEL: ; %bb.0: ; %entry +; GFX9-GISEL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-GISEL-NEXT: s_load_dword s2, s[2:3], 0x0 +; GFX9-GISEL-NEXT: s_mov_b32 s3, 0xf000 +; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-GISEL-NEXT: v_cvt_f16_f32_e64 v0, -|s2| +; GFX9-GISEL-NEXT: s_mov_b32 s2, -1 +; GFX9-GISEL-NEXT: buffer_store_short v0, off, s[0:3], 0 +; GFX9-GISEL-NEXT: s_endpgm +; +; GFX11-SDAG-LABEL: fneg_fabs_fptrunc_f32_to_f16: +; GFX11-SDAG: ; %bb.0: ; %entry +; GFX11-SDAG-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 +; GFX11-SDAG-NEXT: s_mov_b32 s6, -1 +; GFX11-SDAG-NEXT: s_mov_b32 s7, 0x31016000 +; GFX11-SDAG-NEXT: s_mov_b32 s10, s6 +; GFX11-SDAG-NEXT: s_mov_b32 s11, s7 +; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-SDAG-NEXT: s_mov_b32 s8, s2 +; GFX11-SDAG-NEXT: s_mov_b32 s9, s3 +; GFX11-SDAG-NEXT: s_mov_b32 s4, s0 +; GFX11-SDAG-NEXT: buffer_load_b32 v0, off, s[8:11], 0 +; GFX11-SDAG-NEXT: s_mov_b32 s5, s1 +; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) +; GFX11-SDAG-NEXT: v_cvt_f16_f32_e64 v0, -|v0| +; GFX11-SDAG-NEXT: buffer_store_b16 v0, off, s[4:7], 0 +; GFX11-SDAG-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-SDAG-NEXT: s_endpgm +; +; GFX11-GISEL-LABEL: fneg_fabs_fptrunc_f32_to_f16: +; GFX11-GISEL: ; %bb.0: ; %entry +; GFX11-GISEL-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 +; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-GISEL-NEXT: s_load_b32 s2, s[2:3], 0x0 +; GFX11-GISEL-NEXT: s_mov_b32 s3, 0x31016000 +; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-GISEL-NEXT: v_cvt_f16_f32_e64 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-GISEL-NEXT: s_endpgm ptr addrspace(1) %r, ptr addrspace(1) %a) #0 { entry: @@ -618,80 +1276,119 @@ } define amdgpu_kernel void @fptrunc_f32_to_f16_zext_i32( -; SI-LABEL: fptrunc_f32_to_f16_zext_i32: -; SI: ; %bb.0: ; %entry -; SI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 -; SI-NEXT: s_mov_b32 s7, 0xf000 -; SI-NEXT: s_mov_b32 s6, -1 -; SI-NEXT: s_mov_b32 s10, s6 -; SI-NEXT: s_mov_b32 s11, s7 -; SI-NEXT: s_waitcnt lgkmcnt(0) -; SI-NEXT: s_mov_b32 s8, s2 -; SI-NEXT: s_mov_b32 s9, s3 -; SI-NEXT: buffer_load_dword v0, off, s[8:11], 0 -; SI-NEXT: s_mov_b32 s4, s0 -; SI-NEXT: s_mov_b32 s5, s1 -; SI-NEXT: s_waitcnt vmcnt(0) -; SI-NEXT: v_cvt_f16_f32_e32 v0, v0 -; SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 -; SI-NEXT: s_endpgm -; -; VI-LABEL: fptrunc_f32_to_f16_zext_i32: -; VI: ; %bb.0: ; %entry -; VI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 -; VI-NEXT: s_mov_b32 s7, 0xf000 -; VI-NEXT: s_mov_b32 s6, -1 -; VI-NEXT: s_mov_b32 s10, s6 -; VI-NEXT: s_mov_b32 s11, s7 -; VI-NEXT: s_waitcnt lgkmcnt(0) -; VI-NEXT: s_mov_b32 s8, s2 -; VI-NEXT: s_mov_b32 s9, s3 -; VI-NEXT: buffer_load_dword v0, off, s[8:11], 0 -; VI-NEXT: s_mov_b32 s4, s0 -; VI-NEXT: s_mov_b32 s5, s1 -; VI-NEXT: s_waitcnt vmcnt(0) -; VI-NEXT: v_cvt_f16_f32_e32 v0, v0 -; VI-NEXT: buffer_store_dword v0, off, s[4:7], 0 -; VI-NEXT: s_endpgm -; -; GFX9-LABEL: fptrunc_f32_to_f16_zext_i32: -; GFX9: ; %bb.0: ; %entry -; GFX9-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 -; GFX9-NEXT: s_mov_b32 s7, 0xf000 -; GFX9-NEXT: s_mov_b32 s6, -1 -; GFX9-NEXT: s_mov_b32 s10, s6 -; GFX9-NEXT: s_mov_b32 s11, s7 -; GFX9-NEXT: s_waitcnt lgkmcnt(0) -; GFX9-NEXT: s_mov_b32 s8, s2 -; GFX9-NEXT: s_mov_b32 s9, s3 -; GFX9-NEXT: buffer_load_dword v0, off, s[8:11], 0 -; GFX9-NEXT: s_mov_b32 s4, s0 -; GFX9-NEXT: s_mov_b32 s5, s1 -; GFX9-NEXT: s_waitcnt vmcnt(0) -; GFX9-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX9-NEXT: buffer_store_dword v0, off, s[4:7], 0 -; GFX9-NEXT: s_endpgm -; -; GFX11-LABEL: fptrunc_f32_to_f16_zext_i32: -; GFX11: ; %bb.0: ; %entry -; GFX11-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 -; GFX11-NEXT: s_mov_b32 s6, -1 -; GFX11-NEXT: s_mov_b32 s7, 0x31016000 -; GFX11-NEXT: s_mov_b32 s10, s6 -; GFX11-NEXT: s_mov_b32 s11, s7 -; GFX11-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NEXT: s_mov_b32 s8, s2 -; GFX11-NEXT: s_mov_b32 s9, s3 -; GFX11-NEXT: s_mov_b32 s4, s0 -; GFX11-NEXT: buffer_load_b32 v0, off, s[8:11], 0 -; GFX11-NEXT: s_mov_b32 s5, s1 -; GFX11-NEXT: s_waitcnt vmcnt(0) -; GFX11-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-NEXT: buffer_store_b32 v0, off, s[4:7], 0 -; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) -; GFX11-NEXT: s_endpgm +; SI-SDAG-LABEL: fptrunc_f32_to_f16_zext_i32: +; SI-SDAG: ; %bb.0: ; %entry +; SI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 +; SI-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; SI-SDAG-NEXT: s_mov_b32 s6, -1 +; SI-SDAG-NEXT: s_mov_b32 s10, s6 +; SI-SDAG-NEXT: s_mov_b32 s11, s7 +; SI-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; SI-SDAG-NEXT: s_mov_b32 s8, s2 +; SI-SDAG-NEXT: s_mov_b32 s9, s3 +; SI-SDAG-NEXT: buffer_load_dword v0, off, s[8:11], 0 +; SI-SDAG-NEXT: s_mov_b32 s4, s0 +; SI-SDAG-NEXT: s_mov_b32 s5, s1 +; SI-SDAG-NEXT: s_waitcnt vmcnt(0) +; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; SI-SDAG-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; SI-SDAG-NEXT: s_endpgm +; +; VI-SDAG-LABEL: fptrunc_f32_to_f16_zext_i32: +; VI-SDAG: ; %bb.0: ; %entry +; VI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; VI-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; VI-SDAG-NEXT: s_mov_b32 s6, -1 +; VI-SDAG-NEXT: s_mov_b32 s10, s6 +; VI-SDAG-NEXT: s_mov_b32 s11, s7 +; VI-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; VI-SDAG-NEXT: s_mov_b32 s8, s2 +; VI-SDAG-NEXT: s_mov_b32 s9, s3 +; VI-SDAG-NEXT: buffer_load_dword v0, off, s[8:11], 0 +; VI-SDAG-NEXT: s_mov_b32 s4, s0 +; VI-SDAG-NEXT: s_mov_b32 s5, s1 +; VI-SDAG-NEXT: s_waitcnt vmcnt(0) +; VI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; VI-SDAG-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; VI-SDAG-NEXT: s_endpgm +; +; VI-GISEL-LABEL: fptrunc_f32_to_f16_zext_i32: +; VI-GISEL: ; %bb.0: ; %entry +; VI-GISEL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: s_load_dword s2, s[2:3], 0x0 +; VI-GISEL-NEXT: s_mov_b32 s3, 0xf000 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: v_cvt_f16_f32_e32 v0, s2 +; VI-GISEL-NEXT: s_mov_b32 s2, -1 +; VI-GISEL-NEXT: buffer_store_dword v0, off, s[0:3], 0 +; VI-GISEL-NEXT: s_endpgm +; +; GFX9-SDAG-LABEL: fptrunc_f32_to_f16_zext_i32: +; GFX9-SDAG: ; %bb.0: ; %entry +; GFX9-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; GFX9-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; GFX9-SDAG-NEXT: s_mov_b32 s6, -1 +; GFX9-SDAG-NEXT: s_mov_b32 s10, s6 +; GFX9-SDAG-NEXT: s_mov_b32 s11, s7 +; GFX9-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-SDAG-NEXT: s_mov_b32 s8, s2 +; GFX9-SDAG-NEXT: s_mov_b32 s9, s3 +; GFX9-SDAG-NEXT: buffer_load_dword v0, off, s[8:11], 0 +; GFX9-SDAG-NEXT: s_mov_b32 s4, s0 +; GFX9-SDAG-NEXT: s_mov_b32 s5, s1 +; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) +; GFX9-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; GFX9-SDAG-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; GFX9-SDAG-NEXT: s_endpgm +; +; GFX9-GISEL-LABEL: fptrunc_f32_to_f16_zext_i32: +; GFX9-GISEL: ; %bb.0: ; %entry +; GFX9-GISEL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-GISEL-NEXT: s_load_dword s2, s[2:3], 0x0 +; GFX9-GISEL-NEXT: s_mov_b32 s3, 0xf000 +; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-GISEL-NEXT: v_cvt_f16_f32_e32 v0, s2 +; GFX9-GISEL-NEXT: s_mov_b32 s2, -1 +; GFX9-GISEL-NEXT: buffer_store_dword v0, off, s[0:3], 0 +; GFX9-GISEL-NEXT: s_endpgm +; +; GFX11-SDAG-LABEL: fptrunc_f32_to_f16_zext_i32: +; GFX11-SDAG: ; %bb.0: ; %entry +; GFX11-SDAG-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 +; GFX11-SDAG-NEXT: s_mov_b32 s6, -1 +; GFX11-SDAG-NEXT: s_mov_b32 s7, 0x31016000 +; GFX11-SDAG-NEXT: s_mov_b32 s10, s6 +; GFX11-SDAG-NEXT: s_mov_b32 s11, s7 +; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-SDAG-NEXT: s_mov_b32 s8, s2 +; GFX11-SDAG-NEXT: s_mov_b32 s9, s3 +; GFX11-SDAG-NEXT: s_mov_b32 s4, s0 +; GFX11-SDAG-NEXT: buffer_load_b32 v0, off, s[8:11], 0 +; GFX11-SDAG-NEXT: s_mov_b32 s5, s1 +; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) +; GFX11-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0 +; GFX11-SDAG-NEXT: buffer_store_b32 v0, off, s[4:7], 0 +; GFX11-SDAG-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-SDAG-NEXT: s_endpgm +; +; GFX11-GISEL-LABEL: fptrunc_f32_to_f16_zext_i32: +; GFX11-GISEL: ; %bb.0: ; %entry +; GFX11-GISEL-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 +; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-GISEL-NEXT: s_load_b32 s2, s[2:3], 0x0 +; GFX11-GISEL-NEXT: s_mov_b32 s3, 0x31016000 +; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-GISEL-NEXT: v_cvt_f16_f32_e32 v0, s2 +; GFX11-GISEL-NEXT: s_mov_b32 s2, -1 +; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0 +; GFX11-GISEL-NEXT: buffer_store_b32 v0, off, s[0:3], 0 +; GFX11-GISEL-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-GISEL-NEXT: s_endpgm ptr addrspace(1) %r, ptr addrspace(1) %a) #0 { entry: @@ -704,80 +1401,119 @@ } define amdgpu_kernel void @fptrunc_fabs_f32_to_f16_zext_i32( -; SI-LABEL: fptrunc_fabs_f32_to_f16_zext_i32: -; SI: ; %bb.0: ; %entry -; SI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 -; SI-NEXT: s_mov_b32 s7, 0xf000 -; SI-NEXT: s_mov_b32 s6, -1 -; SI-NEXT: s_mov_b32 s10, s6 -; SI-NEXT: s_mov_b32 s11, s7 -; SI-NEXT: s_waitcnt lgkmcnt(0) -; SI-NEXT: s_mov_b32 s8, s2 -; SI-NEXT: s_mov_b32 s9, s3 -; SI-NEXT: buffer_load_dword v0, off, s[8:11], 0 -; SI-NEXT: s_mov_b32 s4, s0 -; SI-NEXT: s_mov_b32 s5, s1 -; SI-NEXT: s_waitcnt vmcnt(0) -; SI-NEXT: v_cvt_f16_f32_e64 v0, |v0| -; SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 -; SI-NEXT: s_endpgm -; -; VI-LABEL: fptrunc_fabs_f32_to_f16_zext_i32: -; VI: ; %bb.0: ; %entry -; VI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 -; VI-NEXT: s_mov_b32 s7, 0xf000 -; VI-NEXT: s_mov_b32 s6, -1 -; VI-NEXT: s_mov_b32 s10, s6 -; VI-NEXT: s_mov_b32 s11, s7 -; VI-NEXT: s_waitcnt lgkmcnt(0) -; VI-NEXT: s_mov_b32 s8, s2 -; VI-NEXT: s_mov_b32 s9, s3 -; VI-NEXT: buffer_load_dword v0, off, s[8:11], 0 -; VI-NEXT: s_mov_b32 s4, s0 -; VI-NEXT: s_mov_b32 s5, s1 -; VI-NEXT: s_waitcnt vmcnt(0) -; VI-NEXT: v_cvt_f16_f32_e64 v0, |v0| -; VI-NEXT: buffer_store_dword v0, off, s[4:7], 0 -; VI-NEXT: s_endpgm -; -; GFX9-LABEL: fptrunc_fabs_f32_to_f16_zext_i32: -; GFX9: ; %bb.0: ; %entry -; GFX9-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 -; GFX9-NEXT: s_mov_b32 s7, 0xf000 -; GFX9-NEXT: s_mov_b32 s6, -1 -; GFX9-NEXT: s_mov_b32 s10, s6 -; GFX9-NEXT: s_mov_b32 s11, s7 -; GFX9-NEXT: s_waitcnt lgkmcnt(0) -; GFX9-NEXT: s_mov_b32 s8, s2 -; GFX9-NEXT: s_mov_b32 s9, s3 -; GFX9-NEXT: buffer_load_dword v0, off, s[8:11], 0 -; GFX9-NEXT: s_mov_b32 s4, s0 -; GFX9-NEXT: s_mov_b32 s5, s1 -; GFX9-NEXT: s_waitcnt vmcnt(0) -; GFX9-NEXT: v_cvt_f16_f32_e64 v0, |v0| -; GFX9-NEXT: buffer_store_dword v0, off, s[4:7], 0 -; GFX9-NEXT: s_endpgm -; -; GFX11-LABEL: fptrunc_fabs_f32_to_f16_zext_i32: -; GFX11: ; %bb.0: ; %entry -; GFX11-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 -; GFX11-NEXT: s_mov_b32 s6, -1 -; GFX11-NEXT: s_mov_b32 s7, 0x31016000 -; GFX11-NEXT: s_mov_b32 s10, s6 -; GFX11-NEXT: s_mov_b32 s11, s7 -; GFX11-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NEXT: s_mov_b32 s8, s2 -; GFX11-NEXT: s_mov_b32 s9, s3 -; GFX11-NEXT: s_mov_b32 s4, s0 -; GFX11-NEXT: buffer_load_b32 v0, off, s[8:11], 0 -; GFX11-NEXT: s_mov_b32 s5, s1 -; GFX11-NEXT: s_waitcnt vmcnt(0) -; GFX11-NEXT: v_cvt_f16_f32_e64 v0, |v0| -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-NEXT: buffer_store_b32 v0, off, s[4:7], 0 -; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) -; GFX11-NEXT: s_endpgm +; SI-SDAG-LABEL: fptrunc_fabs_f32_to_f16_zext_i32: +; SI-SDAG: ; %bb.0: ; %entry +; SI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 +; SI-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; SI-SDAG-NEXT: s_mov_b32 s6, -1 +; SI-SDAG-NEXT: s_mov_b32 s10, s6 +; SI-SDAG-NEXT: s_mov_b32 s11, s7 +; SI-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; SI-SDAG-NEXT: s_mov_b32 s8, s2 +; SI-SDAG-NEXT: s_mov_b32 s9, s3 +; SI-SDAG-NEXT: buffer_load_dword v0, off, s[8:11], 0 +; SI-SDAG-NEXT: s_mov_b32 s4, s0 +; SI-SDAG-NEXT: s_mov_b32 s5, s1 +; SI-SDAG-NEXT: s_waitcnt vmcnt(0) +; SI-SDAG-NEXT: v_cvt_f16_f32_e64 v0, |v0| +; SI-SDAG-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; SI-SDAG-NEXT: s_endpgm +; +; VI-SDAG-LABEL: fptrunc_fabs_f32_to_f16_zext_i32: +; VI-SDAG: ; %bb.0: ; %entry +; VI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; VI-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; VI-SDAG-NEXT: s_mov_b32 s6, -1 +; VI-SDAG-NEXT: s_mov_b32 s10, s6 +; VI-SDAG-NEXT: s_mov_b32 s11, s7 +; VI-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; VI-SDAG-NEXT: s_mov_b32 s8, s2 +; VI-SDAG-NEXT: s_mov_b32 s9, s3 +; VI-SDAG-NEXT: buffer_load_dword v0, off, s[8:11], 0 +; VI-SDAG-NEXT: s_mov_b32 s4, s0 +; VI-SDAG-NEXT: s_mov_b32 s5, s1 +; VI-SDAG-NEXT: s_waitcnt vmcnt(0) +; VI-SDAG-NEXT: v_cvt_f16_f32_e64 v0, |v0| +; VI-SDAG-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; VI-SDAG-NEXT: s_endpgm +; +; VI-GISEL-LABEL: fptrunc_fabs_f32_to_f16_zext_i32: +; VI-GISEL: ; %bb.0: ; %entry +; VI-GISEL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: s_load_dword s2, s[2:3], 0x0 +; VI-GISEL-NEXT: s_mov_b32 s3, 0xf000 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: v_cvt_f16_f32_e64 v0, |s2| +; VI-GISEL-NEXT: s_mov_b32 s2, -1 +; VI-GISEL-NEXT: buffer_store_dword v0, off, s[0:3], 0 +; VI-GISEL-NEXT: s_endpgm +; +; GFX9-SDAG-LABEL: fptrunc_fabs_f32_to_f16_zext_i32: +; GFX9-SDAG: ; %bb.0: ; %entry +; GFX9-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; GFX9-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; GFX9-SDAG-NEXT: s_mov_b32 s6, -1 +; GFX9-SDAG-NEXT: s_mov_b32 s10, s6 +; GFX9-SDAG-NEXT: s_mov_b32 s11, s7 +; GFX9-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-SDAG-NEXT: s_mov_b32 s8, s2 +; GFX9-SDAG-NEXT: s_mov_b32 s9, s3 +; GFX9-SDAG-NEXT: buffer_load_dword v0, off, s[8:11], 0 +; GFX9-SDAG-NEXT: s_mov_b32 s4, s0 +; GFX9-SDAG-NEXT: s_mov_b32 s5, s1 +; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) +; GFX9-SDAG-NEXT: v_cvt_f16_f32_e64 v0, |v0| +; GFX9-SDAG-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; GFX9-SDAG-NEXT: s_endpgm +; +; GFX9-GISEL-LABEL: fptrunc_fabs_f32_to_f16_zext_i32: +; GFX9-GISEL: ; %bb.0: ; %entry +; GFX9-GISEL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-GISEL-NEXT: s_load_dword s2, s[2:3], 0x0 +; GFX9-GISEL-NEXT: s_mov_b32 s3, 0xf000 +; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-GISEL-NEXT: v_cvt_f16_f32_e64 v0, |s2| +; GFX9-GISEL-NEXT: s_mov_b32 s2, -1 +; GFX9-GISEL-NEXT: buffer_store_dword v0, off, s[0:3], 0 +; GFX9-GISEL-NEXT: s_endpgm +; +; GFX11-SDAG-LABEL: fptrunc_fabs_f32_to_f16_zext_i32: +; GFX11-SDAG: ; %bb.0: ; %entry +; GFX11-SDAG-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 +; GFX11-SDAG-NEXT: s_mov_b32 s6, -1 +; GFX11-SDAG-NEXT: s_mov_b32 s7, 0x31016000 +; GFX11-SDAG-NEXT: s_mov_b32 s10, s6 +; GFX11-SDAG-NEXT: s_mov_b32 s11, s7 +; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-SDAG-NEXT: s_mov_b32 s8, s2 +; GFX11-SDAG-NEXT: s_mov_b32 s9, s3 +; GFX11-SDAG-NEXT: s_mov_b32 s4, s0 +; GFX11-SDAG-NEXT: buffer_load_b32 v0, off, s[8:11], 0 +; GFX11-SDAG-NEXT: s_mov_b32 s5, s1 +; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) +; GFX11-SDAG-NEXT: v_cvt_f16_f32_e64 v0, |v0| +; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0 +; GFX11-SDAG-NEXT: buffer_store_b32 v0, off, s[4:7], 0 +; GFX11-SDAG-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-SDAG-NEXT: s_endpgm +; +; GFX11-GISEL-LABEL: fptrunc_fabs_f32_to_f16_zext_i32: +; GFX11-GISEL: ; %bb.0: ; %entry +; GFX11-GISEL-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 +; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-GISEL-NEXT: s_load_b32 s2, s[2:3], 0x0 +; GFX11-GISEL-NEXT: s_mov_b32 s3, 0x31016000 +; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-GISEL-NEXT: v_cvt_f16_f32_e64 v0, |s2| +; GFX11-GISEL-NEXT: s_mov_b32 s2, -1 +; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0 +; GFX11-GISEL-NEXT: buffer_store_b32 v0, off, s[0:3], 0 +; GFX11-GISEL-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-GISEL-NEXT: s_endpgm ptr addrspace(1) %r, ptr addrspace(1) %a) #0 { entry: @@ -791,83 +1527,124 @@ } define amdgpu_kernel void @fptrunc_f32_to_f16_sext_i32( -; SI-LABEL: fptrunc_f32_to_f16_sext_i32: -; SI: ; %bb.0: ; %entry -; SI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 -; SI-NEXT: s_mov_b32 s7, 0xf000 -; SI-NEXT: s_mov_b32 s6, -1 -; SI-NEXT: s_mov_b32 s10, s6 -; SI-NEXT: s_mov_b32 s11, s7 -; SI-NEXT: s_waitcnt lgkmcnt(0) -; SI-NEXT: s_mov_b32 s8, s2 -; SI-NEXT: s_mov_b32 s9, s3 -; SI-NEXT: buffer_load_dword v0, off, s[8:11], 0 -; SI-NEXT: s_mov_b32 s4, s0 -; SI-NEXT: s_mov_b32 s5, s1 -; SI-NEXT: s_waitcnt vmcnt(0) -; SI-NEXT: v_cvt_f16_f32_e32 v0, v0 -; SI-NEXT: v_bfe_i32 v0, v0, 0, 16 -; SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 -; SI-NEXT: s_endpgm -; -; VI-LABEL: fptrunc_f32_to_f16_sext_i32: -; VI: ; %bb.0: ; %entry -; VI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 -; VI-NEXT: s_mov_b32 s7, 0xf000 -; VI-NEXT: s_mov_b32 s6, -1 -; VI-NEXT: s_mov_b32 s10, s6 -; VI-NEXT: s_mov_b32 s11, s7 -; VI-NEXT: s_waitcnt lgkmcnt(0) -; VI-NEXT: s_mov_b32 s8, s2 -; VI-NEXT: s_mov_b32 s9, s3 -; VI-NEXT: buffer_load_dword v0, off, s[8:11], 0 -; VI-NEXT: s_mov_b32 s4, s0 -; VI-NEXT: s_mov_b32 s5, s1 -; VI-NEXT: s_waitcnt vmcnt(0) -; VI-NEXT: v_cvt_f16_f32_e32 v0, v0 -; VI-NEXT: v_bfe_i32 v0, v0, 0, 16 -; VI-NEXT: buffer_store_dword v0, off, s[4:7], 0 -; VI-NEXT: s_endpgm -; -; GFX9-LABEL: fptrunc_f32_to_f16_sext_i32: -; GFX9: ; %bb.0: ; %entry -; GFX9-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 -; GFX9-NEXT: s_mov_b32 s7, 0xf000 -; GFX9-NEXT: s_mov_b32 s6, -1 -; GFX9-NEXT: s_mov_b32 s10, s6 -; GFX9-NEXT: s_mov_b32 s11, s7 -; GFX9-NEXT: s_waitcnt lgkmcnt(0) -; GFX9-NEXT: s_mov_b32 s8, s2 -; GFX9-NEXT: s_mov_b32 s9, s3 -; GFX9-NEXT: buffer_load_dword v0, off, s[8:11], 0 -; GFX9-NEXT: s_mov_b32 s4, s0 -; GFX9-NEXT: s_mov_b32 s5, s1 -; GFX9-NEXT: s_waitcnt vmcnt(0) -; GFX9-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX9-NEXT: v_bfe_i32 v0, v0, 0, 16 -; GFX9-NEXT: buffer_store_dword v0, off, s[4:7], 0 -; GFX9-NEXT: s_endpgm -; -; GFX11-LABEL: fptrunc_f32_to_f16_sext_i32: -; GFX11: ; %bb.0: ; %entry -; GFX11-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 -; GFX11-NEXT: s_mov_b32 s6, -1 -; GFX11-NEXT: s_mov_b32 s7, 0x31016000 -; GFX11-NEXT: s_mov_b32 s10, s6 -; GFX11-NEXT: s_mov_b32 s11, s7 -; GFX11-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NEXT: s_mov_b32 s8, s2 -; GFX11-NEXT: s_mov_b32 s9, s3 -; GFX11-NEXT: s_mov_b32 s4, s0 -; GFX11-NEXT: buffer_load_b32 v0, off, s[8:11], 0 -; GFX11-NEXT: s_mov_b32 s5, s1 -; GFX11-NEXT: s_waitcnt vmcnt(0) -; GFX11-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NEXT: v_bfe_i32 v0, v0, 0, 16 -; GFX11-NEXT: buffer_store_b32 v0, off, s[4:7], 0 -; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) -; GFX11-NEXT: s_endpgm +; SI-SDAG-LABEL: fptrunc_f32_to_f16_sext_i32: +; SI-SDAG: ; %bb.0: ; %entry +; SI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9 +; SI-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; SI-SDAG-NEXT: s_mov_b32 s6, -1 +; SI-SDAG-NEXT: s_mov_b32 s10, s6 +; SI-SDAG-NEXT: s_mov_b32 s11, s7 +; SI-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; SI-SDAG-NEXT: s_mov_b32 s8, s2 +; SI-SDAG-NEXT: s_mov_b32 s9, s3 +; SI-SDAG-NEXT: buffer_load_dword v0, off, s[8:11], 0 +; SI-SDAG-NEXT: s_mov_b32 s4, s0 +; SI-SDAG-NEXT: s_mov_b32 s5, s1 +; SI-SDAG-NEXT: s_waitcnt vmcnt(0) +; SI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; SI-SDAG-NEXT: v_bfe_i32 v0, v0, 0, 16 +; SI-SDAG-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; SI-SDAG-NEXT: s_endpgm +; +; VI-SDAG-LABEL: fptrunc_f32_to_f16_sext_i32: +; VI-SDAG: ; %bb.0: ; %entry +; VI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; VI-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; VI-SDAG-NEXT: s_mov_b32 s6, -1 +; VI-SDAG-NEXT: s_mov_b32 s10, s6 +; VI-SDAG-NEXT: s_mov_b32 s11, s7 +; VI-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; VI-SDAG-NEXT: s_mov_b32 s8, s2 +; VI-SDAG-NEXT: s_mov_b32 s9, s3 +; VI-SDAG-NEXT: buffer_load_dword v0, off, s[8:11], 0 +; VI-SDAG-NEXT: s_mov_b32 s4, s0 +; VI-SDAG-NEXT: s_mov_b32 s5, s1 +; VI-SDAG-NEXT: s_waitcnt vmcnt(0) +; VI-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; VI-SDAG-NEXT: v_bfe_i32 v0, v0, 0, 16 +; VI-SDAG-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; VI-SDAG-NEXT: s_endpgm +; +; VI-GISEL-LABEL: fptrunc_f32_to_f16_sext_i32: +; VI-GISEL: ; %bb.0: ; %entry +; VI-GISEL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: s_load_dword s2, s[2:3], 0x0 +; VI-GISEL-NEXT: s_mov_b32 s3, 0xf000 +; VI-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; VI-GISEL-NEXT: v_cvt_f16_f32_e32 v0, s2 +; VI-GISEL-NEXT: s_mov_b32 s2, -1 +; VI-GISEL-NEXT: v_bfe_i32 v0, v0, 0, 16 +; VI-GISEL-NEXT: buffer_store_dword v0, off, s[0:3], 0 +; VI-GISEL-NEXT: s_endpgm +; +; GFX9-SDAG-LABEL: fptrunc_f32_to_f16_sext_i32: +; GFX9-SDAG: ; %bb.0: ; %entry +; GFX9-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; GFX9-SDAG-NEXT: s_mov_b32 s7, 0xf000 +; GFX9-SDAG-NEXT: s_mov_b32 s6, -1 +; GFX9-SDAG-NEXT: s_mov_b32 s10, s6 +; GFX9-SDAG-NEXT: s_mov_b32 s11, s7 +; GFX9-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-SDAG-NEXT: s_mov_b32 s8, s2 +; GFX9-SDAG-NEXT: s_mov_b32 s9, s3 +; GFX9-SDAG-NEXT: buffer_load_dword v0, off, s[8:11], 0 +; GFX9-SDAG-NEXT: s_mov_b32 s4, s0 +; GFX9-SDAG-NEXT: s_mov_b32 s5, s1 +; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) +; GFX9-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; GFX9-SDAG-NEXT: v_bfe_i32 v0, v0, 0, 16 +; GFX9-SDAG-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; GFX9-SDAG-NEXT: s_endpgm +; +; GFX9-GISEL-LABEL: fptrunc_f32_to_f16_sext_i32: +; GFX9-GISEL: ; %bb.0: ; %entry +; GFX9-GISEL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 +; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-GISEL-NEXT: s_load_dword s2, s[2:3], 0x0 +; GFX9-GISEL-NEXT: s_mov_b32 s3, 0xf000 +; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-GISEL-NEXT: v_cvt_f16_f32_e32 v0, s2 +; GFX9-GISEL-NEXT: s_mov_b32 s2, -1 +; GFX9-GISEL-NEXT: v_bfe_i32 v0, v0, 0, 16 +; GFX9-GISEL-NEXT: buffer_store_dword v0, off, s[0:3], 0 +; GFX9-GISEL-NEXT: s_endpgm +; +; GFX11-SDAG-LABEL: fptrunc_f32_to_f16_sext_i32: +; GFX11-SDAG: ; %bb.0: ; %entry +; GFX11-SDAG-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 +; GFX11-SDAG-NEXT: s_mov_b32 s6, -1 +; GFX11-SDAG-NEXT: s_mov_b32 s7, 0x31016000 +; GFX11-SDAG-NEXT: s_mov_b32 s10, s6 +; GFX11-SDAG-NEXT: s_mov_b32 s11, s7 +; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-SDAG-NEXT: s_mov_b32 s8, s2 +; GFX11-SDAG-NEXT: s_mov_b32 s9, s3 +; GFX11-SDAG-NEXT: s_mov_b32 s4, s0 +; GFX11-SDAG-NEXT: buffer_load_b32 v0, off, s[8:11], 0 +; GFX11-SDAG-NEXT: s_mov_b32 s5, s1 +; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) +; GFX11-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-SDAG-NEXT: v_bfe_i32 v0, v0, 0, 16 +; GFX11-SDAG-NEXT: buffer_store_b32 v0, off, s[4:7], 0 +; GFX11-SDAG-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-SDAG-NEXT: s_endpgm +; +; GFX11-GISEL-LABEL: fptrunc_f32_to_f16_sext_i32: +; GFX11-GISEL: ; %bb.0: ; %entry +; GFX11-GISEL-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 +; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-GISEL-NEXT: s_load_b32 s2, s[2:3], 0x0 +; GFX11-GISEL-NEXT: s_mov_b32 s3, 0x31016000 +; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-GISEL-NEXT: v_cvt_f16_f32_e32 v0, s2 +; GFX11-GISEL-NEXT: s_mov_b32 s2, -1 +; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-GISEL-NEXT: v_bfe_i32 v0, v0, 0, 16 +; GFX11-GISEL-NEXT: buffer_store_b32 v0, off, s[0:3], 0 +; GFX11-GISEL-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-GISEL-NEXT: s_endpgm ptr addrspace(1) %r, ptr addrspace(1) %a) #0 { entry: diff --git a/llvm/test/CodeGen/AMDGPU/fptrunc.ll b/llvm/test/CodeGen/AMDGPU/fptrunc.ll --- a/llvm/test/CodeGen/AMDGPU/fptrunc.ll +++ b/llvm/test/CodeGen/AMDGPU/fptrunc.ll @@ -1,9 +1,17 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2 ; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefixes=SI %s -; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefixes=VI %s -; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefixes=VI-UNSAFE %s -; RUN: llc -march=amdgcn -mcpu=gfx1100 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX11,GFX11-SAFE %s -; RUN: llc -march=amdgcn -mcpu=gfx1100 -mattr=-flat-for-global -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX11,GFX11-UNSAFE %s +; RUN: llc -march=amdgcn -mcpu=tonga -global-isel=0 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefixes=VI-SDAG,VI-SAFE-SDAG %s +; RUN: llc -march=amdgcn -mcpu=tonga -global-isel=1 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefixes=VI-GISEL %s +; RUN: llc -march=amdgcn -mcpu=tonga -global-isel=0 -mattr=-flat-for-global -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefixes=VI-SDAG,VI-UNSAFE-SDAG %s +; RUN: llc -march=amdgcn -mcpu=tonga -global-isel=1 -mattr=-flat-for-global -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefixes=VI-GISEL %s +; RUN: llc -march=amdgcn -mcpu=gfx1030 -global-isel=0 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10-SDAG,GFX10-SAFE-SDAG %s +; RUN: llc -march=amdgcn -mcpu=gfx1030 -global-isel=1 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10-GISEL %s +; RUN: llc -march=amdgcn -mcpu=gfx1030 -global-isel=0 -mattr=-flat-for-global -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10-SDAG,GFX10-UNSAFE-SDAG %s +; RUN: llc -march=amdgcn -mcpu=gfx1030 -global-isel=1 -mattr=-flat-for-global -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10-GISEL %s +; RUN: llc -march=amdgcn -mcpu=gfx1100 -global-isel=0 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX11-SDAG,GFX11-SAFE-SDAG %s +; RUN: llc -march=amdgcn -mcpu=gfx1100 -global-isel=1 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX11-GISEL %s +; RUN: llc -march=amdgcn -mcpu=gfx1100 -global-isel=0 -mattr=-flat-for-global -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX11-SDAG,GFX11-UNSAFE-SDAG %s +; RUN: llc -march=amdgcn -mcpu=gfx1100 -global-isel=1 -mattr=-flat-for-global -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX11-GISEL %s define amdgpu_kernel void @fptrunc_f64_to_f32(ptr addrspace(1) %out, double %in) { ; SI-LABEL: fptrunc_f64_to_f32: @@ -18,40 +26,69 @@ ; SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 ; SI-NEXT: s_endpgm ; -; VI-LABEL: fptrunc_f64_to_f32: -; VI: ; %bb.0: -; VI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 -; VI-NEXT: s_mov_b32 s7, 0xf000 -; VI-NEXT: s_mov_b32 s6, -1 -; VI-NEXT: s_waitcnt lgkmcnt(0) -; VI-NEXT: v_cvt_f32_f64_e32 v0, s[2:3] -; VI-NEXT: s_mov_b32 s4, s0 -; VI-NEXT: s_mov_b32 s5, s1 -; VI-NEXT: buffer_store_dword v0, off, s[4:7], 0 -; VI-NEXT: s_endpgm -; -; VI-UNSAFE-LABEL: fptrunc_f64_to_f32: -; VI-UNSAFE: ; %bb.0: -; VI-UNSAFE-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 -; VI-UNSAFE-NEXT: s_mov_b32 s7, 0xf000 -; VI-UNSAFE-NEXT: s_mov_b32 s6, -1 -; VI-UNSAFE-NEXT: s_waitcnt lgkmcnt(0) -; VI-UNSAFE-NEXT: v_cvt_f32_f64_e32 v0, s[2:3] -; VI-UNSAFE-NEXT: s_mov_b32 s4, s0 -; VI-UNSAFE-NEXT: s_mov_b32 s5, s1 -; VI-UNSAFE-NEXT: buffer_store_dword v0, off, s[4:7], 0 -; VI-UNSAFE-NEXT: s_endpgm -; -; GFX11-LABEL: fptrunc_f64_to_f32: -; GFX11: ; %bb.0: -; GFX11-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 -; GFX11-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NEXT: v_cvt_f32_f64_e32 v0, s[2:3] -; GFX11-NEXT: s_mov_b32 s3, 0x31016000 -; GFX11-NEXT: s_mov_b32 s2, -1 -; GFX11-NEXT: buffer_store_b32 v0, off, s[0:3], 0 -; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) -; GFX11-NEXT: s_endpgm +; VI-SDAG-LABEL: fptrunc_f64_to_f32: +; VI-SDAG: ; %bb.0: +; VI-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 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: +; VI-GISEL: ; %bb.0: +; VI-GISEL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 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: +; GFX10-SDAG: ; %bb.0: +; GFX10-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 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: +; GFX10-GISEL: ; %bb.0: +; GFX10-GISEL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 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: +; GFX11-SDAG: ; %bb.0: +; GFX11-SDAG-NEXT: s_load_b128 s[0:3], s[0:1], 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-SDAG-NEXT: s_endpgm +; +; GFX11-GISEL-LABEL: fptrunc_f64_to_f32: +; GFX11-GISEL: ; %bb.0: +; GFX11-GISEL-NEXT: s_load_b128 s[0:3], s[0:1], 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-GISEL-NEXT: s_endpgm %result = fptrunc double %in to float store float %result, ptr addrspace(1) %out ret void @@ -118,154 +155,381 @@ ; SI-NEXT: buffer_store_short v0, off, s[0:3], 0 ; SI-NEXT: s_endpgm ; -; VI-LABEL: fptrunc_f64_to_f16: -; VI: ; %bb.0: -; VI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24 -; VI-NEXT: s_mov_b32 s3, 0xf000 -; VI-NEXT: s_mov_b32 s2, -1 -; VI-NEXT: s_waitcnt lgkmcnt(0) -; VI-NEXT: s_mov_b32 s0, s4 -; VI-NEXT: s_lshr_b32 s4, s7, 8 -; VI-NEXT: s_and_b32 s8, s4, 0xffe -; VI-NEXT: s_and_b32 s4, s7, 0x1ff -; VI-NEXT: s_or_b32 s4, s4, s6 -; VI-NEXT: s_cmp_lg_u32 s4, 0 -; VI-NEXT: s_mov_b32 s1, s5 -; VI-NEXT: s_cselect_b64 s[4:5], -1, 0 -; VI-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[4:5] -; VI-NEXT: v_readfirstlane_b32 s4, v0 -; VI-NEXT: s_bfe_u32 s5, s7, 0xb0014 -; VI-NEXT: s_or_b32 s6, s8, s4 -; VI-NEXT: s_sub_i32 s8, 0x3f1, s5 -; VI-NEXT: v_med3_i32 v0, s8, 0, 13 -; VI-NEXT: s_or_b32 s4, s6, 0x1000 -; VI-NEXT: v_readfirstlane_b32 s8, v0 -; VI-NEXT: s_lshr_b32 s8, s4, s8 -; VI-NEXT: v_lshlrev_b32_e64 v0, v0, s8 -; VI-NEXT: v_cmp_ne_u32_e32 vcc, s4, v0 -; VI-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc -; VI-NEXT: s_add_i32 s10, s5, 0xfffffc10 -; VI-NEXT: v_readfirstlane_b32 s4, v0 -; VI-NEXT: s_lshl_b32 s5, s10, 12 -; VI-NEXT: s_or_b32 s4, s8, s4 -; VI-NEXT: s_or_b32 s5, s6, s5 -; VI-NEXT: s_cmp_lt_i32 s10, 1 -; VI-NEXT: s_cselect_b32 s11, s4, s5 -; VI-NEXT: s_and_b32 s8, s11, 7 -; VI-NEXT: s_cmp_gt_i32 s8, 5 -; VI-NEXT: s_cselect_b64 s[4:5], -1, 0 -; VI-NEXT: s_cmp_eq_u32 s8, 3 -; VI-NEXT: s_cselect_b64 s[8:9], -1, 0 -; VI-NEXT: s_or_b64 s[4:5], s[8:9], s[4:5] -; VI-NEXT: s_lshr_b32 s8, s11, 2 -; VI-NEXT: s_cmp_lg_u64 s[4:5], 0 -; VI-NEXT: s_addc_u32 s4, s8, 0 -; VI-NEXT: s_cmp_lt_i32 s10, 31 -; VI-NEXT: s_cselect_b32 s8, s4, 0x7c00 -; VI-NEXT: s_cmp_lg_u32 s6, 0 -; VI-NEXT: s_cselect_b64 s[4:5], -1, 0 -; VI-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[4:5] -; VI-NEXT: v_lshlrev_b32_e32 v0, 9, v0 -; VI-NEXT: s_cmpk_eq_i32 s10, 0x40f -; VI-NEXT: v_or_b32_e32 v0, 0x7c00, v0 -; VI-NEXT: v_mov_b32_e32 v1, s8 -; VI-NEXT: s_cselect_b64 vcc, -1, 0 -; VI-NEXT: s_lshr_b32 s4, s7, 16 -; VI-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc -; VI-NEXT: s_and_b32 s4, s4, 0x8000 -; VI-NEXT: v_or_b32_e32 v0, s4, v0 -; VI-NEXT: buffer_store_short v0, off, s[0:3], 0 -; VI-NEXT: s_endpgm -; -; VI-UNSAFE-LABEL: fptrunc_f64_to_f16: -; VI-UNSAFE: ; %bb.0: -; VI-UNSAFE-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x24 -; VI-UNSAFE-NEXT: s_waitcnt lgkmcnt(0) -; VI-UNSAFE-NEXT: v_cvt_f32_f64_e32 v0, s[2:3] -; VI-UNSAFE-NEXT: s_mov_b32 s3, 0xf000 -; VI-UNSAFE-NEXT: s_mov_b32 s2, -1 -; VI-UNSAFE-NEXT: v_cvt_f16_f32_e32 v0, v0 -; VI-UNSAFE-NEXT: buffer_store_short v0, off, s[0:3], 0 -; VI-UNSAFE-NEXT: s_endpgm -; -; GFX11-SAFE-LABEL: fptrunc_f64_to_f16: -; GFX11-SAFE: ; %bb.0: -; GFX11-SAFE-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 -; GFX11-SAFE-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-SAFE-NEXT: s_and_b32 s4, s3, 0x1ff -; GFX11-SAFE-NEXT: s_lshr_b32 s5, s3, 8 -; GFX11-SAFE-NEXT: s_or_b32 s2, s4, s2 -; GFX11-SAFE-NEXT: s_and_b32 s4, s5, 0xffe -; GFX11-SAFE-NEXT: s_cmp_lg_u32 s2, 0 -; GFX11-SAFE-NEXT: s_cselect_b32 s2, -1, 0 -; GFX11-SAFE-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) -; GFX11-SAFE-NEXT: v_cndmask_b32_e64 v0, 0, 1, s2 -; GFX11-SAFE-NEXT: s_bfe_u32 s2, s3, 0xb0014 -; GFX11-SAFE-NEXT: s_sub_i32 s5, 0x3f1, s2 -; GFX11-SAFE-NEXT: s_addk_i32 s2, 0xfc10 -; GFX11-SAFE-NEXT: v_med3_i32 v1, s5, 0, 13 -; GFX11-SAFE-NEXT: v_readfirstlane_b32 s5, v0 -; GFX11-SAFE-NEXT: s_lshl_b32 s7, s2, 12 -; GFX11-SAFE-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-SAFE-NEXT: v_readfirstlane_b32 s6, v1 -; GFX11-SAFE-NEXT: s_or_b32 s4, s4, s5 -; GFX11-SAFE-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-SAFE-NEXT: s_or_b32 s5, s4, 0x1000 -; GFX11-SAFE-NEXT: s_or_b32 s7, s4, s7 -; GFX11-SAFE-NEXT: s_lshr_b32 s6, s5, s6 -; GFX11-SAFE-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SAFE-NEXT: v_lshlrev_b32_e64 v0, v1, s6 -; GFX11-SAFE-NEXT: v_cmp_ne_u32_e32 vcc_lo, s5, v0 -; GFX11-SAFE-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo -; GFX11-SAFE-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SAFE-NEXT: v_readfirstlane_b32 s5, v0 -; GFX11-SAFE-NEXT: s_or_b32 s5, s6, s5 -; GFX11-SAFE-NEXT: s_cmp_lt_i32 s2, 1 -; GFX11-SAFE-NEXT: s_cselect_b32 s5, s5, s7 -; GFX11-SAFE-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) -; GFX11-SAFE-NEXT: s_and_b32 s6, s5, 7 -; GFX11-SAFE-NEXT: s_cmp_gt_i32 s6, 5 -; GFX11-SAFE-NEXT: s_cselect_b32 s7, -1, 0 -; GFX11-SAFE-NEXT: s_cmp_eq_u32 s6, 3 -; GFX11-SAFE-NEXT: s_cselect_b32 s6, -1, 0 -; GFX11-SAFE-NEXT: s_lshr_b32 s5, s5, 2 -; GFX11-SAFE-NEXT: s_or_b32 s6, s6, s7 -; GFX11-SAFE-NEXT: s_delay_alu instid0(SALU_CYCLE_1) -; GFX11-SAFE-NEXT: s_cmp_lg_u32 s6, 0 -; GFX11-SAFE-NEXT: s_addc_u32 s5, s5, 0 -; GFX11-SAFE-NEXT: s_cmp_lt_i32 s2, 31 -; GFX11-SAFE-NEXT: s_cselect_b32 s5, s5, 0x7c00 -; GFX11-SAFE-NEXT: s_cmp_lg_u32 s4, 0 -; GFX11-SAFE-NEXT: s_cselect_b32 s4, -1, 0 -; GFX11-SAFE-NEXT: s_cmpk_eq_i32 s2, 0x40f -; GFX11-SAFE-NEXT: v_cndmask_b32_e64 v0, 0, 1, s4 -; GFX11-SAFE-NEXT: s_cselect_b32 vcc_lo, -1, 0 -; GFX11-SAFE-NEXT: s_lshr_b32 s2, s3, 16 -; GFX11-SAFE-NEXT: s_mov_b32 s3, 0x31016000 -; GFX11-SAFE-NEXT: s_and_b32 s2, s2, 0x8000 -; GFX11-SAFE-NEXT: v_lshlrev_b32_e32 v0, 9, v0 -; GFX11-SAFE-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SAFE-NEXT: v_or_b32_e32 v0, 0x7c00, v0 -; GFX11-SAFE-NEXT: v_cndmask_b32_e32 v0, s5, v0, vcc_lo -; GFX11-SAFE-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-SAFE-NEXT: v_or_b32_e32 v0, s2, v0 -; GFX11-SAFE-NEXT: s_mov_b32 s2, -1 -; GFX11-SAFE-NEXT: buffer_store_b16 v0, off, s[0:3], 0 -; GFX11-SAFE-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) -; GFX11-SAFE-NEXT: s_endpgm -; -; GFX11-UNSAFE-LABEL: fptrunc_f64_to_f16: -; GFX11-UNSAFE: ; %bb.0: -; GFX11-UNSAFE-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 -; GFX11-UNSAFE-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-UNSAFE-NEXT: v_cvt_f32_f64_e32 v0, s[2:3] -; GFX11-UNSAFE-NEXT: s_mov_b32 s3, 0x31016000 -; GFX11-UNSAFE-NEXT: s_mov_b32 s2, -1 -; GFX11-UNSAFE-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-UNSAFE-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX11-UNSAFE-NEXT: buffer_store_b16 v0, off, s[0:3], 0 -; GFX11-UNSAFE-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) -; GFX11-UNSAFE-NEXT: s_endpgm +; VI-SAFE-SDAG-LABEL: fptrunc_f64_to_f16: +; VI-SAFE-SDAG: ; %bb.0: +; VI-SAFE-SDAG-NEXT: s_load_dwordx4 s[4:7], s[0:1], 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 s5, s7, 0xb0014 +; VI-SAFE-SDAG-NEXT: s_or_b32 s6, s8, s4 +; VI-SAFE-SDAG-NEXT: s_sub_i32 s8, 0x3f1, s5 +; VI-SAFE-SDAG-NEXT: v_med3_i32 v0, s8, 0, 13 +; VI-SAFE-SDAG-NEXT: s_or_b32 s4, s6, 0x1000 +; VI-SAFE-SDAG-NEXT: v_readfirstlane_b32 s8, v0 +; VI-SAFE-SDAG-NEXT: s_lshr_b32 s8, s4, s8 +; VI-SAFE-SDAG-NEXT: v_lshlrev_b32_e64 v0, v0, s8 +; VI-SAFE-SDAG-NEXT: v_cmp_ne_u32_e32 vcc, s4, v0 +; VI-SAFE-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc +; VI-SAFE-SDAG-NEXT: s_add_i32 s10, s5, 0xfffffc10 +; VI-SAFE-SDAG-NEXT: v_readfirstlane_b32 s4, v0 +; VI-SAFE-SDAG-NEXT: s_lshl_b32 s5, s10, 12 +; VI-SAFE-SDAG-NEXT: s_or_b32 s4, s8, s4 +; VI-SAFE-SDAG-NEXT: s_or_b32 s5, s6, s5 +; VI-SAFE-SDAG-NEXT: s_cmp_lt_i32 s10, 1 +; VI-SAFE-SDAG-NEXT: s_cselect_b32 s11, s4, s5 +; VI-SAFE-SDAG-NEXT: s_and_b32 s8, s11, 7 +; VI-SAFE-SDAG-NEXT: s_cmp_gt_i32 s8, 5 +; VI-SAFE-SDAG-NEXT: s_cselect_b64 s[4:5], -1, 0 +; VI-SAFE-SDAG-NEXT: s_cmp_eq_u32 s8, 3 +; VI-SAFE-SDAG-NEXT: s_cselect_b64 s[8:9], -1, 0 +; VI-SAFE-SDAG-NEXT: s_or_b64 s[4:5], s[8:9], s[4:5] +; VI-SAFE-SDAG-NEXT: s_lshr_b32 s8, s11, 2 +; VI-SAFE-SDAG-NEXT: s_cmp_lg_u64 s[4:5], 0 +; VI-SAFE-SDAG-NEXT: s_addc_u32 s4, s8, 0 +; VI-SAFE-SDAG-NEXT: s_cmp_lt_i32 s10, 31 +; VI-SAFE-SDAG-NEXT: s_cselect_b32 s8, s4, 0x7c00 +; VI-SAFE-SDAG-NEXT: s_cmp_lg_u32 s6, 0 +; 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_lshlrev_b32_e32 v0, 9, v0 +; VI-SAFE-SDAG-NEXT: s_cmpk_eq_i32 s10, 0x40f +; VI-SAFE-SDAG-NEXT: v_or_b32_e32 v0, 0x7c00, v0 +; VI-SAFE-SDAG-NEXT: v_mov_b32_e32 v1, s8 +; VI-SAFE-SDAG-NEXT: s_cselect_b64 vcc, -1, 0 +; VI-SAFE-SDAG-NEXT: s_lshr_b32 s4, s7, 16 +; VI-SAFE-SDAG-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc +; VI-SAFE-SDAG-NEXT: s_and_b32 s4, s4, 0x8000 +; VI-SAFE-SDAG-NEXT: v_or_b32_e32 v0, s4, v0 +; 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: +; VI-GISEL: ; %bb.0: +; VI-GISEL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 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_movk_i32 s5, 0x7e00 +; VI-GISEL-NEXT: s_cselect_b32 s5, s5, 0x7c00 +; 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_lshr_b32 s8, s2, s7 +; 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_and_b32 s6, s6, 1 +; 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: +; VI-UNSAFE-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 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: +; GFX10-SAFE-SDAG: ; %bb.0: +; GFX10-SAFE-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 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: s_addk_i32 s2, 0xfc10 +; GFX10-SAFE-SDAG-NEXT: v_med3_i32 v1, s5, 0, 13 +; GFX10-SAFE-SDAG-NEXT: v_readfirstlane_b32 s5, v0 +; GFX10-SAFE-SDAG-NEXT: s_lshl_b32 s7, s2, 12 +; 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_or_b32 s7, s4, s7 +; GFX10-SAFE-SDAG-NEXT: s_lshr_b32 s6, s5, s6 +; GFX10-SAFE-SDAG-NEXT: v_lshlrev_b32_e64 v0, v1, s6 +; GFX10-SAFE-SDAG-NEXT: v_cmp_ne_u32_e32 vcc_lo, s5, v0 +; GFX10-SAFE-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX10-SAFE-SDAG-NEXT: v_readfirstlane_b32 s5, v0 +; GFX10-SAFE-SDAG-NEXT: s_or_b32 s5, s6, s5 +; GFX10-SAFE-SDAG-NEXT: s_cmp_lt_i32 s2, 1 +; GFX10-SAFE-SDAG-NEXT: s_cselect_b32 s5, s5, s7 +; 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_cmp_lg_u32 s6, 0 +; GFX10-SAFE-SDAG-NEXT: s_addc_u32 s5, s5, 0 +; GFX10-SAFE-SDAG-NEXT: s_cmp_lt_i32 s2, 31 +; 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, -1, 0 +; GFX10-SAFE-SDAG-NEXT: s_cmpk_eq_i32 s2, 0x40f +; GFX10-SAFE-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, s4 +; GFX10-SAFE-SDAG-NEXT: s_cselect_b32 vcc_lo, -1, 0 +; GFX10-SAFE-SDAG-NEXT: s_lshr_b32 s2, s3, 16 +; GFX10-SAFE-SDAG-NEXT: s_mov_b32 s3, 0x31016000 +; GFX10-SAFE-SDAG-NEXT: s_and_b32 s2, s2, 0x8000 +; GFX10-SAFE-SDAG-NEXT: v_lshlrev_b32_e32 v0, 9, v0 +; GFX10-SAFE-SDAG-NEXT: v_or_b32_e32 v0, 0x7c00, v0 +; GFX10-SAFE-SDAG-NEXT: v_cndmask_b32_e32 v0, s5, v0, vcc_lo +; GFX10-SAFE-SDAG-NEXT: v_or_b32_e32 v0, s2, v0 +; 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: +; GFX10-GISEL: ; %bb.0: +; GFX10-GISEL-NEXT: s_load_dwordx4 s[0:3], s[0:1], 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_movk_i32 s5, 0x7e00 +; GFX10-GISEL-NEXT: s_cmp_lg_u32 s2, 0 +; GFX10-GISEL-NEXT: s_cselect_b32 s5, s5, 0x7c00 +; GFX10-GISEL-NEXT: s_sub_i32 s6, 1, s4 +; GFX10-GISEL-NEXT: s_or_b32 s7, s2, 0x1000 +; GFX10-GISEL-NEXT: s_max_i32 s6, s6, 0 +; GFX10-GISEL-NEXT: s_lshl_b32 s9, s4, 12 +; GFX10-GISEL-NEXT: s_min_i32 s6, s6, 13 +; GFX10-GISEL-NEXT: s_or_b32 s2, s2, s9 +; GFX10-GISEL-NEXT: s_lshr_b32 s8, s7, s6 +; GFX10-GISEL-NEXT: s_lshl_b32 s6, s8, s6 +; GFX10-GISEL-NEXT: s_cmp_lg_u32 s6, s7 +; GFX10-GISEL-NEXT: s_cselect_b32 s6, 1, 0 +; GFX10-GISEL-NEXT: s_or_b32 s6, s8, 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_and_b32 s6, s6, 1 +; 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: +; GFX10-UNSAFE-SDAG-NEXT: s_load_dwordx4 s[0:3], s[0:1], 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: +; GFX11-SAFE-SDAG: ; %bb.0: +; GFX11-SAFE-SDAG-NEXT: s_load_b128 s[0:3], s[0:1], 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_addk_i32 s2, 0xfc10 +; 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_lshl_b32 s7, s2, 12 +; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-SAFE-SDAG-NEXT: v_readfirstlane_b32 s6, v1 +; GFX11-SAFE-SDAG-NEXT: s_or_b32 s4, s4, s5 +; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-SAFE-SDAG-NEXT: s_or_b32 s5, s4, 0x1000 +; GFX11-SAFE-SDAG-NEXT: s_or_b32 s7, s4, s7 +; GFX11-SAFE-SDAG-NEXT: s_lshr_b32 s6, s5, s6 +; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-SAFE-SDAG-NEXT: v_lshlrev_b32_e64 v0, v1, s6 +; GFX11-SAFE-SDAG-NEXT: v_cmp_ne_u32_e32 vcc_lo, s5, v0 +; GFX11-SAFE-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-SAFE-SDAG-NEXT: v_readfirstlane_b32 s5, v0 +; GFX11-SAFE-SDAG-NEXT: s_or_b32 s5, s6, s5 +; GFX11-SAFE-SDAG-NEXT: s_cmp_lt_i32 s2, 1 +; GFX11-SAFE-SDAG-NEXT: s_cselect_b32 s5, s5, s7 +; 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_cmp_lg_u32 s6, 0 +; GFX11-SAFE-SDAG-NEXT: s_addc_u32 s5, s5, 0 +; GFX11-SAFE-SDAG-NEXT: s_cmp_lt_i32 s2, 31 +; 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, -1, 0 +; GFX11-SAFE-SDAG-NEXT: s_cmpk_eq_i32 s2, 0x40f +; GFX11-SAFE-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, s4 +; GFX11-SAFE-SDAG-NEXT: s_cselect_b32 vcc_lo, -1, 0 +; GFX11-SAFE-SDAG-NEXT: s_lshr_b32 s2, s3, 16 +; GFX11-SAFE-SDAG-NEXT: s_mov_b32 s3, 0x31016000 +; GFX11-SAFE-SDAG-NEXT: s_and_b32 s2, s2, 0x8000 +; GFX11-SAFE-SDAG-NEXT: v_lshlrev_b32_e32 v0, 9, v0 +; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-SAFE-SDAG-NEXT: v_or_b32_e32 v0, 0x7c00, v0 +; GFX11-SAFE-SDAG-NEXT: v_cndmask_b32_e32 v0, s5, v0, vcc_lo +; GFX11-SAFE-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-SAFE-SDAG-NEXT: v_or_b32_e32 v0, s2, v0 +; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-SAFE-SDAG-NEXT: s_endpgm +; +; GFX11-GISEL-LABEL: fptrunc_f64_to_f16: +; GFX11-GISEL: ; %bb.0: +; GFX11-GISEL-NEXT: s_load_b128 s[0:3], s[0:1], 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) +; GFX11-GISEL-NEXT: s_or_b32 s2, s5, s2 +; GFX11-GISEL-NEXT: s_movk_i32 s5, 0x7e00 +; GFX11-GISEL-NEXT: s_cmp_lg_u32 s2, 0 +; GFX11-GISEL-NEXT: s_cselect_b32 s5, s5, 0x7c00 +; GFX11-GISEL-NEXT: s_sub_i32 s6, 1, s4 +; GFX11-GISEL-NEXT: s_or_b32 s7, s2, 0x1000 +; GFX11-GISEL-NEXT: s_max_i32 s6, s6, 0 +; GFX11-GISEL-NEXT: s_lshl_b32 s9, s4, 12 +; GFX11-GISEL-NEXT: s_min_i32 s6, s6, 13 +; GFX11-GISEL-NEXT: s_or_b32 s2, s2, s9 +; GFX11-GISEL-NEXT: s_lshr_b32 s8, s7, s6 +; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) +; GFX11-GISEL-NEXT: s_lshl_b32 s6, s8, s6 +; GFX11-GISEL-NEXT: s_cmp_lg_u32 s6, s7 +; 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, s8, 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_and_b32 s6, s6, 1 +; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) +; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-GISEL-NEXT: s_endpgm +; +; GFX11-UNSAFE-SDAG-LABEL: fptrunc_f64_to_f16: +; GFX11-UNSAFE-SDAG: ; %bb.0: +; GFX11-UNSAFE-SDAG-NEXT: s_load_b128 s[0:3], s[0:1], 0x24 +; GFX11-UNSAFE-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-UNSAFE-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[2:3] +; GFX11-UNSAFE-SDAG-NEXT: s_mov_b32 s3, 0x31016000 +; GFX11-UNSAFE-SDAG-NEXT: s_mov_b32 s2, -1 +; GFX11-UNSAFE-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-UNSAFE-SDAG-NEXT: v_cvt_f16_f32_e32 v0, v0 +; GFX11-UNSAFE-SDAG-NEXT: buffer_store_b16 v0, off, s[0:3], 0 +; GFX11-UNSAFE-SDAG-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-UNSAFE-SDAG-NEXT: s_endpgm %result = fptrunc double %in to half %result_i16 = bitcast half %result to i16 store i16 %result_i16, ptr addrspace(1) %out @@ -285,43 +549,83 @@ ; SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 ; SI-NEXT: s_endpgm ; -; VI-LABEL: fptrunc_v2f64_to_v2f32: -; VI: ; %bb.0: -; VI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x34 -; VI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24 -; VI-NEXT: s_mov_b32 s3, 0xf000 -; VI-NEXT: s_mov_b32 s2, -1 -; VI-NEXT: s_waitcnt lgkmcnt(0) -; VI-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] -; VI-NEXT: v_cvt_f32_f64_e32 v0, s[4:5] -; VI-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 -; VI-NEXT: s_endpgm -; -; VI-UNSAFE-LABEL: fptrunc_v2f64_to_v2f32: -; VI-UNSAFE: ; %bb.0: -; VI-UNSAFE-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x34 -; VI-UNSAFE-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24 -; VI-UNSAFE-NEXT: s_mov_b32 s3, 0xf000 -; VI-UNSAFE-NEXT: s_mov_b32 s2, -1 -; VI-UNSAFE-NEXT: s_waitcnt lgkmcnt(0) -; VI-UNSAFE-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] -; VI-UNSAFE-NEXT: v_cvt_f32_f64_e32 v0, s[4:5] -; VI-UNSAFE-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 -; VI-UNSAFE-NEXT: s_endpgm -; -; GFX11-LABEL: fptrunc_v2f64_to_v2f32: -; GFX11: ; %bb.0: -; GFX11-NEXT: s_clause 0x1 -; GFX11-NEXT: s_load_b128 s[4:7], s[0:1], 0x34 -; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x24 -; GFX11-NEXT: s_mov_b32 s3, 0x31016000 -; GFX11-NEXT: s_mov_b32 s2, -1 -; GFX11-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] -; GFX11-NEXT: v_cvt_f32_f64_e32 v0, s[4:5] -; GFX11-NEXT: buffer_store_b64 v[0:1], off, s[0:3], 0 -; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) -; GFX11-NEXT: s_endpgm +; VI-SDAG-LABEL: fptrunc_v2f64_to_v2f32: +; VI-SDAG: ; %bb.0: +; VI-SDAG-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x34 +; VI-SDAG-NEXT: s_load_dwordx2 s[0:1], s[0:1], 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 v1, s[6:7] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[4:5] +; VI-SDAG-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 +; VI-SDAG-NEXT: s_endpgm +; +; VI-GISEL-LABEL: fptrunc_v2f64_to_v2f32: +; VI-GISEL: ; %bb.0: +; VI-GISEL-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x34 +; VI-GISEL-NEXT: s_load_dwordx2 s[0:1], s[0:1], 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[4:5] +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] +; VI-GISEL-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 +; VI-GISEL-NEXT: s_endpgm +; +; GFX10-SDAG-LABEL: fptrunc_v2f64_to_v2f32: +; GFX10-SDAG: ; %bb.0: +; GFX10-SDAG-NEXT: s_clause 0x1 +; GFX10-SDAG-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x34 +; GFX10-SDAG-NEXT: s_load_dwordx2 s[0:1], s[0:1], 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 v1, s[6:7] +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[4:5] +; 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: +; GFX10-GISEL: ; %bb.0: +; GFX10-GISEL-NEXT: s_clause 0x1 +; GFX10-GISEL-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x34 +; GFX10-GISEL-NEXT: s_load_dwordx2 s[0:1], s[0:1], 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[4:5] +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] +; 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: +; GFX11-SDAG: ; %bb.0: +; GFX11-SDAG-NEXT: s_clause 0x1 +; GFX11-SDAG-NEXT: s_load_b128 s[4:7], s[0:1], 0x34 +; GFX11-SDAG-NEXT: s_load_b64 s[0:1], s[0:1], 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 v1, s[6:7] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[4:5] +; GFX11-SDAG-NEXT: buffer_store_b64 v[0:1], off, s[0:3], 0 +; GFX11-SDAG-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-SDAG-NEXT: s_endpgm +; +; GFX11-GISEL-LABEL: fptrunc_v2f64_to_v2f32: +; GFX11-GISEL: ; %bb.0: +; GFX11-GISEL-NEXT: s_clause 0x1 +; GFX11-GISEL-NEXT: s_load_b128 s[4:7], s[0:1], 0x34 +; GFX11-GISEL-NEXT: s_load_b64 s[0:1], s[0:1], 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[4:5] +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] +; GFX11-GISEL-NEXT: buffer_store_b64 v[0:1], off, s[0:3], 0 +; GFX11-GISEL-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-GISEL-NEXT: s_endpgm %result = fptrunc <2 x double> %in to <2 x float> store <2 x float> %result, ptr addrspace(1) %out ret void @@ -343,49 +647,92 @@ ; SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 ; SI-NEXT: s_endpgm ; -; VI-LABEL: fptrunc_v3f64_to_v3f32: -; VI: ; %bb.0: -; VI-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x54 -; VI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x44 -; VI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24 -; VI-NEXT: s_waitcnt lgkmcnt(0) -; VI-NEXT: v_cvt_f32_f64_e32 v2, s[2:3] -; VI-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] -; VI-NEXT: v_cvt_f32_f64_e32 v0, s[4:5] -; VI-NEXT: s_mov_b32 s3, 0xf000 -; VI-NEXT: s_mov_b32 s2, -1 -; VI-NEXT: buffer_store_dwordx3 v[0:2], off, s[0:3], 0 -; VI-NEXT: s_endpgm -; -; VI-UNSAFE-LABEL: fptrunc_v3f64_to_v3f32: -; VI-UNSAFE: ; %bb.0: -; VI-UNSAFE-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x54 -; VI-UNSAFE-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x44 -; VI-UNSAFE-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24 -; VI-UNSAFE-NEXT: s_waitcnt lgkmcnt(0) -; VI-UNSAFE-NEXT: v_cvt_f32_f64_e32 v2, s[2:3] -; VI-UNSAFE-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] -; VI-UNSAFE-NEXT: v_cvt_f32_f64_e32 v0, s[4:5] -; VI-UNSAFE-NEXT: s_mov_b32 s3, 0xf000 -; VI-UNSAFE-NEXT: s_mov_b32 s2, -1 -; VI-UNSAFE-NEXT: buffer_store_dwordx3 v[0:2], off, s[0:3], 0 -; VI-UNSAFE-NEXT: s_endpgm -; -; GFX11-LABEL: fptrunc_v3f64_to_v3f32: -; GFX11: ; %bb.0: -; GFX11-NEXT: s_clause 0x2 -; GFX11-NEXT: s_load_b64 s[2:3], s[0:1], 0x54 -; GFX11-NEXT: s_load_b128 s[4:7], s[0:1], 0x44 -; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x24 -; GFX11-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NEXT: v_cvt_f32_f64_e32 v2, s[2:3] -; GFX11-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] -; GFX11-NEXT: v_cvt_f32_f64_e32 v0, s[4:5] -; GFX11-NEXT: s_mov_b32 s3, 0x31016000 -; GFX11-NEXT: s_mov_b32 s2, -1 -; GFX11-NEXT: buffer_store_b96 v[0:2], off, s[0:3], 0 -; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) -; GFX11-NEXT: s_endpgm +; VI-SDAG-LABEL: fptrunc_v3f64_to_v3f32: +; VI-SDAG: ; %bb.0: +; VI-SDAG-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x54 +; VI-SDAG-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x44 +; VI-SDAG-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24 +; VI-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v2, s[2:3] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[4:5] +; VI-SDAG-NEXT: s_mov_b32 s3, 0xf000 +; VI-SDAG-NEXT: s_mov_b32 s2, -1 +; VI-SDAG-NEXT: buffer_store_dwordx3 v[0:2], off, s[0:3], 0 +; VI-SDAG-NEXT: s_endpgm +; +; VI-GISEL-LABEL: fptrunc_v3f64_to_v3f32: +; VI-GISEL: ; %bb.0: +; VI-GISEL-NEXT: s_load_dwordx8 s[4:11], s[0:1], 0x44 +; VI-GISEL-NEXT: s_load_dwordx2 s[0:1], s[0:1], 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[4:5] +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v2, s[8:9] +; 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: +; GFX10-SDAG: ; %bb.0: +; GFX10-SDAG-NEXT: s_clause 0x2 +; GFX10-SDAG-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x54 +; GFX10-SDAG-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x44 +; GFX10-SDAG-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24 +; GFX10-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v2, s[2:3] +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[4:5] +; GFX10-SDAG-NEXT: s_mov_b32 s3, 0x31016000 +; GFX10-SDAG-NEXT: s_mov_b32 s2, -1 +; 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: +; GFX10-GISEL: ; %bb.0: +; GFX10-GISEL-NEXT: s_clause 0x1 +; GFX10-GISEL-NEXT: s_load_dwordx8 s[4:11], s[0:1], 0x44 +; GFX10-GISEL-NEXT: s_load_dwordx2 s[0:1], s[0:1], 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[4:5] +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v2, s[8:9] +; 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: +; GFX11-SDAG: ; %bb.0: +; GFX11-SDAG-NEXT: s_clause 0x2 +; GFX11-SDAG-NEXT: s_load_b64 s[2:3], s[0:1], 0x54 +; GFX11-SDAG-NEXT: s_load_b128 s[4:7], s[0:1], 0x44 +; GFX11-SDAG-NEXT: s_load_b64 s[0:1], s[0:1], 0x24 +; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v2, s[2:3] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[4:5] +; GFX11-SDAG-NEXT: s_mov_b32 s3, 0x31016000 +; GFX11-SDAG-NEXT: s_mov_b32 s2, -1 +; GFX11-SDAG-NEXT: buffer_store_b96 v[0:2], off, s[0:3], 0 +; GFX11-SDAG-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-SDAG-NEXT: s_endpgm +; +; GFX11-GISEL-LABEL: fptrunc_v3f64_to_v3f32: +; GFX11-GISEL: ; %bb.0: +; GFX11-GISEL-NEXT: s_clause 0x1 +; GFX11-GISEL-NEXT: s_load_b256 s[4:11], s[0:1], 0x44 +; GFX11-GISEL-NEXT: s_load_b64 s[0:1], s[0:1], 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[4:5] +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v2, s[8:9] +; GFX11-GISEL-NEXT: buffer_store_b96 v[0:2], off, s[0:3], 0 +; GFX11-GISEL-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-GISEL-NEXT: s_endpgm %result = fptrunc <3 x double> %in to <3 x float> store <3 x float> %result, ptr addrspace(1) %out ret void @@ -406,49 +753,95 @@ ; SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 ; SI-NEXT: s_endpgm ; -; VI-LABEL: fptrunc_v4f64_to_v4f32: -; VI: ; %bb.0: -; VI-NEXT: s_load_dwordx8 s[4:11], s[0:1], 0x44 -; VI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24 -; VI-NEXT: s_mov_b32 s3, 0xf000 -; VI-NEXT: s_mov_b32 s2, -1 -; VI-NEXT: s_waitcnt lgkmcnt(0) -; VI-NEXT: v_cvt_f32_f64_e32 v3, s[10:11] -; VI-NEXT: v_cvt_f32_f64_e32 v2, s[8:9] -; VI-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] -; VI-NEXT: v_cvt_f32_f64_e32 v0, s[4:5] -; VI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 -; VI-NEXT: s_endpgm -; -; VI-UNSAFE-LABEL: fptrunc_v4f64_to_v4f32: -; VI-UNSAFE: ; %bb.0: -; VI-UNSAFE-NEXT: s_load_dwordx8 s[4:11], s[0:1], 0x44 -; VI-UNSAFE-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24 -; VI-UNSAFE-NEXT: s_mov_b32 s3, 0xf000 -; VI-UNSAFE-NEXT: s_mov_b32 s2, -1 -; VI-UNSAFE-NEXT: s_waitcnt lgkmcnt(0) -; VI-UNSAFE-NEXT: v_cvt_f32_f64_e32 v3, s[10:11] -; VI-UNSAFE-NEXT: v_cvt_f32_f64_e32 v2, s[8:9] -; VI-UNSAFE-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] -; VI-UNSAFE-NEXT: v_cvt_f32_f64_e32 v0, s[4:5] -; VI-UNSAFE-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 -; VI-UNSAFE-NEXT: s_endpgm -; -; GFX11-LABEL: fptrunc_v4f64_to_v4f32: -; GFX11: ; %bb.0: -; GFX11-NEXT: s_clause 0x1 -; GFX11-NEXT: s_load_b256 s[4:11], s[0:1], 0x44 -; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x24 -; GFX11-NEXT: s_mov_b32 s3, 0x31016000 -; GFX11-NEXT: s_mov_b32 s2, -1 -; GFX11-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NEXT: v_cvt_f32_f64_e32 v3, s[10:11] -; GFX11-NEXT: v_cvt_f32_f64_e32 v2, s[8:9] -; GFX11-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] -; GFX11-NEXT: v_cvt_f32_f64_e32 v0, s[4:5] -; GFX11-NEXT: buffer_store_b128 v[0:3], off, s[0:3], 0 -; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) -; GFX11-NEXT: s_endpgm +; VI-SDAG-LABEL: fptrunc_v4f64_to_v4f32: +; VI-SDAG: ; %bb.0: +; VI-SDAG-NEXT: s_load_dwordx8 s[4:11], s[0:1], 0x44 +; VI-SDAG-NEXT: s_load_dwordx2 s[0:1], s[0:1], 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[10:11] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v2, s[8:9] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[4:5] +; 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: +; VI-GISEL: ; %bb.0: +; VI-GISEL-NEXT: s_load_dwordx8 s[4:11], s[0:1], 0x44 +; VI-GISEL-NEXT: s_load_dwordx2 s[0:1], s[0:1], 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[4:5] +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v2, s[8:9] +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v3, s[10:11] +; 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: +; GFX10-SDAG: ; %bb.0: +; GFX10-SDAG-NEXT: s_clause 0x1 +; GFX10-SDAG-NEXT: s_load_dwordx8 s[4:11], s[0:1], 0x44 +; GFX10-SDAG-NEXT: s_load_dwordx2 s[0:1], s[0:1], 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[10:11] +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v2, s[8:9] +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[4:5] +; 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: +; GFX10-GISEL: ; %bb.0: +; GFX10-GISEL-NEXT: s_clause 0x1 +; GFX10-GISEL-NEXT: s_load_dwordx8 s[4:11], s[0:1], 0x44 +; GFX10-GISEL-NEXT: s_load_dwordx2 s[0:1], s[0:1], 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[4:5] +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v2, s[8:9] +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v3, s[10:11] +; 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: +; GFX11-SDAG: ; %bb.0: +; GFX11-SDAG-NEXT: s_clause 0x1 +; GFX11-SDAG-NEXT: s_load_b256 s[4:11], s[0:1], 0x44 +; GFX11-SDAG-NEXT: s_load_b64 s[0:1], s[0:1], 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[10:11] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v2, s[8:9] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[4:5] +; GFX11-SDAG-NEXT: buffer_store_b128 v[0:3], off, s[0:3], 0 +; GFX11-SDAG-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-SDAG-NEXT: s_endpgm +; +; GFX11-GISEL-LABEL: fptrunc_v4f64_to_v4f32: +; GFX11-GISEL: ; %bb.0: +; GFX11-GISEL-NEXT: s_clause 0x1 +; GFX11-GISEL-NEXT: s_load_b256 s[4:11], s[0:1], 0x44 +; GFX11-GISEL-NEXT: s_load_b64 s[0:1], s[0:1], 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[4:5] +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v2, s[8:9] +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v3, s[10:11] +; GFX11-GISEL-NEXT: buffer_store_b128 v[0:3], off, s[0:3], 0 +; GFX11-GISEL-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-GISEL-NEXT: s_endpgm %result = fptrunc <4 x double> %in to <4 x float> store <4 x float> %result, ptr addrspace(1) %out ret void @@ -474,65 +867,127 @@ ; SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 ; SI-NEXT: s_endpgm ; -; VI-LABEL: fptrunc_v8f64_to_v8f32: -; VI: ; %bb.0: -; VI-NEXT: s_load_dwordx16 s[4:19], s[0:1], 0x64 -; VI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24 -; VI-NEXT: s_mov_b32 s3, 0xf000 -; VI-NEXT: s_mov_b32 s2, -1 -; VI-NEXT: s_waitcnt lgkmcnt(0) -; VI-NEXT: v_cvt_f32_f64_e32 v7, s[18:19] -; VI-NEXT: v_cvt_f32_f64_e32 v6, s[16:17] -; VI-NEXT: v_cvt_f32_f64_e32 v5, s[14:15] -; VI-NEXT: v_cvt_f32_f64_e32 v4, s[12:13] -; VI-NEXT: v_cvt_f32_f64_e32 v3, s[10:11] -; VI-NEXT: v_cvt_f32_f64_e32 v2, s[8:9] -; VI-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] -; VI-NEXT: v_cvt_f32_f64_e32 v0, s[4:5] -; VI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16 -; VI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 -; VI-NEXT: s_endpgm -; -; VI-UNSAFE-LABEL: fptrunc_v8f64_to_v8f32: -; VI-UNSAFE: ; %bb.0: -; VI-UNSAFE-NEXT: s_load_dwordx16 s[4:19], s[0:1], 0x64 -; VI-UNSAFE-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x24 -; VI-UNSAFE-NEXT: s_mov_b32 s3, 0xf000 -; VI-UNSAFE-NEXT: s_mov_b32 s2, -1 -; VI-UNSAFE-NEXT: s_waitcnt lgkmcnt(0) -; VI-UNSAFE-NEXT: v_cvt_f32_f64_e32 v7, s[18:19] -; VI-UNSAFE-NEXT: v_cvt_f32_f64_e32 v6, s[16:17] -; VI-UNSAFE-NEXT: v_cvt_f32_f64_e32 v5, s[14:15] -; VI-UNSAFE-NEXT: v_cvt_f32_f64_e32 v4, s[12:13] -; VI-UNSAFE-NEXT: v_cvt_f32_f64_e32 v3, s[10:11] -; VI-UNSAFE-NEXT: v_cvt_f32_f64_e32 v2, s[8:9] -; VI-UNSAFE-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] -; VI-UNSAFE-NEXT: v_cvt_f32_f64_e32 v0, s[4:5] -; VI-UNSAFE-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:16 -; VI-UNSAFE-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 -; VI-UNSAFE-NEXT: s_endpgm -; -; GFX11-LABEL: fptrunc_v8f64_to_v8f32: -; GFX11: ; %bb.0: -; GFX11-NEXT: s_clause 0x1 -; GFX11-NEXT: s_load_b512 s[4:19], s[0:1], 0x64 -; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x24 -; GFX11-NEXT: s_mov_b32 s3, 0x31016000 -; GFX11-NEXT: s_mov_b32 s2, -1 -; GFX11-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NEXT: v_cvt_f32_f64_e32 v7, s[18:19] -; GFX11-NEXT: v_cvt_f32_f64_e32 v6, s[16:17] -; GFX11-NEXT: v_cvt_f32_f64_e32 v5, s[14:15] -; GFX11-NEXT: v_cvt_f32_f64_e32 v4, s[12:13] -; GFX11-NEXT: v_cvt_f32_f64_e32 v3, s[10:11] -; GFX11-NEXT: v_cvt_f32_f64_e32 v2, s[8:9] -; GFX11-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] -; GFX11-NEXT: v_cvt_f32_f64_e32 v0, s[4:5] -; GFX11-NEXT: s_clause 0x1 -; GFX11-NEXT: buffer_store_b128 v[4:7], off, s[0:3], 0 offset:16 -; GFX11-NEXT: buffer_store_b128 v[0:3], off, s[0:3], 0 -; GFX11-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS) -; GFX11-NEXT: s_endpgm +; VI-SDAG-LABEL: fptrunc_v8f64_to_v8f32: +; VI-SDAG: ; %bb.0: +; VI-SDAG-NEXT: s_load_dwordx16 s[4:19], s[0:1], 0x64 +; VI-SDAG-NEXT: s_load_dwordx2 s[0:1], s[0:1], 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[18:19] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v6, s[16:17] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v5, s[14:15] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v4, s[12:13] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v3, s[10:11] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v2, s[8:9] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] +; VI-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[4:5] +; 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: +; VI-GISEL: ; %bb.0: +; VI-GISEL-NEXT: s_load_dwordx16 s[4:19], s[0:1], 0x64 +; VI-GISEL-NEXT: s_load_dwordx2 s[0:1], s[0:1], 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[4:5] +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v2, s[8:9] +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v3, s[10:11] +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v4, s[12:13] +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v5, s[14:15] +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v6, s[16:17] +; VI-GISEL-NEXT: v_cvt_f32_f64_e32 v7, s[18:19] +; 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: +; GFX10-SDAG: ; %bb.0: +; GFX10-SDAG-NEXT: s_clause 0x1 +; GFX10-SDAG-NEXT: s_load_dwordx16 s[4:19], s[0:1], 0x64 +; GFX10-SDAG-NEXT: s_load_dwordx2 s[0:1], s[0:1], 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[18:19] +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v6, s[16:17] +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v5, s[14:15] +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v4, s[12:13] +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v3, s[10:11] +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v2, s[8:9] +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] +; GFX10-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[4:5] +; 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: +; GFX10-GISEL: ; %bb.0: +; GFX10-GISEL-NEXT: s_clause 0x1 +; GFX10-GISEL-NEXT: s_load_dwordx16 s[4:19], s[0:1], 0x64 +; GFX10-GISEL-NEXT: s_load_dwordx2 s[0:1], s[0:1], 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[4:5] +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v2, s[8:9] +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v3, s[10:11] +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v4, s[12:13] +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v5, s[14:15] +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v6, s[16:17] +; GFX10-GISEL-NEXT: v_cvt_f32_f64_e32 v7, s[18:19] +; 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: +; GFX11-SDAG: ; %bb.0: +; GFX11-SDAG-NEXT: s_clause 0x1 +; GFX11-SDAG-NEXT: s_load_b512 s[4:19], s[0:1], 0x64 +; GFX11-SDAG-NEXT: s_load_b64 s[0:1], s[0:1], 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[18:19] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v6, s[16:17] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v5, s[14:15] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v4, s[12:13] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v3, s[10:11] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v2, s[8:9] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] +; GFX11-SDAG-NEXT: v_cvt_f32_f64_e32 v0, s[4:5] +; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; GFX11-SDAG-NEXT: s_endpgm +; +; GFX11-GISEL-LABEL: fptrunc_v8f64_to_v8f32: +; GFX11-GISEL: ; %bb.0: +; GFX11-GISEL-NEXT: s_clause 0x1 +; GFX11-GISEL-NEXT: s_load_b512 s[4:19], s[0:1], 0x64 +; GFX11-GISEL-NEXT: s_load_b64 s[0:1], s[0:1], 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[4:5] +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v1, s[6:7] +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v2, s[8:9] +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v3, s[10:11] +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v4, s[12:13] +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v5, s[14:15] +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v6, s[16:17] +; GFX11-GISEL-NEXT: v_cvt_f32_f64_e32 v7, s[18:19] +; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS) +; 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