From: Lei Huang Date: Mon, 14 Aug 2017 18:09:29 +0000 (+0000) Subject: [PowerPC] Add codegen for VSX word extract convert to FP X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=af34a3a6d6b83a5d9e0b50fcf54fc70bf61be47d;p=llvm [PowerPC] Add codegen for VSX word extract convert to FP Add codegen for VSX word extract conversion from signed/unsigned to single/double precision. For UINT_TO_FP: Extract word unsigned and convert to float was implemented in https://reviews.llvm.org/D20239. Here we will add the missing extract integer and conversion to double. This utilizes the new P9 instruction xxextractuw to extracting an integer element when the result will be converted to double thereby saving 2 direct moves (VSR <-> GPR). For SINT_TO_FP: We will implement the following sequence which will also reduce the number of instructions by saving 2 direct moves. v4i32->f32: xxspltw xvcvsxwsp xscvspdpn v4i32->f64: xxspltw xvcvsxwdp Differential Revision: https://reviews.llvm.org/D35859 git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310866 91177308-0d34-0410-b5e6-96231b3b80d8 --- diff --git a/lib/Target/PowerPC/PPCInstrVSX.td b/lib/Target/PowerPC/PPCInstrVSX.td index 942e8b392b8..21e4a9d4bf0 100644 --- a/lib/Target/PowerPC/PPCInstrVSX.td +++ b/lib/Target/PowerPC/PPCInstrVSX.td @@ -2550,6 +2550,44 @@ let AddedComplexity = 400, Predicates = [HasP9Vector] in { UseVSXReg; } // mayStore + let Predicates = [IsLittleEndian] in { + def: Pat<(f32 (PPCfcfids (PPCmtvsra (i32 (extractelt v4i32:$A, 0))))), + (f32 (XSCVSPDPN (XVCVSXWSP (XXSPLTW $A, 3))))>; + def: Pat<(f32 (PPCfcfids (PPCmtvsra (i32 (extractelt v4i32:$A, 1))))), + (f32 (XSCVSPDPN (XVCVSXWSP (XXSPLTW $A, 2))))>; + def: Pat<(f32 (PPCfcfids (PPCmtvsra (i32 (extractelt v4i32:$A, 2))))), + (f32 (XSCVSPDPN (XVCVSXWSP (XXSPLTW $A, 1))))>; + def: Pat<(f32 (PPCfcfids (PPCmtvsra (i32 (extractelt v4i32:$A, 3))))), + (f32 (XSCVSPDPN (XVCVSXWSP (XXSPLTW $A, 0))))>; + def: Pat<(f64 (PPCfcfid (PPCmtvsra (i32 (extractelt v4i32:$A, 0))))), + (f64 (COPY_TO_REGCLASS (XVCVSXWDP (XXSPLTW $A, 3)), VSFRC))>; + def: Pat<(f64 (PPCfcfid (PPCmtvsra (i32 (extractelt v4i32:$A, 1))))), + (f64 (COPY_TO_REGCLASS (XVCVSXWDP (XXSPLTW $A, 2)), VSFRC))>; + def: Pat<(f64 (PPCfcfid (PPCmtvsra (i32 (extractelt v4i32:$A, 2))))), + (f64 (COPY_TO_REGCLASS (XVCVSXWDP (XXSPLTW $A, 1)), VSFRC))>; + def: Pat<(f64 (PPCfcfid (PPCmtvsra (i32 (extractelt v4i32:$A, 3))))), + (f64 (COPY_TO_REGCLASS (XVCVSXWDP (XXSPLTW $A, 0)), VSFRC))>; + } + + let Predicates = [IsBigEndian] in { + def: Pat<(f32 (PPCfcfids (PPCmtvsra (i32 (extractelt v4i32:$A, 0))))), + (f32 (XSCVSPDPN (XVCVSXWSP (XXSPLTW $A, 0))))>; + def: Pat<(f32 (PPCfcfids (PPCmtvsra (i32 (extractelt v4i32:$A, 1))))), + (f32 (XSCVSPDPN (XVCVSXWSP (XXSPLTW $A, 1))))>; + def: Pat<(f32 (PPCfcfids (PPCmtvsra (i32 (extractelt v4i32:$A, 2))))), + (f32 (XSCVSPDPN (XVCVSXWSP (XXSPLTW $A, 2))))>; + def: Pat<(f32 (PPCfcfids (PPCmtvsra (i32 (extractelt v4i32:$A, 3))))), + (f32 (XSCVSPDPN (XVCVSXWSP (XXSPLTW $A, 3))))>; + def: Pat<(f64 (PPCfcfid (PPCmtvsra (i32 (extractelt v4i32:$A, 0))))), + (f64 (COPY_TO_REGCLASS (XVCVSXWDP (XXSPLTW $A, 0)), VSFRC))>; + def: Pat<(f64 (PPCfcfid (PPCmtvsra (i32 (extractelt v4i32:$A, 1))))), + (f64 (COPY_TO_REGCLASS (XVCVSXWDP (XXSPLTW $A, 1)), VSFRC))>; + def: Pat<(f64 (PPCfcfid (PPCmtvsra (i32 (extractelt v4i32:$A, 2))))), + (f64 (COPY_TO_REGCLASS (XVCVSXWDP (XXSPLTW $A, 2)), VSFRC))>; + def: Pat<(f64 (PPCfcfid (PPCmtvsra (i32 (extractelt v4i32:$A, 3))))), + (f64 (COPY_TO_REGCLASS (XVCVSXWDP (XXSPLTW $A, 3)), VSFRC))>; + } + // Patterns for which instructions from ISA 3.0 are a better match let Predicates = [IsLittleEndian, HasP9Vector] in { def : Pat<(f32 (PPCfcfidus (PPCmtvsrz (i32 (extractelt v4i32:$A, 0))))), @@ -2560,6 +2598,14 @@ let AddedComplexity = 400, Predicates = [HasP9Vector] in { (f32 (XSCVUXDSP (XXEXTRACTUW $A, 4)))>; def : Pat<(f32 (PPCfcfidus (PPCmtvsrz (i32 (extractelt v4i32:$A, 3))))), (f32 (XSCVUXDSP (XXEXTRACTUW $A, 0)))>; + def : Pat<(f64 (PPCfcfidu (PPCmtvsrz (i32 (extractelt v4i32:$A, 0))))), + (f64 (XSCVUXDDP (XXEXTRACTUW $A, 12)))>; + def : Pat<(f64 (PPCfcfidu (PPCmtvsrz (i32 (extractelt v4i32:$A, 1))))), + (f64 (XSCVUXDDP (XXEXTRACTUW $A, 8)))>; + def : Pat<(f64 (PPCfcfidu (PPCmtvsrz (i32 (extractelt v4i32:$A, 2))))), + (f64 (XSCVUXDDP (XXEXTRACTUW $A, 4)))>; + def : Pat<(f64 (PPCfcfidu (PPCmtvsrz (i32 (extractelt v4i32:$A, 3))))), + (f64 (XSCVUXDDP (XXEXTRACTUW $A, 0)))>; def : Pat<(v4i32 (insertelt v4i32:$A, i32:$B, 0)), (v4i32 (XXINSERTW v4i32:$A, AlignValues.I32_TO_BE_WORD1, 12))>; def : Pat<(v4i32 (insertelt v4i32:$A, i32:$B, 1)), @@ -2587,6 +2633,14 @@ let AddedComplexity = 400, Predicates = [HasP9Vector] in { (f32 (XSCVUXDSP (XXEXTRACTUW $A, 8)))>; def : Pat<(f32 (PPCfcfidus (PPCmtvsrz (i32 (extractelt v4i32:$A, 3))))), (f32 (XSCVUXDSP (XXEXTRACTUW $A, 12)))>; + def : Pat<(f64 (PPCfcfidu (PPCmtvsrz (i32 (extractelt v4i32:$A, 0))))), + (f64 (XSCVUXDDP (XXEXTRACTUW $A, 0)))>; + def : Pat<(f64 (PPCfcfidu (PPCmtvsrz (i32 (extractelt v4i32:$A, 1))))), + (f64 (XSCVUXDDP (XXEXTRACTUW $A, 4)))>; + def : Pat<(f64 (PPCfcfidu (PPCmtvsrz (i32 (extractelt v4i32:$A, 2))))), + (f64 (XSCVUXDDP (XXEXTRACTUW $A, 8)))>; + def : Pat<(f64 (PPCfcfidu (PPCmtvsrz (i32 (extractelt v4i32:$A, 3))))), + (f64 (XSCVUXDDP (XXEXTRACTUW $A, 12)))>; def : Pat<(v4i32 (insertelt v4i32:$A, i32:$B, 0)), (v4i32 (XXINSERTW v4i32:$A, AlignValues.I32_TO_BE_WORD1, 0))>; def : Pat<(v4i32 (insertelt v4i32:$A, i32:$B, 1)), diff --git a/test/CodeGen/PowerPC/p9-xxinsertw-xxextractuw.ll b/test/CodeGen/PowerPC/p9-xxinsertw-xxextractuw.ll index fe34bcb8563..783833d6b02 100644 --- a/test/CodeGen/PowerPC/p9-xxinsertw-xxextractuw.ll +++ b/test/CodeGen/PowerPC/p9-xxinsertw-xxextractuw.ll @@ -439,6 +439,69 @@ entry: ret float %conv } +; Verify we generate optimal code for unsigned vector int elem extract followed +; by conversion to double + +define double @conv2dlbTestui0(<4 x i32> %a) { +entry: +; CHECK-LABEL: conv2dlbTestui0 +; CHECK: xxextractuw [[SW:[0-9]+]], 34, 12 +; CHECK: xscvuxddp 1, [[SW]] +; CHECK-BE-LABEL: conv2dlbTestui0 +; CHECK-BE: xxextractuw [[CP:[0-9]+]], 34, 0 +; CHECK-BE: xscvuxddp 1, [[CP]] + %0 = extractelement <4 x i32> %a, i32 0 + %1 = uitofp i32 %0 to double + ret double %1 +} + +define double @conv2dlbTestui1(<4 x i32> %a) { +entry: +; CHECK-LABEL: conv2dlbTestui1 +; CHECK: xxextractuw [[SW:[0-9]+]], 34, 8 +; CHECK: xscvuxddp 1, [[SW]] +; CHECK-BE-LABEL: conv2dlbTestui1 +; CHECK-BE: xxextractuw [[CP:[0-9]+]], 34, 4 +; CHECK-BE: xscvuxddp 1, [[CP]] + %0 = extractelement <4 x i32> %a, i32 1 + %1 = uitofp i32 %0 to double + ret double %1 +} + +define double @conv2dlbTestui2(<4 x i32> %a) { +entry: +; CHECK-LABEL: conv2dlbTestui2 +; CHECK: xxextractuw [[SW:[0-9]+]], 34, 4 +; CHECK: xscvuxddp 1, [[SW]] +; CHECK-BE-LABEL: conv2dlbTestui2 +; CHECK-BE: xxextractuw [[CP:[0-9]+]], 34, 8 +; CHECK-BE: xscvuxddp 1, [[CP]] + %0 = extractelement <4 x i32> %a, i32 2 + %1 = uitofp i32 %0 to double + ret double %1 +} + +define double @conv2dlbTestui3(<4 x i32> %a) { +entry: +; CHECK-LABEL: conv2dlbTestui3 +; CHECK: xxextractuw [[SW:[0-9]+]], 34, 0 +; CHECK: xscvuxddp 1, [[SW]] +; CHECK-BE-LABEL: conv2dlbTestui3 +; CHECK-BE: xxextractuw [[CP:[0-9]+]], 34, 12 +; CHECK-BE: xscvuxddp 1, [[CP]] + %0 = extractelement <4 x i32> %a, i32 3 + %1 = uitofp i32 %0 to double + ret double %1 +} + +; verify we don't crash for variable elem extract +define double @conv2dlbTestuiVar(<4 x i32> %a, i32 zeroext %elem) { +entry: + %vecext = extractelement <4 x i32> %a, i32 %elem + %conv = uitofp i32 %vecext to double + ret double %conv +} + define <4 x float> @_Z10testInsEltILj0EDv4_ffET0_S1_T1_(<4 x float> %a, float %b) { entry: ; CHECK-LABEL: _Z10testInsEltILj0EDv4_ffET0_S1_T1_ diff --git a/test/CodeGen/PowerPC/remove-redundant-moves.ll b/test/CodeGen/PowerPC/remove-redundant-moves.ll index 6b845cbf380..40e39c64dea 100644 --- a/test/CodeGen/PowerPC/remove-redundant-moves.ll +++ b/test/CodeGen/PowerPC/remove-redundant-moves.ll @@ -105,3 +105,131 @@ entry: %1 = uitofp i64 %0 to float ret float %1 } + +define float @conv2fltTesti0(<4 x i32> %a) { +entry: +; CHECK-LABEL: conv2fltTesti0 +; CHECK: xxspltw [[SW:[0-9]+]], 34, 3 +; CHECK: xvcvsxwsp [[SW]], [[SW]] +; CHECK: xscvspdpn 1, [[SW]] +; CHECK-BE-LABEL: conv2fltTesti0 +; CHECK-BE: xxspltw [[CP:[0-9]+]], 34, 0 +; CHECK-BE: xvcvsxwsp [[CP]], [[CP]] +; CHECK-BE: xscvspdpn 1, [[CP]] + %vecext = extractelement <4 x i32> %a, i32 0 + %conv = sitofp i32 %vecext to float + ret float %conv +} + +define float @conv2fltTesti1(<4 x i32> %a) { +entry: +; CHECK-LABEL: conv2fltTesti1 +; CHECK: xxspltw [[SW:[0-9]+]], 34, 2 +; CHECK: xvcvsxwsp [[SW]], [[SW]] +; CHECK: xscvspdpn 1, [[SW]] +; CHECK-BE-LABEL: conv2fltTesti1 +; CHECK-BE: xxspltw [[CP:[0-9]+]], 34, 1 +; CHECK-BE: xvcvsxwsp [[CP]], [[CP]] +; CHECK-BE: xscvspdpn 1, [[CP]] + %vecext = extractelement <4 x i32> %a, i32 1 + %conv = sitofp i32 %vecext to float + ret float %conv +} + +define float @conv2fltTesti2(<4 x i32> %a) { +entry: +; CHECK-LABEL: conv2fltTesti2 +; CHECK: xxspltw [[SW:[0-9]+]], 34, 1 +; CHECK: xvcvsxwsp [[SW]], [[SW]] +; CHECK: xscvspdpn 1, [[SW]] +; CHECK-BE-LABEL: conv2fltTesti2 +; CHECK-BE: xxspltw [[CP:[0-9]+]], 34, 2 +; CHECK-BE: xvcvsxwsp [[CP]], [[CP]] +; CHECK-BE: xscvspdpn 1, [[CP]] + %vecext = extractelement <4 x i32> %a, i32 2 + %conv = sitofp i32 %vecext to float + ret float %conv +} + +define float @conv2fltTesti3(<4 x i32> %a) { +entry: +; CHECK-LABEL: conv2fltTesti3 +; CHECK: xxspltw [[SW:[0-9]+]], 34, 0 +; CHECK: xvcvsxwsp [[SW]], [[SW]] +; CHECK: xscvspdpn 1, [[SW]] +; CHECK-BE-LABEL: conv2fltTesti3 +; CHECK-BE: xxspltw [[CP:[0-9]+]], 34, 3 +; CHECK-BE: xvcvsxwsp [[CP]], [[CP]] +; CHECK-BE: xscvspdpn 1, [[CP]] + %vecext = extractelement <4 x i32> %a, i32 3 + %conv = sitofp i32 %vecext to float + ret float %conv +} + +; verify we don't crash for variable elem extract +define float @conv2fltTestiVar(<4 x i32> %a, i32 zeroext %elem) { +entry: + %vecext = extractelement <4 x i32> %a, i32 %elem + %conv = sitofp i32 %vecext to float + ret float %conv +} + +define double @conv2dblTesti0(<4 x i32> %a) { +entry: +; CHECK-LABEL: conv2dblTesti0 +; CHECK: xxspltw [[SW:[0-9]+]], 34, 3 +; CHECK: xvcvsxwdp 1, [[SW]] +; CHECK-BE-LABEL: conv2dblTesti0 +; CHECK-BE: xxspltw [[CP:[0-9]+]], 34, 0 +; CHECK-BE: xvcvsxwdp 1, [[CP]] + %vecext = extractelement <4 x i32> %a, i32 0 + %conv = sitofp i32 %vecext to double + ret double %conv +} + +define double @conv2dblTesti1(<4 x i32> %a) { +entry: +; CHECK-LABEL: conv2dblTesti1 +; CHECK: xxspltw [[SW:[0-9]+]], 34, 2 +; CHECK: xvcvsxwdp 1, [[SW]] +; CHECK-BE-LABEL: conv2dblTesti1 +; CHECK-BE: xxspltw [[CP:[0-9]+]], 34, 1 +; CHECK-BE: xvcvsxwdp 1, [[CP]] + %vecext = extractelement <4 x i32> %a, i32 1 + %conv = sitofp i32 %vecext to double + ret double %conv +} + +define double @conv2dblTesti2(<4 x i32> %a) { +entry: +; CHECK-LABEL: conv2dblTesti2 +; CHECK: xxspltw [[SW:[0-9]+]], 34, 1 +; CHECK: xvcvsxwdp 1, [[SW]] +; CHECK-BE-LABEL: conv2dblTesti2 +; CHECK-BE: xxspltw [[CP:[0-9]+]], 34, 2 +; CHECK-BE: xvcvsxwdp 1, [[CP]] + %vecext = extractelement <4 x i32> %a, i32 2 + %conv = sitofp i32 %vecext to double + ret double %conv +} + +define double @conv2dblTesti3(<4 x i32> %a) { +entry: +; CHECK-LABEL: conv2dblTesti3 +; CHECK: xxspltw [[SW:[0-9]+]], 34, 0 +; CHECK: xvcvsxwdp 1, [[SW]] +; CHECK-BE-LABEL: conv2dblTesti3 +; CHECK-BE: xxspltw [[CP:[0-9]+]], 34, 3 +; CHECK-BE: xvcvsxwdp 1, [[CP]] + %vecext = extractelement <4 x i32> %a, i32 3 + %conv = sitofp i32 %vecext to double + ret double %conv +} + +; verify we don't crash for variable elem extract +define double @conv2dblTestiVar(<4 x i32> %a, i32 zeroext %elem) { +entry: + %vecext = extractelement <4 x i32> %a, i32 %elem + %conv = sitofp i32 %vecext to double + ret double %conv +}