llvmorg-github-actions[bot] wrote:
<!--LLVM PR SUMMARY COMMENT--> @llvm/pr-subscribers-backend-loongarch Author: Jay Foad (jayfoad) <details> <summary>Changes</summary> Several patterns widen a 128-bit LSX vector to a 256-bit LASX vector with SUBREG_TO_REG, for instructions that only use the low 128 bits: vext2xv.*, and xvpermi.d followed by an instruction that only uses the low half of each 128-bit lane. SUBREG_TO_REG means that the instruction defining the LSX vector also defined the upper 128 bits, which is not the case and is not needed. Add an lsx_to_lasx helper that uses INSERT_SUBREG into an IMPLICIT_DEF instead, as other patterns already do for this kind of widening. Co-Authored-By: Claude Opus 5.5 <noreply@<!-- -->anthropic.com> --- <sub>Stack created with <a href="https://github.com/github/gh-stack">GitHub Stacks CLI</a> • <a href="https://gh.io/stacks-feedback">Give Feedback 💬</a></sub> --- Full diff: https://github.com/llvm/llvm-project/pull/242331.diff 1 Files Affected: - (modified) llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td (+34-31) ``````````diff diff --git a/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td b/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td index d34738f75ca64..464fa3ff5ab06 100644 --- a/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td +++ b/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td @@ -2042,6 +2042,11 @@ multiclass XVmulwusPat<SDPatternOperator OpNode, string Inst> { (!cast<LAInst>(Inst#"_Q_DU_D") LASX256:$xj, LASX256:$xk)>; } +// Widen a 128-bit LSX vector to a 256-bit LASX vector. The upper 128 bits are +// undefined. +def lsx_to_lasx : OutPatFrag<(ops node:$vj), + (INSERT_SUBREG (IMPLICIT_DEF), $vj, sub_128)>; + let Predicates = [HasExtLASX] in { // XVADD_{B/H/W/D} @@ -2851,15 +2856,13 @@ def : Pat<(fneg (v4f64 LASX256:$xj)), (XVBITREVI_D LASX256:$xj, 63)>; def : Pat<(v8f32 (sint_to_fp v8i32:$vj)), (XVFFINT_S_W v8i32:$vj)>; def : Pat<(v4f64 (sint_to_fp v4i64:$vj)), (XVFFINT_D_L v4i64:$vj)>; def : Pat<(v4f64 (sint_to_fp v4i32:$vj)), - (XVFFINT_D_L (VEXT2XV_D_W (SUBREG_TO_REG v4i32:$vj, - sub_128)))>; + (XVFFINT_D_L (VEXT2XV_D_W (lsx_to_lasx v4i32:$vj)))>; // XVFFINT_{S_WU/D_LU} def : Pat<(v8f32 (uint_to_fp v8i32:$vj)), (XVFFINT_S_WU v8i32:$vj)>; def : Pat<(v4f64 (uint_to_fp v4i64:$vj)), (XVFFINT_D_LU v4i64:$vj)>; def : Pat<(v4f64 (uint_to_fp v4i32:$vj)), - (XVFFINT_D_LU (VEXT2XV_DU_WU (SUBREG_TO_REG v4i32:$vj, - sub_128)))>; + (XVFFINT_D_LU (VEXT2XV_DU_WU (lsx_to_lasx v4i32:$vj)))>; // XVFFINT_S_L def : Pat<(v8f32 (loongarch_vffint_s_l (v4i64 LASX256:$xj), (v4i64 LASX256:$xk))), @@ -2872,7 +2875,7 @@ def : Pat<(v8i32 (fp_to_sint_sat v8f32:$vj, i32)), (XVFTINTRZ_W_S v8f32:$vj)>; def : Pat<(v4i64 (fp_to_sint_sat v4f64:$vj, i64)), (XVFTINTRZ_L_D v4f64:$vj)>; def : Pat<(v4i64 (fp_to_sint v4f32:$vj)), (XVFTINTRZL_L_S - (XVPERMI_D (SUBREG_TO_REG v4f32:$vj, sub_128), 216))>; + (XVPERMI_D (lsx_to_lasx v4f32:$vj), 216))>; def : Pat<(v4i32(fp_to_sint v4f64:$vj)), (EXTRACT_SUBREG(XVPICKEV_W(XVPERMI_D(XVFTINTRZ_L_D v4f64:$vj), 238), (XVFTINTRZ_L_D v4f64:$vj)), @@ -2895,7 +2898,7 @@ multiclass PatVrFToIntRoundLASX<SDNode ToInt, SDNode ToIntSat, SDNode Round, (!cast<LAInst>(InstWD) v4f64:$vj, v4f64:$vk)>; def : Pat<(v4i64 (ToInt (Round v4f32:$vj))), (!cast<LAInst>(InstLS) - (XVPERMI_D (SUBREG_TO_REG v4f32:$vj, sub_128), 216))>; + (XVPERMI_D (lsx_to_lasx v4f32:$vj), 216))>; } defm : PatVrFToIntRoundLASX<fp_to_sint, fp_to_sint_sat, ffloor, @@ -2919,7 +2922,7 @@ def : Pat<(v4i64 (fp_to_uint (ftrunc v4f64:$vj))), (XVFTINTRZ_LU_D v4f64:$vj)>; def : Pat<(v4i64 (fp_to_uint v4f32:$vj)), (XVFTINTRZ_LU_D (XVFCVTL_D_S - (XVPERMI_D (SUBREG_TO_REG v4f32:$vj, sub_128), 216)))>; + (XVPERMI_D (lsx_to_lasx v4f32:$vj), 216)))>; def : Pat<(v4i32(fp_to_uint v4f64:$vj)), (EXTRACT_SUBREG(XVPICKEV_W(XVPERMI_D(XVFTINTRZ_LU_D v4f64:$vj), 238), (XVFTINTRZ_LU_D v4f64:$vj)), @@ -3052,29 +3055,29 @@ def : Pat<(int_loongarch_lasx_insert_128_hi (v4i64 LASX256:$src), (v2i64 LSX128: // Sign extensions def : Pat<(v4i64 (sext v4i32:$vj)), - (v4i64 (VEXT2XV_D_W (SUBREG_TO_REG v4i32:$vj, sub_128)))>; + (v4i64 (VEXT2XV_D_W (lsx_to_lasx v4i32:$vj)))>; def : Pat<(v8i32 (sext v8i16:$vj)), - (v8i32 (VEXT2XV_W_H (SUBREG_TO_REG v8i16:$vj, sub_128)))>; + (v8i32 (VEXT2XV_W_H (lsx_to_lasx v8i16:$vj)))>; def : Pat<(v16i16 (sext v16i8:$vj)), - (v16i16 (VEXT2XV_H_B (SUBREG_TO_REG v16i8:$vj, sub_128)))>; + (v16i16 (VEXT2XV_H_B (lsx_to_lasx v16i8:$vj)))>; def : Pat<(v2i64 (sext_invec v16i8:$vj)), - (v2i64 (EXTRACT_SUBREG (VEXT2XV_D_B (SUBREG_TO_REG v16i8:$vj, sub_128)), + (v2i64 (EXTRACT_SUBREG (VEXT2XV_D_B (lsx_to_lasx v16i8:$vj)), sub_128))>; def : Pat<(v2i64 (sext_invec v8i16:$vj)), - (v2i64 (EXTRACT_SUBREG (VEXT2XV_D_H (SUBREG_TO_REG v8i16:$vj, sub_128)), + (v2i64 (EXTRACT_SUBREG (VEXT2XV_D_H (lsx_to_lasx v8i16:$vj)), sub_128))>; def : Pat<(v2i64 (sext_invec v4i32:$vj)), - (v2i64 (EXTRACT_SUBREG (VEXT2XV_D_W (SUBREG_TO_REG v4i32:$vj, sub_128)), + (v2i64 (EXTRACT_SUBREG (VEXT2XV_D_W (lsx_to_lasx v4i32:$vj)), sub_128))>; def : Pat<(v4i32 (sext_invec v16i8:$vj)), - (v4i32 (EXTRACT_SUBREG (VEXT2XV_W_B (SUBREG_TO_REG v16i8:$vj, sub_128)), + (v4i32 (EXTRACT_SUBREG (VEXT2XV_W_B (lsx_to_lasx v16i8:$vj)), sub_128))>; def : Pat<(v4i32 (sext_invec v8i16:$vj)), - (v4i32 (EXTRACT_SUBREG (VEXT2XV_W_H (SUBREG_TO_REG v8i16:$vj, sub_128)), + (v4i32 (EXTRACT_SUBREG (VEXT2XV_W_H (lsx_to_lasx v8i16:$vj)), sub_128))>; def : Pat<(v8i16 (sext_invec v16i8:$vj)), - (v8i16 (EXTRACT_SUBREG (VEXT2XV_H_B (SUBREG_TO_REG v16i8:$vj, sub_128)), + (v8i16 (EXTRACT_SUBREG (VEXT2XV_H_B (lsx_to_lasx v16i8:$vj)), sub_128))>; def : Pat<(v4i64 (sext_invec v32i8:$xj)), (v4i64 (VEXT2XV_D_B v32i8:$xj))>; @@ -3085,37 +3088,37 @@ def : Pat<(v8i32 (sext_invec v16i16:$xj)), (v8i32 (VEXT2XV_W_H v16i16:$xj))>; def : Pat<(v16i16 (sext_invec v32i8:$xj)), (v16i16 (VEXT2XV_H_B v32i8:$xj))>; def : Pat<(v4i64 (sext_invec v16i8:$vj)), - (v4i64 (VEXT2XV_D_B (SUBREG_TO_REG v16i8:$vj, sub_128)))>; + (v4i64 (VEXT2XV_D_B (lsx_to_lasx v16i8:$vj)))>; def : Pat<(v4i64 (sext_invec v8i16:$vj)), - (v4i64 (VEXT2XV_D_H (SUBREG_TO_REG v8i16:$vj, sub_128)))>; + (v4i64 (VEXT2XV_D_H (lsx_to_lasx v8i16:$vj)))>; def : Pat<(v8i32 (sext_invec v16i8:$vj)), - (v8i32 (VEXT2XV_W_B (SUBREG_TO_REG v16i8:$vj, sub_128)))>; + (v8i32 (VEXT2XV_W_B (lsx_to_lasx v16i8:$vj)))>; // Zero extensions def : Pat<(v4i64 (zext v4i32:$vj)), - (v4i64 (VEXT2XV_DU_WU (SUBREG_TO_REG v4i32:$vj, sub_128)))>; + (v4i64 (VEXT2XV_DU_WU (lsx_to_lasx v4i32:$vj)))>; def : Pat<(v8i32 (zext v8i16:$vj)), - (v8i32 (VEXT2XV_WU_HU (SUBREG_TO_REG v8i16:$vj, sub_128)))>; + (v8i32 (VEXT2XV_WU_HU (lsx_to_lasx v8i16:$vj)))>; def : Pat<(v16i16 (zext v16i8:$vj)), - (v16i16 (VEXT2XV_HU_BU (SUBREG_TO_REG v16i8:$vj, sub_128)))>; + (v16i16 (VEXT2XV_HU_BU (lsx_to_lasx v16i8:$vj)))>; def : Pat<(v2i64 (zext_invec v16i8:$vj)), - (v2i64 (EXTRACT_SUBREG (VEXT2XV_DU_BU (SUBREG_TO_REG v16i8:$vj, sub_128)), + (v2i64 (EXTRACT_SUBREG (VEXT2XV_DU_BU (lsx_to_lasx v16i8:$vj)), sub_128))>; def : Pat<(v2i64 (zext_invec v8i16:$vj)), - (v2i64 (EXTRACT_SUBREG (VEXT2XV_DU_HU (SUBREG_TO_REG v8i16:$vj, sub_128)), + (v2i64 (EXTRACT_SUBREG (VEXT2XV_DU_HU (lsx_to_lasx v8i16:$vj)), sub_128))>; def : Pat<(v2i64 (zext_invec v4i32:$vj)), - (v2i64 (EXTRACT_SUBREG (VEXT2XV_DU_WU (SUBREG_TO_REG v4i32:$vj, sub_128)), + (v2i64 (EXTRACT_SUBREG (VEXT2XV_DU_WU (lsx_to_lasx v4i32:$vj)), sub_128))>; def : Pat<(v4i32 (zext_invec v16i8:$vj)), - (v4i32 (EXTRACT_SUBREG (VEXT2XV_WU_BU (SUBREG_TO_REG v16i8:$vj, sub_128)), + (v4i32 (EXTRACT_SUBREG (VEXT2XV_WU_BU (lsx_to_lasx v16i8:$vj)), sub_128))>; def : Pat<(v4i32 (zext_invec v8i16:$vj)), - (v4i32 (EXTRACT_SUBREG (VEXT2XV_WU_HU (SUBREG_TO_REG v8i16:$vj, sub_128)), + (v4i32 (EXTRACT_SUBREG (VEXT2XV_WU_HU (lsx_to_lasx v8i16:$vj)), sub_128))>; def : Pat<(v8i16 (zext_invec v16i8:$vj)), - (v8i16 (EXTRACT_SUBREG (VEXT2XV_HU_BU (SUBREG_TO_REG v16i8:$vj, sub_128)), + (v8i16 (EXTRACT_SUBREG (VEXT2XV_HU_BU (lsx_to_lasx v16i8:$vj)), sub_128))>; def : Pat<(v4i64 (zext_invec v32i8:$xj)), (v4i64 (VEXT2XV_DU_BU v32i8:$xj))>; @@ -3126,11 +3129,11 @@ def : Pat<(v8i32 (zext_invec v16i16:$xj)), (v8i32 (VEXT2XV_WU_HU v16i16:$xj))>; def : Pat<(v16i16 (zext_invec v32i8:$xj)), (v16i16 (VEXT2XV_HU_BU v32i8:$xj))>; def : Pat<(v4i64 (zext_invec v16i8:$vj)), - (v4i64 (VEXT2XV_DU_BU (SUBREG_TO_REG v16i8:$vj, sub_128)))>; + (v4i64 (VEXT2XV_DU_BU (lsx_to_lasx v16i8:$vj)))>; def : Pat<(v4i64 (zext_invec v8i16:$vj)), - (v4i64 (VEXT2XV_DU_HU (SUBREG_TO_REG v8i16:$vj, sub_128)))>; + (v4i64 (VEXT2XV_DU_HU (lsx_to_lasx v8i16:$vj)))>; def : Pat<(v8i32 (zext_invec v16i8:$vj)), - (v8i32 (VEXT2XV_WU_BU (SUBREG_TO_REG v16i8:$vj, sub_128)))>; + (v8i32 (VEXT2XV_WU_BU (lsx_to_lasx v16i8:$vj)))>; } // Predicates = [HasExtLASX] `````````` </details> https://github.com/llvm/llvm-project/pull/242331 _______________________________________________ llvm-branch-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
