GlobalISel: Fold out G_FPTRUNC(G_FPEXT)

[arsenm]

62eac3e06891937a648014ccd49d55c0072bb92e

llvm/include/llvm/Target/GlobalISel/Combine.td

def sub_add_reg: GICombineRule <
   (apply [{ Helper.applyBuildFn(*${root}, ${matchinfo}); }])>;
 
 def bitcast_bitcast_fold : GICombineRule<
   (defs root:$dst),
   (match (G_BITCAST $dst, $src1):$op, (G_BITCAST $src1, $src0),
       [{ return MRI.getType(${src0}.getReg()) == MRI.getType(${dst}.getReg()); }]),
   (apply [{ Helper.replaceSingleDefInstWithReg(*${op}, ${src0}.getReg()); }])>;
 
+
+def fptrunc_fpext_fold : GICombineRule<
+  (defs root:$dst),
+  (match (G_FPTRUNC $dst, $src1):$op, (G_FPEXT $src1, $src0),
+      [{ return MRI.getType(${src0}.getReg()) == MRI.getType(${dst}.getReg()); }]),
+  (apply [{ Helper.replaceSingleDefInstWithReg(*${op}, ${src0}.getReg()); }])>;
+
+
 def select_to_minmax: GICombineRule<
   (defs root:$root, build_fn_matchinfo:$info),
   (match (wip_match_opcode G_SELECT):$root,
          [{ return Helper.matchSimplifySelectToMinMax(*${root}, ${info}); }]),
   (apply [{ Helper.applyBuildFn(*${root}, ${info}); }])>;
 
 // FIXME: These should use the custom predicate feature once it lands.
 def undef_combines : GICombineGroup<[undef_to_fp_zero, undef_to_int_zero,

 def identity_combines : GICombineGroup<[select_same_val, right_identity_zero,
                                         binop_same_val, binop_left_to_zero,
                                         binop_right_to_zero, p2i_to_i2p,
                                         i2p_to_p2i, anyext_trunc_fold,
                                         fneg_fneg_fold, right_identity_one,
                                         add_sub_reg, buildvector_identity_fold,
                                         trunc_buildvector_fold,
                                         trunc_lshr_buildvector_fold,
-                                        bitcast_bitcast_fold]>;
+                                        bitcast_bitcast_fold, fptrunc_fpext_fold]>;
 
 def const_combines : GICombineGroup<[constant_fp_op, const_ptradd_to_i2p,
                                      overlapping_and, mulo_by_2, mulo_by_0,
                                      addo_by_0, adde_to_addo,
                                      combine_minmax_nan]>;
 
 def known_bits_simplifications : GICombineGroup<[
   redundant_and, redundant_sext_inreg, redundant_or, urem_pow2_to_mask,

llvm/test/CodeGen/AMDGPU/GlobalISel/prelegalizer-combiner-fptrunc_fpext.mir

+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -run-pass=amdgpu-prelegalizer-combiner -o - %s | FileCheck %s
+
+---
+name: fptrunc_fpext_s16_to_s32_to_s16
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $vgpr0
+    ; CHECK-LABEL: name: fptrunc_fpext_s16_to_s32_to_s16
+    ; CHECK: liveins: $vgpr0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: %reg:_(s32) = COPY $vgpr0
+    ; CHECK-NEXT: $vgpr0 = COPY %reg(s32)
+    %reg:_(s32) = COPY $vgpr0
+    %src:_(s16) = G_TRUNC %reg
+    %fpext:_(s32) = G_FPEXT %src
+    %fptrunc:_(s16) = G_FPTRUNC %fpext
+    %anyext:_(s32) = G_ANYEXT %fptrunc
+    $vgpr0 = COPY %anyext
+...
+
+---
+name: fptrunc_fpext_s16_to_s64_to_s32
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $vgpr0
+    ; CHECK-LABEL: name: fptrunc_fpext_s16_to_s64_to_s32
+    ; CHECK: liveins: $vgpr0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: %reg:_(s32) = COPY $vgpr0
+    ; CHECK-NEXT: %src:_(s16) = G_TRUNC %reg(s32)
+    ; CHECK-NEXT: %fpext:_(s64) = G_FPEXT %src(s16)
+    ; CHECK-NEXT: %fptrunc:_(s32) = G_FPTRUNC %fpext(s64)
+    ; CHECK-NEXT: $vgpr0 = COPY %fptrunc(s32)
+    %reg:_(s32) = COPY $vgpr0
+    %src:_(s16) = G_TRUNC %reg
+    %fpext:_(s64) = G_FPEXT %src
+    %fptrunc:_(s32) = G_FPTRUNC %fpext
+    $vgpr0 = COPY %fptrunc
+...
+
+---
+name: fptrunc_fpext_v2s16_to_v2s32_to_v2s16
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $vgpr0
+    ; CHECK-LABEL: name: fptrunc_fpext_v2s16_to_v2s32_to_v2s16
+    ; CHECK: liveins: $vgpr0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: %src:_(<2 x s16>) = COPY $vgpr0
+    ; CHECK-NEXT: $vgpr0 = COPY %src(<2 x s16>)
+    %src:_(<2 x s16>) = COPY $vgpr0
+    %fpext:_(<2 x s32>) = G_FPEXT %src
+    %fptrunc:_(<2 x s16>) = G_FPTRUNC %fpext
+    $vgpr0 = COPY %fptrunc
+...
+
+---
+name: fptrunc_fpext_v2s16_to_v2s64_to_v2s32
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $vgpr0
+    ; CHECK-LABEL: name: fptrunc_fpext_v2s16_to_v2s64_to_v2s32
+    ; CHECK: liveins: $vgpr0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: %src:_(<2 x s16>) = COPY $vgpr0
+    ; CHECK-NEXT: %fpext:_(<2 x s64>) = G_FPEXT %src(<2 x s16>)
+    ; CHECK-NEXT: %fptrunc:_(<2 x s32>) = G_FPTRUNC %fpext(<2 x s64>)
+    ; CHECK-NEXT: $vgpr0_vgpr1 = COPY %fptrunc(<2 x s32>)
+    %src:_(<2 x s16>) = COPY $vgpr0
+    %fpext:_(<2 x s64>) = G_FPEXT %src
+    %fptrunc:_(<2 x s32>) = G_FPTRUNC %fpext
+    $vgpr0_vgpr1 = COPY %fptrunc
+...