GlobalISel: Fold out G_FPTRUNC(G_FPEXT)

diff --git a/llvm/include/llvm/Target/GlobalISel/Combine.td b/llvm/include/llvm/Target/GlobalISel/Combine.td
index 0c70483..40ccfb0 100644 (file)
--- a/llvm/include/llvm/Target/GlobalISel/Combine.td
+++ b/llvm/include/llvm/Target/GlobalISel/Combine.td
@@ -1036,6 +1036,14 @@ def bitcast_bitcast_fold : GICombineRule<
       [{ 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,
@@ -1063,7 +1071,7 @@ def identity_combines : GICombineGroup<[select_same_val, right_identity_zero,
                                         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,

diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/prelegalizer-combiner-fptrunc_fpext.mir b/llvm/test/CodeGen/AMDGPU/GlobalISel/prelegalizer-combiner-fptrunc_fpext.mir
new file mode 100644 (file)
index 0000000..b6434c6
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/prelegalizer-combiner-fptrunc_fpext.mir
@@ -0,0 +1,78 @@
+# 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
+...