[llvm] r320562 - [mips] Provide additional DSP bitconvert patterns

[Stefan Maksimovic via llvm-commits]

Author: smaksimovic
Date: Wed Dec 13 02:13:35 2017
New Revision: 320562

URL: http://llvm.org/viewvc/llvm-project?rev=320562&view=rev
Log:
[mips] Provide additional DSP bitconvert patterns

Previously, v2i16 -> f32 bitcast could not be matched.
Add patterns to support matching this and similar types of bitcasts.

Differential revision: https://reviews.llvm.org/D40959

Added:
    llvm/trunk/test/CodeGen/Mips/v2i16tof32.ll
Modified:
    llvm/trunk/lib/Target/Mips/MipsDSPInstrInfo.td

Modified: llvm/trunk/lib/Target/Mips/MipsDSPInstrInfo.td
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/Mips/MipsDSPInstrInfo.td?rev=320562&r1=320561&r2=320562&view=diff
==============================================================================
--- llvm/trunk/lib/Target/Mips/MipsDSPInstrInfo.td (original)
+++ llvm/trunk/lib/Target/Mips/MipsDSPInstrInfo.td Wed Dec 13 02:13:35 2017
@@ -1325,6 +1325,10 @@ def : BitconvertPat<i32, v2i16, GPR32, D
 def : BitconvertPat<i32, v4i8, GPR32, DSPR>;
 def : BitconvertPat<v2i16, i32, DSPR, GPR32>;
 def : BitconvertPat<v4i8, i32, DSPR, GPR32>;
+def : BitconvertPat<f32, v2i16, FGR32, DSPR>;
+def : BitconvertPat<f32, v4i8, FGR32, DSPR>;
+def : BitconvertPat<v2i16, f32, DSPR, FGR32>;
+def : BitconvertPat<v4i8, f32, DSPR, FGR32>;
 
 def : DSPPat<(v2i16 (load addr:$a)),
              (v2i16 (COPY_TO_REGCLASS (LW addr:$a), DSPR))>;

Added: llvm/trunk/test/CodeGen/Mips/v2i16tof32.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/Mips/v2i16tof32.ll?rev=320562&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/Mips/v2i16tof32.ll (added)
+++ llvm/trunk/test/CodeGen/Mips/v2i16tof32.ll Wed Dec 13 02:13:35 2017
@@ -0,0 +1,47 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -O2 -mtriple mipsel--linux-android -mattr=+dsp -verify-machineinstrs | FileCheck %s
+
+; Function below generates a v2i16 to f32 bitcast.
+; Test that we are able to match it.
+
+define float @f(<8 x i16>* %a) {
+; CHECK-LABEL: f:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    addiu $sp, $sp, -32
+; CHECK-NEXT:    .cfi_def_cfa_offset 32
+; CHECK-NEXT:    sw $fp, 28($sp) # 4-byte Folded Spill
+; CHECK-NEXT:    .cfi_offset 30, -4
+; CHECK-NEXT:    move $fp, $sp
+; CHECK-NEXT:    .cfi_def_cfa_register 30
+; CHECK-NEXT:    addiu $1, $zero, -16
+; CHECK-NEXT:    and $sp, $sp, $1
+; CHECK-NEXT:    lw $1, 8($4)
+; CHECK-NEXT:    lw $2, 4($4)
+; CHECK-NEXT:    lw $3, 12($4)
+; CHECK-NEXT:    sw $3, 12($sp)
+; CHECK-NEXT:    sw $1, 8($sp)
+; CHECK-NEXT:    sw $2, 4($sp)
+; CHECK-NEXT:    lw $1, 0($4)
+; CHECK-NEXT:    sw $1, 0($sp)
+; CHECK-NEXT:    mtc1 $1, $f0
+; CHECK-NEXT:    move $sp, $fp
+; CHECK-NEXT:    lw $fp, 28($sp) # 4-byte Folded Reload
+; CHECK-NEXT:    jr $ra
+; CHECK-NEXT:    addiu $sp, $sp, 32
+; CHECK-NEXT:    .set at
+; CHECK-NEXT:    .set macro
+; CHECK-NEXT:    .set reorder
+; CHECK-NEXT:    .end f
+entry:
+  %m = alloca <8 x i16>
+  %0 = load <8 x i16>, <8 x i16>* %a
+  store <8 x i16> %0, <8 x i16>* %m
+  %1 = bitcast <8 x i16> %0 to <4 x float>
+  %2 = shufflevector <4 x float> %1, <4 x float> undef, <8 x i32> <i32 0, i32 3, i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef>
+  %3 = shufflevector <8 x float> zeroinitializer, <8 x float> %2, <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 4, i32 5, i32 6, i32 7>
+  %4 = bitcast <8 x float> %3 to <8 x i32>
+  %5 = extractelement <8 x i32> %4, i32 0
+  %6 = bitcast i32 %5 to float
+  ret float %6
+}
+