From 945fab34477773c75ca6414eb92f5989091716aa Mon Sep 17 00:00:00 2001
From: Marek Olsak <marek.olsak@amd.com>
Date: Tue, 24 Mar 2015 13:40:21 +0000
Subject: [PATCH] R600/SI: Improve BFM support

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@233077 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/R600/SIInstrInfo.cpp      |  1 +
 lib/Target/R600/SIInstructions.td    | 22 +++++++++++++++++++---
 test/CodeGen/R600/llvm.AMDGPU.bfm.ll | 27 +++++++++++++++++++++++----
 3 files changed, 43 insertions(+), 7 deletions(-)

diff --git a/lib/Target/R600/SIInstrInfo.cpp b/lib/Target/R600/SIInstrInfo.cpp
index c5eda9a387b..ba98ad7dd70 100644
--- a/lib/Target/R600/SIInstrInfo.cpp
+++ b/lib/Target/R600/SIInstrInfo.cpp
@@ -1406,6 +1406,7 @@ unsigned SIInstrInfo::getVALUOp(const MachineInstr &MI) {
   case AMDGPU::S_SEXT_I32_I16: return AMDGPU::V_BFE_I32;
   case AMDGPU::S_BFE_U32: return AMDGPU::V_BFE_U32;
   case AMDGPU::S_BFE_I32: return AMDGPU::V_BFE_I32;
+  case AMDGPU::S_BFM_B32: return AMDGPU::V_BFM_B32_e64;
   case AMDGPU::S_BREV_B32: return AMDGPU::V_BFREV_B32_e32;
   case AMDGPU::S_NOT_B32: return AMDGPU::V_NOT_B32_e32;
   case AMDGPU::S_NOT_B64: return AMDGPU::V_NOT_B32_e32;
diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
index b27d1424ff2..764e79b0fb3 100644
--- a/lib/Target/R600/SIInstructions.td
+++ b/lib/Target/R600/SIInstructions.td
@@ -308,7 +308,8 @@ defm S_ASHR_I64 : SOP2_64_32 <sop2<0x23, 0x21>, "s_ashr_i64",
 >;
 } // End Defs = [SCC]
 
-defm S_BFM_B32 : SOP2_32 <sop2<0x24, 0x22>, "s_bfm_b32", []>;
+defm S_BFM_B32 : SOP2_32 <sop2<0x24, 0x22>, "s_bfm_b32",
+  [(set i32:$dst, (AMDGPUbfm i32:$src0, i32:$src1))]>;
 defm S_BFM_B64 : SOP2_64 <sop2<0x25, 0x23>, "s_bfm_b64", []>;
 defm S_MUL_I32 : SOP2_32 <sop2<0x26, 0x24>, "s_mul_i32",
   [(set i32:$dst, (mul i32:$src0, i32:$src1))]
@@ -1613,8 +1614,8 @@ defm V_MAC_LEGACY_F32 : VOP2_VI3_Inst <vop23<0x6, 0x28e>, "v_mac_legacy_f32",
 >;
 } // End isCommutable = 1
 
-defm V_BFM_B32 : VOP2_VI3_Inst <vop23<0x1e, 0x293>, "v_bfm_b32", VOP_I32_I32_I32,
-  AMDGPUbfm
+defm V_BFM_B32 : VOP2_VI3_Inst <vop23<0x1e, 0x293>, "v_bfm_b32",
+  VOP_I32_I32_I32
 >;
 defm V_BCNT_U32_B32 : VOP2_VI3_Inst <vop23<0x22, 0x28b>, "v_bcnt_u32_b32",
   VOP_I32_I32_I32
@@ -3323,6 +3324,21 @@ def : Pat <
   (V_CNDMASK_B32_e64 $src0, $src1, $src2)
 >;
 
+multiclass BFMPatterns <ValueType vt, InstSI BFM, InstSI MOV> {
+  def : Pat <
+    (vt (shl (vt (add (vt (shl 1, vt:$a)), -1)), vt:$b)),
+    (BFM $a, $b)
+  >;
+
+  def : Pat <
+    (vt (add (vt (shl 1, vt:$a)), -1)),
+    (BFM $a, (MOV 0))
+  >;
+}
+
+defm : BFMPatterns <i32, S_BFM_B32, S_MOV_B32>;
+// FIXME: defm : BFMPatterns <i64, S_BFM_B64, S_MOV_B64>;
+
 //===----------------------------------------------------------------------===//
 // Fract Patterns
 //===----------------------------------------------------------------------===//
diff --git a/test/CodeGen/R600/llvm.AMDGPU.bfm.ll b/test/CodeGen/R600/llvm.AMDGPU.bfm.ll
index 2346f408ec4..50492289d74 100644
--- a/test/CodeGen/R600/llvm.AMDGPU.bfm.ll
+++ b/test/CodeGen/R600/llvm.AMDGPU.bfm.ll
@@ -5,7 +5,7 @@
 declare i32 @llvm.AMDGPU.bfm(i32, i32) nounwind readnone
 
 ; FUNC-LABEL: {{^}}bfm_arg_arg:
-; SI: v_bfm
+; SI: s_bfm_b32 {{s[0-9]+}}, {{s[0-9]+}}, {{s[0-9]+}}
 ; EG: BFM_INT
 define void @bfm_arg_arg(i32 addrspace(1)* %out, i32 %src0, i32 %src1) nounwind {
   %bfm = call i32 @llvm.AMDGPU.bfm(i32 %src0, i32 %src1) nounwind readnone
@@ -14,7 +14,7 @@ define void @bfm_arg_arg(i32 addrspace(1)* %out, i32 %src0, i32 %src1) nounwind
 }
 
 ; FUNC-LABEL: {{^}}bfm_arg_imm:
-; SI: v_bfm
+; SI: s_bfm_b32 {{s[0-9]+}}, {{s[0-9]+}}, 0x7b
 ; EG: BFM_INT
 define void @bfm_arg_imm(i32 addrspace(1)* %out, i32 %src0) nounwind {
   %bfm = call i32 @llvm.AMDGPU.bfm(i32 %src0, i32 123) nounwind readnone
@@ -23,7 +23,7 @@ define void @bfm_arg_imm(i32 addrspace(1)* %out, i32 %src0) nounwind {
 }
 
 ; FUNC-LABEL: {{^}}bfm_imm_arg:
-; SI: v_bfm
+; SI: s_bfm_b32 {{s[0-9]+}}, 0x7b, {{s[0-9]+}}
 ; EG: BFM_INT
 define void @bfm_imm_arg(i32 addrspace(1)* %out, i32 %src1) nounwind {
   %bfm = call i32 @llvm.AMDGPU.bfm(i32 123, i32 %src1) nounwind readnone
@@ -32,10 +32,29 @@ define void @bfm_imm_arg(i32 addrspace(1)* %out, i32 %src1) nounwind {
 }
 
 ; FUNC-LABEL: {{^}}bfm_imm_imm:
-; SI: v_bfm
+; SI: s_bfm_b32 {{s[0-9]+}}, 0x7b, 0x1c8
 ; EG: BFM_INT
 define void @bfm_imm_imm(i32 addrspace(1)* %out) nounwind {
   %bfm = call i32 @llvm.AMDGPU.bfm(i32 123, i32 456) nounwind readnone
   store i32 %bfm, i32 addrspace(1)* %out, align 4
   ret void
 }
+
+; FUNC-LABEL: {{^}}bfm_pattern:
+; SI: s_bfm_b32 {{s[0-9]+}}, {{s[0-9]+}}, {{s[0-9]+}}
+define void @bfm_pattern(i32 addrspace(1)* %out, i32 %x, i32 %y) {
+  %a = shl i32 1, %x
+  %b = sub i32 %a, 1
+  %c = shl i32 %b, %y
+  store i32 %c, i32 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: {{^}}bfm_pattern_simple:
+; SI: s_bfm_b32 {{s[0-9]+}}, {{s[0-9]+}}, 0
+define void @bfm_pattern_simple(i32 addrspace(1)* %out, i32 %x) {
+  %a = shl i32 1, %x
+  %b = sub i32 %a, 1
+  store i32 %b, i32 addrspace(1)* %out
+  ret void
+}
-- 
2.34.1