R600: Add support for 24-bit MAD instructions

Reviewed-by: Vincent Lejeune <vljn at ovi.com>
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@186923 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/R600/R600Instructions.td b/lib/Target/R600/R600Instructions.td
index 56015ea80a4ee5a3641fb58e65a441693af232ab..3511dcd1ead36af6b2a326ca2c14690237f2c2c4 100644 (file)
--- a/lib/Target/R600/R600Instructions.td
+++ b/lib/Target/R600/R600Instructions.td
@@ -1460,6 +1460,9 @@ let Predicates = [isEGorCayman] in {
   def BFI_INT_eg : R600_3OP <0x06, "BFI_INT", [], VecALU>;
   defm : BFIPatterns <BFI_INT_eg>;
 
+  def MULADD_UINT24_eg : R600_3OP <0x10, "MULADD_UINT24",
+    [(set i32:$dst, (add (mul U24:$src0, U24:$src1), i32:$src2))], VecALU
+  >;
   def BIT_ALIGN_INT_eg : R600_3OP <0xC, "BIT_ALIGN_INT", [], VecALU>;
   def : ROTRPattern <BIT_ALIGN_INT_eg>;
 
@@ -1706,6 +1709,9 @@ defm R600_ : RegisterLoadStore <R600_Reg32, FRAMEri, ADDRIndirect>;
 
 let Predicates = [isCayman] in {
 
+def MULADD_INT24_cm : R600_3OP <0x08, "MULADD_INT24",
+  [(set i32:$dst, (add (mul I24:$src0, I24:$src1), i32:$src2))], VecALU
+>;
 def MUL_INT24_cm : R600_2OP <0x5B, "MUL_INT24",
   [(set i32:$dst, (mul I24:$src0, I24:$src1))], VecALU
 >;

diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
index 8f3baaab615c685f527afc5457a16fe497206d85..7ae0ffdd7a6238b24da56af508cf60693be9e9c8 100644 (file)
--- a/lib/Target/R600/SIInstructions.td
+++ b/lib/Target/R600/SIInstructions.td
@@ -983,8 +983,12 @@ let neverHasSideEffects = 1 in {
 
 def V_MAD_LEGACY_F32 : VOP3_32 <0x00000140, "V_MAD_LEGACY_F32", []>;
 def V_MAD_F32 : VOP3_32 <0x00000141, "V_MAD_F32", []>;
-//def V_MAD_I32_I24 : VOP3_32 <0x00000142, "V_MAD_I32_I24", []>;
-//def V_MAD_U32_U24 : VOP3_32 <0x00000143, "V_MAD_U32_U24", []>;
+def V_MAD_I32_I24 : VOP3_32 <0x00000142, "V_MAD_I32_I24",
+  [(set i32:$dst, (add (mul I24:$src0, I24:$src1), i32:$src2))]
+>;
+def V_MAD_U32_U24 : VOP3_32 <0x00000143, "V_MAD_U32_U24",
+  [(set i32:$dst, (add (mul U24:$src0, U24:$src1), i32:$src2))]
+>;
 
 } // End neverHasSideEffects
 def V_CUBEID_F32 : VOP3_32 <0x00000144, "V_CUBEID_F32", []>;

diff --git a/test/CodeGen/R600/mad_int24.ll b/test/CodeGen/R600/mad_int24.ll
new file mode 100644 (file)
index 0000000..ce42ae7
--- /dev/null
+++ b/test/CodeGen/R600/mad_int24.ll
@@ -0,0 +1,20 @@
+; RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s --check-prefix=EG-CHECK
+; RUN: llc < %s -march=r600 -mcpu=cayman | FileCheck %s --check-prefix=CM-CHECK
+; RUN: llc < %s -march=r600 -mcpu=SI | FileCheck %s --check-prefix=SI-CHECK
+
+; EG-CHECK: @i32_mad24
+; Signed 24-bit multiply is not supported on pre-Cayman GPUs.
+; EG-CHECK: MULLO_INT
+; CM-CHECK: MULADD_INT24 {{[ *]*}}T{{[0-9].[XYZW]}}, KC0[2].Z, KC0[2].W, KC0[3].X
+; SI-CHECK: V_MAD_I32_I24
+define void @i32_mad24(i32 addrspace(1)* %out, i32 %a, i32 %b, i32 %c) {
+entry:
+  %0 = shl i32 %a, 8
+  %a_24 = ashr i32 %0, 8
+  %1 = shl i32 %b, 8
+  %b_24 = ashr i32 %1, 8
+  %2 = mul i32 %a_24, %b_24
+  %3 = add i32 %2, %c
+  store i32 %3, i32 addrspace(1)* %out
+  ret void
+}

diff --git a/test/CodeGen/R600/mad_uint24.ll b/test/CodeGen/R600/mad_uint24.ll
new file mode 100644 (file)
index 0000000..00aa64a
--- /dev/null
+++ b/test/CodeGen/R600/mad_uint24.ll
@@ -0,0 +1,70 @@
+; RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s --check-prefix=EG-CHECK
+; RUN: llc < %s -march=r600 -mcpu=cayman | FileCheck %s --check-prefix=EG-CHECK
+; RUN: llc < %s -march=r600 -mcpu=SI | FileCheck %s --check-prefix=SI-CHECK
+
+; EG-CHECK: @u32_mad24
+; EG-CHECK: MULADD_UINT24 {{[* ]*}}T{{[0-9]\.[XYZW]}}, KC0[2].Z, KC0[2].W, KC0[3].X
+; SI-CHECK: @u32_mad24
+; SI-CHECK: V_MAD_U32_U24
+
+define void @u32_mad24(i32 addrspace(1)* %out, i32 %a, i32 %b, i32 %c) {
+entry:
+  %0 = shl i32 %a, 8
+  %a_24 = lshr i32 %0, 8
+  %1 = shl i32 %b, 8
+  %b_24 = lshr i32 %1, 8
+  %2 = mul i32 %a_24, %b_24
+  %3 = add i32 %2, %c
+  store i32 %3, i32 addrspace(1)* %out
+  ret void
+}
+
+; EG-CHECK: @i16_mad24
+; EG-CHECK-DAG: VTX_READ_16 [[A:T[0-9]\.X]], T{{[0-9]}}.X, 40
+; EG-CHECK-DAG: VTX_READ_16 [[B:T[0-9]\.X]], T{{[0-9]}}.X, 44
+; EG-CHECK-DAG: VTX_READ_16 [[C:T[0-9]\.X]], T{{[0-9]}}.X, 48
+; The order of A and B does not matter.
+; EG-CHECK: MULADD_UINT24 {{[* ]*}}T{{[0-9]}}.[[MAD_CHAN:[XYZW]]], [[A]], [[B]], [[C]]
+; The result must be sign-extended
+; EG-CHECK: LSHL {{[* ]*}}T{{[0-9]}}.[[LSHL_CHAN:[XYZW]]], PV.[[MAD_CHAN]], literal.x
+; EG-CHECK: 16
+; EG-CHECK: ASHR {{[* ]*}}T{{[0-9]\.[XYZW]}}, PV.[[LSHL_CHAN]], literal.x
+; EG-CHECK: 16
+; SI-CHECK: @i16_mad24
+; SI-CHECK: V_MAD_U32_U24 [[MAD:VGPR[0-9]]], {{[SV]GPR[0-9], [SV]GPR[0-9]}}
+; SI-CHECK: V_LSHLREV_B32_e32 [[LSHL:VGPR[0-9]]], 16, [[MAD]]
+; SI-CHECK: V_ASHRREV_I32_e32 VGPR{{[0-9]}}, 16, [[LSHL]]
+
+define void @i16_mad24(i32 addrspace(1)* %out, i16 %a, i16 %b, i16 %c) {
+entry:
+  %0 = mul i16 %a, %b
+  %1 = add i16 %0, %c
+  %2 = sext i16 %1 to i32
+  store i32 %2, i32 addrspace(1)* %out
+  ret void
+}
+
+; EG-CHECK: @i8_mad24
+; EG-CHECK-DAG: VTX_READ_8 [[A:T[0-9]\.X]], T{{[0-9]}}.X, 40
+; EG-CHECK-DAG: VTX_READ_8 [[B:T[0-9]\.X]], T{{[0-9]}}.X, 44
+; EG-CHECK-DAG: VTX_READ_8 [[C:T[0-9]\.X]], T{{[0-9]}}.X, 48
+; The order of A and B does not matter.
+; EG-CHECK: MULADD_UINT24 {{[* ]*}}T{{[0-9]}}.[[MAD_CHAN:[XYZW]]], [[A]], [[B]], [[C]]
+; The result must be sign-extended
+; EG-CHECK: LSHL {{[* ]*}}T{{[0-9]}}.[[LSHL_CHAN:[XYZW]]], PV.[[MAD_CHAN]], literal.x
+; EG-CHECK: 24
+; EG-CHECK: ASHR {{[* ]*}}T{{[0-9]\.[XYZW]}}, PV.[[LSHL_CHAN]], literal.x
+; EG-CHECK: 24
+; SI-CHECK: @i8_mad24
+; SI-CHECK: V_MAD_U32_U24 [[MUL:VGPR[0-9]]], {{[SV]GPR[0-9], [SV]GPR[0-9]}}
+; SI-CHECK: V_LSHLREV_B32_e32 [[LSHL:VGPR[0-9]]], 24, [[MUL]]
+; SI-CHECK: V_ASHRREV_I32_e32 VGPR{{[0-9]}}, 24, [[LSHL]]
+
+define void @i8_mad24(i32 addrspace(1)* %out, i8 %a, i8 %b, i8 %c) {
+entry:
+  %0 = mul i8 %a, %b
+  %1 = add i8 %0, %c
+  %2 = sext i8 %1 to i32
+  store i32 %2, i32 addrspace(1)* %out
+  ret void
+}