[Mesa-dev] [PATCH v2 2/2] R600/SI: Add intrinsic for BUFFER_LOAD_DWORD* instructions
Michel Dänzer
michel at daenzer.net
Fri Jan 24 00:07:34 PST 2014
From: Michel Dänzer <michel.daenzer at amd.com>
Signed-off-by: Michel Dänzer <michel.daenzer at amd.com>
---
v2: I've confirmed in the meantime that the instruction offset is ignored
when the offen bit is enabled.
lib/Target/R600/SIInstrInfo.td | 60 ++++++++++++++++++++++-----------
lib/Target/R600/SIInstructions.td | 48 ++++++++++++++++++++++++--
lib/Target/R600/SIIntrinsics.td | 14 ++++++++
test/CodeGen/R600/llvm.SI.load.dword.ll | 40 ++++++++++++++++++++++
4 files changed, 141 insertions(+), 21 deletions(-)
create mode 100644 test/CodeGen/R600/llvm.SI.load.dword.ll
diff --git a/lib/Target/R600/SIInstrInfo.td b/lib/Target/R600/SIInstrInfo.td
index d0cc7ce..97d37d8 100644
--- a/lib/Target/R600/SIInstrInfo.td
+++ b/lib/Target/R600/SIInstrInfo.td
@@ -425,26 +425,48 @@ class MTBUF_Store_Helper <bits<3> op, string asm, RegisterClass regClass> : MTBU
multiclass MUBUF_Load_Helper <bits<7> op, string asm, RegisterClass regClass> {
- let glc = 0, lds = 0, slc = 0, tfe = 0, soffset = 128 /* ZERO */,
- mayLoad = 1 in {
-
- let offen = 1, idxen = 0, addr64 = 0, offset = 0 in {
- def _OFFEN : MUBUF <op, (outs regClass:$vdata),
- (ins SReg_128:$srsrc, VReg_32:$vaddr),
- asm#" $vdata, $srsrc + $vaddr", []>;
- }
-
- let offen = 0, idxen = 1, addr64 = 0 in {
- def _IDXEN : MUBUF <op, (outs regClass:$vdata),
- (ins SReg_128:$srsrc, VReg_32:$vaddr, i16imm:$offset),
- asm#" $vdata, $srsrc[$vaddr] + $offset", []>;
- }
+ let lds = 0, mayLoad = 1 in {
+
+ let addr64 = 0 in {
+
+ let offen = 0, idxen = 0 in {
+ def _OFFSET : MUBUF <op, (outs regClass:$vdata),
+ (ins SReg_128:$srsrc, VReg_32:$vaddr,
+ i16imm:$offset, SSrc_32:$soffset, i1imm:$glc,
+ i1imm:$slc, i1imm:$tfe),
+ asm#" $vdata, $srsrc + $offset + $soffset, glc=$glc, slc=$slc, tfe=$tfe", []>;
+ }
+
+ let offen = 1, idxen = 0, offset = 0 in {
+ def _OFFEN : MUBUF <op, (outs regClass:$vdata),
+ (ins SReg_128:$srsrc, VReg_32:$vaddr,
+ SSrc_32:$soffset, i1imm:$glc, i1imm:$slc,
+ i1imm:$tfe),
+ asm#" $vdata, $srsrc + $vaddr + $soffset, glc=$glc, slc=$slc, tfe=$tfe", []>;
+ }
+
+ let offen = 0, idxen = 1 in {
+ def _IDXEN : MUBUF <op, (outs regClass:$vdata),
+ (ins SReg_128:$srsrc, VReg_32:$vaddr,
+ i16imm:$offset, SSrc_32:$soffset, i1imm:$glc,
+ i1imm:$slc, i1imm:$tfe),
+ asm#" $vdata, $srsrc[$vaddr] + $offset + $soffset, glc=$glc, slc=$slc, tfe=$tfe", []>;
+ }
+
+ let offen = 1, idxen = 1 in {
+ def _BOTHEN : MUBUF <op, (outs regClass:$vdata),
+ (ins SReg_128:$srsrc, VReg_64:$vaddr,
+ SSrc_32:$soffset, i1imm:$glc,
+ i1imm:$slc, i1imm:$tfe),
+ asm#" $vdata, $srsrc[$vaddr[0]] + $vaddr[1] + $soffset, glc=$glc, slc=$slc, tfe=$tfe", []>;
+ }
+ }
- let offen = 0, idxen = 0, addr64 = 1 in {
- def _ADDR64 : MUBUF <op, (outs regClass:$vdata),
- (ins SReg_128:$srsrc, VReg_64:$vaddr, i16imm:$offset),
- asm#" $vdata, $srsrc + $vaddr + $offset", []>;
- }
+ let offen = 0, idxen = 0, addr64 = 1, glc = 0, slc = 0, tfe = 0, soffset = 128 /* ZERO */ in {
+ def _ADDR64 : MUBUF <op, (outs regClass:$vdata),
+ (ins SReg_128:$srsrc, VReg_64:$vaddr, i16imm:$offset),
+ asm#" $vdata, $srsrc + $vaddr + $offset", []>;
+ }
}
}
diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
index c0ad398..a38121b 100644
--- a/lib/Target/R600/SIInstructions.td
+++ b/lib/Target/R600/SIInstructions.td
@@ -1428,7 +1428,7 @@ def : Pat <
/* int_SI_vs_load_input */
def : Pat<
(SIload_input i128:$tlst, IMM12bit:$attr_offset, i32:$buf_idx_vgpr),
- (BUFFER_LOAD_FORMAT_XYZW_IDXEN $tlst, $buf_idx_vgpr, imm:$attr_offset)
+ (BUFFER_LOAD_FORMAT_XYZW_IDXEN $tlst, $buf_idx_vgpr, imm:$attr_offset, 0, 0, 0, 0)
>;
/* int_SI_export */
@@ -1834,7 +1834,7 @@ def : Pat <
// 3. Offset in an 32Bit VGPR
def : Pat <
(SIload_constant i128:$sbase, i32:$voff),
- (BUFFER_LOAD_DWORD_OFFEN $sbase, $voff)
+ (BUFFER_LOAD_DWORD_OFFEN $sbase, $voff, 0, 0, 0, 0)
>;
// The multiplication scales from [0,1] to the unsigned integer range
@@ -1995,6 +1995,50 @@ defm : MUBUFStore_Pattern <BUFFER_STORE_DWORDX2, i64, global_store>;
defm : MUBUFStore_Pattern <BUFFER_STORE_DWORDX2, v2i32, global_store>;
defm : MUBUFStore_Pattern <BUFFER_STORE_DWORDX4, v4i32, global_store>;
+// BUFFER_LOAD_DWORD*, addr64=0
+multiclass MUBUF_Load_Dword <ValueType vt, MUBUF offset, MUBUF offen, MUBUF idxen,
+ MUBUF bothen> {
+
+ def : Pat <
+ (vt (int_SI_buffer_load_dword i128:$rsrc, i32:$vaddr, i32:$soffset,
+ imm:$offset, 0, 0, imm:$glc, imm:$slc,
+ imm:$tfe)),
+ (offset $rsrc, $vaddr, (as_i16imm $offset), $soffset, (as_i1imm $glc),
+ (as_i1imm $slc), (as_i1imm $tfe))
+ >;
+
+ def : Pat <
+ (vt (int_SI_buffer_load_dword i128:$rsrc, i32:$vaddr, i32:$soffset,
+ imm, 1, 0, imm:$glc, imm:$slc,
+ imm:$tfe)),
+ (offen $rsrc, $vaddr, $soffset, (as_i1imm $glc), (as_i1imm $slc),
+ (as_i1imm $tfe))
+ >;
+
+ def : Pat <
+ (vt (int_SI_buffer_load_dword i128:$rsrc, i32:$vaddr, i32:$soffset,
+ imm:$offset, 0, 1, imm:$glc, imm:$slc,
+ imm:$tfe)),
+ (idxen $rsrc, $vaddr, (as_i16imm $offset), $soffset, (as_i1imm $glc),
+ (as_i1imm $slc), (as_i1imm $tfe))
+ >;
+
+ def : Pat <
+ (vt (int_SI_buffer_load_dword i128:$rsrc, v2i32:$vaddr, i32:$soffset,
+ imm, 1, 1, imm:$glc, imm:$slc,
+ imm:$tfe)),
+ (bothen $rsrc, $vaddr, $soffset, (as_i1imm $glc), (as_i1imm $slc),
+ (as_i1imm $tfe))
+ >;
+}
+
+defm : MUBUF_Load_Dword <i32, BUFFER_LOAD_DWORD_OFFSET, BUFFER_LOAD_DWORD_OFFEN,
+ BUFFER_LOAD_DWORD_IDXEN, BUFFER_LOAD_DWORD_BOTHEN>;
+defm : MUBUF_Load_Dword <v2i32, BUFFER_LOAD_DWORDX2_OFFSET, BUFFER_LOAD_DWORDX2_OFFEN,
+ BUFFER_LOAD_DWORDX2_IDXEN, BUFFER_LOAD_DWORDX2_BOTHEN>;
+defm : MUBUF_Load_Dword <v4i32, BUFFER_LOAD_DWORDX4_OFFSET, BUFFER_LOAD_DWORDX4_OFFEN,
+ BUFFER_LOAD_DWORDX4_IDXEN, BUFFER_LOAD_DWORDX4_BOTHEN>;
+
//===----------------------------------------------------------------------===//
// MTBUF Patterns
//===----------------------------------------------------------------------===//
diff --git a/lib/Target/R600/SIIntrinsics.td b/lib/Target/R600/SIIntrinsics.td
index efcdc84..00e32c0 100644
--- a/lib/Target/R600/SIIntrinsics.td
+++ b/lib/Target/R600/SIIntrinsics.td
@@ -38,6 +38,20 @@ let TargetPrefix = "SI", isTarget = 1 in {
llvm_i32_ty], // tfe(imm)
[]>;
+ // Fully-flexible BUFFER_LOAD_DWORD_* except for the ADDR64 bit, which is not exposed
+ def int_SI_buffer_load_dword : Intrinsic <
+ [llvm_anyint_ty], // vdata(VGPR), overloaded for types i32, v2i32, v4i32
+ [llvm_anyint_ty, // rsrc(SGPR)
+ llvm_anyint_ty, // vaddr(VGPR)
+ llvm_i32_ty, // soffset(SGPR)
+ llvm_i32_ty, // inst_offset(imm)
+ llvm_i32_ty, // offen(imm)
+ llvm_i32_ty, // idxen(imm)
+ llvm_i32_ty, // glc(imm)
+ llvm_i32_ty, // slc(imm)
+ llvm_i32_ty], // tfe(imm)
+ [IntrReadArgMem]>;
+
def int_SI_sendmsg : Intrinsic <[], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
class Sample : Intrinsic <[llvm_v4f32_ty], [llvm_anyvector_ty, llvm_v32i8_ty, llvm_anyint_ty, llvm_i32_ty], [IntrNoMem]>;
diff --git a/test/CodeGen/R600/llvm.SI.load.dword.ll b/test/CodeGen/R600/llvm.SI.load.dword.ll
new file mode 100644
index 0000000..a622775
--- /dev/null
+++ b/test/CodeGen/R600/llvm.SI.load.dword.ll
@@ -0,0 +1,40 @@
+;RUN: llc < %s -march=r600 -mcpu=verde -verify-machineinstrs | FileCheck %s
+
+; Example of a simple geometry shader loading vertex attributes from the
+; ESGS ring buffer
+
+; CHECK-LABEL: @main
+; CHECK: BUFFER_LOAD_DWORD
+; CHECK: BUFFER_LOAD_DWORD
+; CHECK: BUFFER_LOAD_DWORD
+; CHECK: BUFFER_LOAD_DWORD
+
+define void @main([17 x <16 x i8>] addrspace(2)* byval, [32 x <16 x i8>] addrspace(2)* byval, [16 x <32 x i8>] addrspace(2)* byval, [2 x <16 x i8>] addrspace(2)* byval, [17 x <16 x i8>] addrspace(2)* inreg, [17 x <16 x i8>] addrspace(2)* inreg, i32, i32, i32, i32) #0 {
+main_body:
+ %10 = getelementptr [2 x <16 x i8>] addrspace(2)* %3, i64 0, i32 1
+ %11 = load <16 x i8> addrspace(2)* %10, !tbaa !0
+ %12 = shl i32 %6, 2
+ %13 = call i32 @llvm.SI.buffer.load.dword.i32.i32(<16 x i8> %11, i32 0, i32 0, i32 0, i32 0, i32 0, i32 1, i32 1, i32 0)
+ %14 = bitcast i32 %13 to float
+ %15 = call i32 @llvm.SI.buffer.load.dword.i32.i32(<16 x i8> %11, i32 %12, i32 0, i32 0, i32 1, i32 0, i32 1, i32 1, i32 0)
+ %16 = bitcast i32 %15 to float
+ %17 = call i32 @llvm.SI.buffer.load.dword.i32.i32(<16 x i8> %11, i32 %12, i32 0, i32 0, i32 0, i32 1, i32 1, i32 1, i32 0)
+ %18 = bitcast i32 %17 to float
+ %19 = call i32 @llvm.SI.buffer.load.dword.i32.v2i32(<16 x i8> %11, <2 x i32> <i32 0, i32 0>, i32 0, i32 0, i32 1, i32 1, i32 1, i32 1, i32 0)
+ %20 = bitcast i32 %19 to float
+ call void @llvm.SI.export(i32 15, i32 0, i32 1, i32 12, i32 0, float %14, float %16, float %18, float %20)
+ ret void
+}
+
+; Function Attrs: nounwind readonly
+declare i32 @llvm.SI.buffer.load.dword.i32.i32(<16 x i8>, i32, i32, i32, i32, i32, i32, i32, i32) #1
+
+; Function Attrs: nounwind readonly
+declare i32 @llvm.SI.buffer.load.dword.i32.v2i32(<16 x i8>, <2 x i32>, i32, i32, i32, i32, i32, i32, i32) #1
+
+declare void @llvm.SI.export(i32, i32, i32, i32, i32, float, float, float, float)
+
+attributes #0 = { "ShaderType"="1" }
+attributes #1 = { nounwind readonly }
+
+!0 = metadata !{metadata !"const", null, i32 1}
--
1.8.5.3
More information about the mesa-dev
mailing list