R600/SI: Try to keep mul on SALU

[arsenm]

Summary

Also fix bug this exposed where when legalizing an immediate
operand, a v_mov_b32 would be created with a VSrc dest register.

[arsenm]

Ping

[arsenm]

r217108

lib/Target/R600/SIInstrInfo.cpp

   case AMDGPU::S_ADDC_U32: return AMDGPU::V_ADDC_U32_e32;
   case AMDGPU::S_SUB_I32: return AMDGPU::V_SUB_I32_e32;
   case AMDGPU::S_SUBB_U32: return AMDGPU::V_SUBB_U32_e32;
+  case AMDGPU::S_MUL_I32: return AMDGPU::V_MUL_LO_I32;
   case AMDGPU::S_AND_B32: return AMDGPU::V_AND_B32_e32;
   case AMDGPU::S_OR_B32: return AMDGPU::V_OR_B32_e32;
   case AMDGPU::S_XOR_B32: return AMDGPU::V_XOR_B32_e32;

   unsigned RCID = get(MI->getOpcode()).OpInfo[OpIdx].RegClass;
   const TargetRegisterClass *RC = RI.getRegClass(RCID);
   unsigned Opcode = AMDGPU::V_MOV_B32_e32;
+
   if (MO.isReg()) {
     Opcode = AMDGPU::COPY;
   } else if (RI.isSGPRClass(RC)) {
     Opcode = AMDGPU::S_MOV_B32;
+  } else if (MO.isImm()) {
+    if (RC == &AMDGPU::VSrc_32RegClass)
+      Opcode = AMDGPU::S_MOV_B32;
   }
 
   const TargetRegisterClass *VRC = RI.getEquivalentVGPRClass(RC);

lib/Target/R600/SIInstructions.td

   [(set i64:$dst, (sra i64:$src0, i32:$src1))]
 >;
 
-} // End AddedComplexity = 1
 
 def S_BFM_B32 : SOP2_32 <0x00000024, "S_BFM_B32", []>;
 def S_BFM_B64 : SOP2_64 <0x00000025, "S_BFM_B64", []>;
-def S_MUL_I32 : SOP2_32 <0x00000026, "S_MUL_I32", []>;
+def S_MUL_I32 : SOP2_32 <0x00000026, "S_MUL_I32",
+  [(set i32:$dst, (mul i32:$src0, i32:$src1))]
+>;
+
+} // End AddedComplexity = 1
+
 def S_BFE_U32 : SOP2_32 <0x00000027, "S_BFE_U32", []>;
 def S_BFE_I32 : SOP2_32 <0x00000028, "S_BFE_I32", []>;
 def S_BFE_U64 : SOP2_64 <0x00000029, "S_BFE_U64", []>;

 def : UMad24Pat<V_MAD_U32_U24>;
 
 def : Pat <
-  (mul i32:$src0, i32:$src1),
-  (V_MUL_LO_I32 $src0, $src1)
->;
-
-def : Pat <
   (mulhu i32:$src0, i32:$src1),
   (V_MUL_HI_U32 $src0, $src1)
 >;

   (V_MUL_HI_I32 $src0, $src1)
 >;
 
+def : Pat <
+  (mul i32:$src0, i32:$src1),
+  (V_MUL_LO_I32 $src0, $src1)
+>;
+
 def : Vop3ModPat<V_MAD_F32, VOP_F32_F32_F32_F32, AMDGPUmad>;
 
 

test/CodeGen/R600/codegen-prepare-addrmode-sext.ll

 ; FUNC-LABEL: @test
 ; OPT: mul nsw i32
 ; OPT-NEXT: sext
-; SI-LLC: V_MUL_LO_I32
-; SI-LLC-NOT: V_MUL_HI
+; SI-LLC: S_MUL_I32
+; SI-LLC-NOT: MUL
 define void @test(i8 addrspace(1)* nocapture readonly %in, i32 %a, i8 %b) {
 entry:
   %0 = mul nsw i32 %a, 3

test/CodeGen/R600/mul.ll

 
 ; mul24 and mad24 are affected
 
-; FUNC-LABEL: @test2
+; FUNC-LABEL: @test_mul_v2i32
 ; EG: MULLO_INT {{\*? *}}T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
 ; EG: MULLO_INT {{\*? *}}T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
 
 ; SI: V_MUL_LO_I32 v{{[0-9]+, v[0-9]+, v[0-9]+}}
 ; SI: V_MUL_LO_I32 v{{[0-9]+, v[0-9]+, v[0-9]+}}
 
-define void @test2(<2 x i32> addrspace(1)* %out, <2 x i32> addrspace(1)* %in) {
+define void @test_mul_v2i32(<2 x i32> addrspace(1)* %out, <2 x i32> addrspace(1)* %in) {
   %b_ptr = getelementptr <2 x i32> addrspace(1)* %in, i32 1
   %a = load <2 x i32> addrspace(1) * %in
   %b = load <2 x i32> addrspace(1) * %b_ptr

   ret void
 }
 
-; FUNC-LABEL: @test4
+; FUNC-LABEL: @v_mul_v4i32
 ; EG: MULLO_INT {{\*? *}}T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
 ; EG: MULLO_INT {{\*? *}}T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
 ; EG: MULLO_INT {{\*? *}}T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}

 ; SI: V_MUL_LO_I32 v{{[0-9]+, v[0-9]+, v[0-9]+}}
 ; SI: V_MUL_LO_I32 v{{[0-9]+, v[0-9]+, v[0-9]+}}
 
-define void @test4(<4 x i32> addrspace(1)* %out, <4 x i32> addrspace(1)* %in) {
+define void @v_mul_v4i32(<4 x i32> addrspace(1)* %out, <4 x i32> addrspace(1)* %in) {
   %b_ptr = getelementptr <4 x i32> addrspace(1)* %in, i32 1
   %a = load <4 x i32> addrspace(1) * %in
   %b = load <4 x i32> addrspace(1) * %b_ptr

   ret void
 }
 
-; FUNC-LABEL: @trunc_i64_mul_to_i32
+; FUNC-LABEL: @s_trunc_i64_mul_to_i32
+; SI: S_LOAD_DWORD
+; SI: S_LOAD_DWORD
+; SI: S_MUL_I32
+; SI: BUFFER_STORE_DWORD
+define void @s_trunc_i64_mul_to_i32(i32 addrspace(1)* %out, i64 %a, i64 %b) {
+  %mul = mul i64 %b, %a
+  %trunc = trunc i64 %mul to i32
+  store i32 %trunc, i32 addrspace(1)* %out, align 8
+  ret void
+}
+
+; FUNC-LABEL: @v_trunc_i64_mul_to_i32
 ; SI: S_LOAD_DWORD
 ; SI: S_LOAD_DWORD
 ; SI: V_MUL_LO_I32
 ; SI: BUFFER_STORE_DWORD
-define void @trunc_i64_mul_to_i32(i32 addrspace(1)* %out, i64 %a, i64 %b) {
+define void @v_trunc_i64_mul_to_i32(i32 addrspace(1)* %out, i64 addrspace(1)* %aptr, i64 addrspace(1)* %bptr) nounwind {
+  %a = load i64 addrspace(1)* %aptr, align 8
+  %b = load i64 addrspace(1)* %bptr, align 8
   %mul = mul i64 %b, %a
   %trunc = trunc i64 %mul to i32
   store i32 %trunc, i32 addrspace(1)* %out, align 8

 ; FUNC-LABEL: @mul64_sext_c
 ; EG-DAG: MULLO_INT
 ; EG-DAG: MULHI_INT
-; SI-DAG: V_MUL_LO_I32
+; SI-DAG: S_MUL_I32
 ; SI-DAG: V_MUL_HI_I32
 define void @mul64_sext_c(i64 addrspace(1)* %out, i32 %in) {
 entry:

   ret void
 }
 
+; FUNC-LABEL: @v_mul64_sext_c:
+; EG-DAG: MULLO_INT
+; EG-DAG: MULHI_INT
+; SI-DAG: V_MUL_LO_I32
+; SI-DAG: V_MUL_HI_I32
+; SI: S_ENDPGM
+define void @v_mul64_sext_c(i64 addrspace(1)* %out, i32 addrspace(1)* %in) {
+  %val = load i32 addrspace(1)* %in, align 4
+  %ext = sext i32 %val to i64
+  %mul = mul i64 %ext, 80
+  store i64 %mul, i64 addrspace(1)* %out, align 8
+  ret void
+}
+
+; FUNC-LABEL: @v_mul64_sext_inline_imm:
+; SI-DAG: V_MUL_LO_I32 v{{[0-9]+}}, 9, v{{[0-9]+}}
+; SI-DAG: V_MUL_HI_I32 v{{[0-9]+}}, 9, v{{[0-9]+}}
+; SI: S_ENDPGM
+define void @v_mul64_sext_inline_imm(i64 addrspace(1)* %out, i32 addrspace(1)* %in) {
+  %val = load i32 addrspace(1)* %in, align 4
+  %ext = sext i32 %val to i64
+  %mul = mul i64 %ext, 9
+  store i64 %mul, i64 addrspace(1)* %out, align 8
+  ret void
+}
+
+; FUNC-LABEL: @s_mul_i32:
+; SI: S_LOAD_DWORD [[SRC0:s[0-9]+]],
+; SI: S_LOAD_DWORD [[SRC1:s[0-9]+]],
+; SI: S_MUL_I32 [[SRESULT:s[0-9]+]], [[SRC0]], [[SRC1]]
+; SI: V_MOV_B32_e32 [[VRESULT:v[0-9]+]], [[SRESULT]]
+; SI: BUFFER_STORE_DWORD [[VRESULT]],
+; SI: S_ENDPGM
+define void @s_mul_i32(i32 addrspace(1)* %out, i32 %a, i32 %b) nounwind {
+  %mul = mul i32 %a, %b
+  store i32 %mul, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @v_mul_i32
+; SI: V_MUL_LO_I32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
+define void @v_mul_i32(i32 addrspace(1)* %out, i32 addrspace(1)* %in) {
+  %b_ptr = getelementptr i32 addrspace(1)* %in, i32 1
+  %a = load i32 addrspace(1)* %in
+  %b = load i32 addrspace(1)* %b_ptr
+  %result = mul i32 %a, %b
+  store i32 %result, i32 addrspace(1)* %out
+  ret void
+}
+
 ; A standard 64-bit multiply.  The expansion should be around 6 instructions.
 ; It would be difficult to match the expansion correctly without writing
 ; a really complicated list of FileCheck expressions.  I don't want
 ; to confuse people who may 'break' this test with a correct optimization,
 ; so this test just uses FUNC-LABEL to make sure the compiler does not
 ; crash with a 'failed to select' error.
-; FUNC-LABEL: @mul64
-define void @mul64(i64 addrspace(1)* %out, i64 %a, i64 %b) {
+
+; FUNC-LABEL: @s_mul_i64:
+define void @s_mul_i64(i64 addrspace(1)* %out, i64 %a, i64 %b) nounwind {
+  %mul = mul i64 %a, %b
+  store i64 %mul, i64 addrspace(1)* %out, align 8
+  ret void
+}
+
+; FUNC-LABEL: @v_mul_i64
+; SI: V_MUL_LO_I32
+define void @v_mul_i64(i64 addrspace(1)* %out, i64 addrspace(1)* %aptr, i64 addrspace(1)* %bptr) {
+  %a = load i64 addrspace(1)* %aptr, align 8
+  %b = load i64 addrspace(1)* %bptr, align 8
+  %mul = mul i64 %a, %b
+  store i64 %mul, i64 addrspace(1)* %out, align 8
+  ret void
+}
+
+; FUNC-LABEL: @mul32_in_branch
+; SI: V_MUL_LO_I32
+define void @mul32_in_branch(i32 addrspace(1)* %out, i32 addrspace(1)* %in, i32 %a, i32 %b, i32 %c) {
 entry:
-  %0 = mul i64 %a, %b
-  store i64 %0, i64 addrspace(1)* %out
+  %0 = icmp eq i32 %a, 0
+  br i1 %0, label %if, label %else
+
+if:
+  %1 = load i32 addrspace(1)* %in
+  br label %endif
+
+else:
+  %2 = mul i32 %a, %b
+  br label %endif
+
+endif:
+  %3 = phi i32 [%1, %if], [%2, %else]
+  store i32 %3, i32 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: @mul64_in_branch
+; SI-DAG: V_MUL_LO_I32
+; SI-DAG: V_MUL_HI_U32
+; SI: S_ENDPGM
+define void @mul64_in_branch(i64 addrspace(1)* %out, i64 addrspace(1)* %in, i64 %a, i64 %b, i64 %c) {
+entry:
+  %0 = icmp eq i64 %a, 0
+  br i1 %0, label %if, label %else
+
+if:
+  %1 = load i64 addrspace(1)* %in
+  br label %endif
+
+else:
+  %2 = mul i64 %a, %b
+  br label %endif
+
+endif:
+  %3 = phi i64 [%1, %if], [%2, %else]
+  store i64 %3, i64 addrspace(1)* %out
   ret void
 }

test/CodeGen/R600/sign_extend.ll

   ret void
 }
 
-; SI-LABEL: @test:
-; SI: V_ASHR
+; SI-LABEL: @test_s_sext_i32_to_i64:
+; SI: S_ASHR_I32
 ; SI: S_ENDPG
-define void @test(i64 addrspace(1)* %out, i32 %a, i32 %b, i32 %c) nounwind {
+define void @test_s_sext_i32_to_i64(i64 addrspace(1)* %out, i32 %a, i32 %b, i32 %c) nounwind {
 entry:
   %mul = mul i32 %a, %b
   %add = add i32 %mul, %c