R600: Try to convert BFE back to standard bit ops when possible.

This allows existing DAG combines to work on them, and then
we can re-match to BFE if necessary during instruction selection.

llvm-svn: 209462

4 files changed, 276 insertions, 11 deletions

diff --git a/llvm/test/CodeGen/R600/llvm.AMDGPU.bfe.i32.ll b/llvm/test/CodeGen/R600/llvm.AMDGPU.bfe.i32.ll
index 71d2b6e7c2e..eb509423282 100644
--- a/llvm/test/CodeGen/R600/llvm.AMDGPU.bfe.i32.ll
+++ b/llvm/test/CodeGen/R600/llvm.AMDGPU.bfe.i32.ll
@@ -69,6 +69,115 @@ define void @bfe_i32_arg_0_width_imm_offset(i32 addrspace(1)* %out, i32 %src0, i
   ret void
 }
 
+; FUNC-LABEL: @bfe_i32_test_6
+; SI: V_LSHLREV_B32_e32 v{{[0-9]+}}, 31, v{{[0-9]+}}
+; SI: V_ASHRREV_I32_e32 v{{[0-9]+}}, 1, v{{[0-9]+}}
+; SI: S_ENDPGM
+define void @bfe_i32_test_6(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %shl = shl i32 %x, 31
+  %bfe = call i32 @llvm.AMDGPU.bfe.i32(i32 %shl, i32 1, i32 31)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_i32_test_7
+; SI-NOT: SHL
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+define void @bfe_i32_test_7(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %shl = shl i32 %x, 31
+  %bfe = call i32 @llvm.AMDGPU.bfe.i32(i32 %shl, i32 0, i32 31)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FIXME: The shifts should be 1 BFE
+; FUNC-LABEL: @bfe_i32_test_8
+; SI: BUFFER_LOAD_DWORD
+; SI: V_BFE_I32 v{{[0-9]+}}, v{{[0-9]+}}, 0, 1
+; SI: S_ENDPGM
+define void @bfe_i32_test_8(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %shl = shl i32 %x, 31
+  %bfe = call i32 @llvm.AMDGPU.bfe.i32(i32 %shl, i32 31, i32 1)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_i32_test_9
+; SI-NOT: BFE
+; SI: V_ASHRREV_I32_e32 v{{[0-9]+}}, 31, v{{[0-9]+}}
+; SI-NOT: BFE
+; SI: S_ENDPGM
+define void @bfe_i32_test_9(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %bfe = call i32 @llvm.AMDGPU.bfe.i32(i32 %x, i32 31, i32 1)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_i32_test_10
+; SI-NOT: BFE
+; SI: V_ASHRREV_I32_e32 v{{[0-9]+}}, 1, v{{[0-9]+}}
+; SI-NOT: BFE
+; SI: S_ENDPGM
+define void @bfe_i32_test_10(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %bfe = call i32 @llvm.AMDGPU.bfe.i32(i32 %x, i32 1, i32 31)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_i32_test_11
+; SI-NOT: BFE
+; SI: V_ASHRREV_I32_e32 v{{[0-9]+}}, 8, v{{[0-9]+}}
+; SI-NOT: BFE
+; SI: S_ENDPGM
+define void @bfe_i32_test_11(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %bfe = call i32 @llvm.AMDGPU.bfe.i32(i32 %x, i32 8, i32 24)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_i32_test_12
+; SI-NOT: BFE
+; SI: V_ASHRREV_I32_e32 v{{[0-9]+}}, 24, v{{[0-9]+}}
+; SI-NOT: BFE
+; SI: S_ENDPGM
+define void @bfe_i32_test_12(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %bfe = call i32 @llvm.AMDGPU.bfe.i32(i32 %x, i32 24, i32 8)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_i32_test_13
+; SI: V_ASHRREV_I32_e32 {{v[0-9]+}}, 31, {{v[0-9]+}}
+; SI-NOT: BFE
+; SI: S_ENDPGM
+define void @bfe_i32_test_13(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %shl = ashr i32 %x, 31
+  %bfe = call i32 @llvm.AMDGPU.bfe.i32(i32 %shl, i32 31, i32 1)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4 ret void
+}
+
+; FUNC-LABEL: @bfe_i32_test_14
+; SI-NOT: LSHR
+; SI-NOT: BFE
+; SI: S_ENDPGM
+define void @bfe_i32_test_14(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %shl = lshr i32 %x, 31
+  %bfe = call i32 @llvm.AMDGPU.bfe.i32(i32 %shl, i32 31, i32 1)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4 ret void
+}
+
 ; FUNC-LABEL: @bfe_i32_constant_fold_test_0
 ; SI-NOT: BFE
 ; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0
@@ -296,3 +405,22 @@ define void @bfe_i32_constant_fold_test_18(i32 addrspace(1)* %out) nounwind {
   store i32 %bfe_i32, i32 addrspace(1)* %out, align 4
   ret void
 }
+
+; XXX - This should really be a single BFE, but the sext_inreg of the
+; extended type i24 is never custom lowered.
+; FUNC-LABEL: @bfe_sext_in_reg_i24
+; SI: BUFFER_LOAD_DWORD [[LOAD:v[0-9]+]],
+; SI: V_LSHLREV_B32_e32 {{v[0-9]+}}, 8, {{v[0-9]+}}
+; SI: V_ASHRREV_I32_e32 {{v[0-9]+}}, 8, {{v[0-9]+}}
+; XSI: V_BFE_I32 [[BFE:v[0-9]+]], [[LOAD]], 0, 8
+; XSI-NOT: SHL
+; XSI-NOT: SHR
+; XSI: BUFFER_STORE_DWORD [[BFE]],
+define void @bfe_sext_in_reg_i24(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %bfe = call i32 @llvm.AMDGPU.bfe.i32(i32 %x, i32 0, i32 24)
+  %shl = shl i32 %bfe, 8
+  %ashr = ashr i32 %shl, 8
+  store i32 %ashr, i32 addrspace(1)* %out, align 4
+  ret void
+}
diff --git a/llvm/test/CodeGen/R600/llvm.AMDGPU.bfe.u32.ll b/llvm/test/CodeGen/R600/llvm.AMDGPU.bfe.u32.ll
index 6ed1ad5d2e6..1a62253eeb7 100644
--- a/llvm/test/CodeGen/R600/llvm.AMDGPU.bfe.u32.ll
+++ b/llvm/test/CodeGen/R600/llvm.AMDGPU.bfe.u32.ll
@@ -161,9 +161,9 @@ define void @bfe_u32_zext_in_reg_i16_offset_8(i32 addrspace(1)* %out, i32 addrsp
 
 ; FUNC-LABEL: @bfe_u32_test_1
 ; SI: BUFFER_LOAD_DWORD
-; SI: V_BFE_U32 {{v[0-9]+}}, {{v[0-9]+}}, 0, 1
+; SI: V_AND_B32_e32 {{v[0-9]+}}, 1, {{v[0-9]+}}
 ; SI: S_ENDPGM
-; EG: BFE_UINT
+; EG: AND_INT T{{[0-9]\.[XYZW]}}, T{{[0-9]\.[XYZW]}}, 1,
 define void @bfe_u32_test_1(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
   %x = load i32 addrspace(1)* %in, align 4
   %bfe = call i32 @llvm.AMDGPU.bfe.u32(i32 %x, i32 0, i32 1)
@@ -220,7 +220,7 @@ define void @bfe_u32_test_5(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounw
 
 ; FUNC-LABEL: @bfe_u32_test_6
 ; SI: V_LSHLREV_B32_e32 v{{[0-9]+}}, 31, v{{[0-9]+}}
-; SI: V_BFE_U32 {{v[0-9]+}}, {{v[0-9]+}}, 1, 31
+; SI: V_LSHRREV_B32_e32 v{{[0-9]+}}, 1, v{{[0-9]+}}
 ; SI: S_ENDPGM
 define void @bfe_u32_test_6(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
   %x = load i32 addrspace(1)* %in, align 4
@@ -243,8 +243,9 @@ define void @bfe_u32_test_7(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounw
 }
 
 ; FUNC-LABEL: @bfe_u32_test_8
-; SI: V_LSHLREV_B32_e32 v{{[0-9]+}}, 31, v{{[0-9]+}}
-; SI: V_BFE_U32 {{v[0-9]+}}, {{v[0-9]+}}, 31, 1
+; SI-NOT: BFE
+; SI: V_AND_B32_e32 {{v[0-9]+}}, 1, {{v[0-9]+}}
+; SI-NOT: BFE
 ; SI: S_ENDPGM
 define void @bfe_u32_test_8(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
   %x = load i32 addrspace(1)* %in, align 4
@@ -254,6 +255,76 @@ define void @bfe_u32_test_8(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounw
   ret void
 }
 
+; FUNC-LABEL: @bfe_u32_test_9
+; SI-NOT: BFE
+; SI: V_LSHRREV_B32_e32 v{{[0-9]+}}, 31, v{{[0-9]+}}
+; SI-NOT: BFE
+; SI: S_ENDPGM
+define void @bfe_u32_test_9(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %bfe = call i32 @llvm.AMDGPU.bfe.u32(i32 %x, i32 31, i32 1)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_test_10
+; SI-NOT: BFE
+; SI: V_LSHRREV_B32_e32 v{{[0-9]+}}, 1, v{{[0-9]+}}
+; SI-NOT: BFE
+; SI: S_ENDPGM
+define void @bfe_u32_test_10(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %bfe = call i32 @llvm.AMDGPU.bfe.u32(i32 %x, i32 1, i32 31)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_test_11
+; SI-NOT: BFE
+; SI: V_LSHRREV_B32_e32 v{{[0-9]+}}, 8, v{{[0-9]+}}
+; SI-NOT: BFE
+; SI: S_ENDPGM
+define void @bfe_u32_test_11(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %bfe = call i32 @llvm.AMDGPU.bfe.u32(i32 %x, i32 8, i32 24)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_test_12
+; SI-NOT: BFE
+; SI: V_LSHRREV_B32_e32 v{{[0-9]+}}, 24, v{{[0-9]+}}
+; SI-NOT: BFE
+; SI: S_ENDPGM
+define void @bfe_u32_test_12(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %bfe = call i32 @llvm.AMDGPU.bfe.u32(i32 %x, i32 24, i32 8)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_test_13
+; V_ASHRREV_U32_e32 {{v[0-9]+}}, 31, {{v[0-9]+}}
+; SI-NOT: BFE
+; SI: S_ENDPGM
+define void @bfe_u32_test_13(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %shl = ashr i32 %x, 31
+  %bfe = call i32 @llvm.AMDGPU.bfe.u32(i32 %shl, i32 31, i32 1)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4 ret void
+}
+
+; FUNC-LABEL: @bfe_u32_test_14
+; SI-NOT: LSHR
+; SI-NOT: BFE
+; SI: S_ENDPGM
+define void @bfe_u32_test_14(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %shl = lshr i32 %x, 31
+  %bfe = call i32 @llvm.AMDGPU.bfe.u32(i32 %shl, i32 31, i32 1)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4 ret void
+}
+
 ; FUNC-LABEL: @bfe_u32_constant_fold_test_0
 ; SI-NOT: BFE
 ; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0
@@ -457,3 +528,27 @@ define void @bfe_u32_constant_fold_test_16(i32 addrspace(1)* %out) nounwind {
   store i32 %bfe_u32, i32 addrspace(1)* %out, align 4
   ret void
 }
+
+; FUNC-LABEL: @bfe_u32_constant_fold_test_17
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0x7f
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_u32_constant_fold_test_17(i32 addrspace(1)* %out) nounwind {
+  %bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 255, i32 1, i32 31) nounwind readnone
+  store i32 %bfe_u32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_constant_fold_test_18
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_u32_constant_fold_test_18(i32 addrspace(1)* %out) nounwind {
+  %bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 255, i32 31, i32 1) nounwind readnone
+  store i32 %bfe_u32, i32 addrspace(1)* %out, align 4
+  ret void
+}
diff --git a/llvm/test/CodeGen/R600/sext-in-reg.ll b/llvm/test/CodeGen/R600/sext-in-reg.ll
index 404c9b8b812..1b02e4bf801 100644
--- a/llvm/test/CodeGen/R600/sext-in-reg.ll
+++ b/llvm/test/CodeGen/R600/sext-in-reg.ll
@@ -417,8 +417,8 @@ define void @bfe_8_bfe_16(i32 addrspace(1)* %out, i32 addrspace(1)* %ptr) nounwi
 
 ; This really should be folded into 1
 ; FUNC-LABEL: @bfe_16_bfe_8
-; SI: V_BFE_I32 v{{[0-9]+}}, v{{[0-9]+}}, 0, 16
 ; SI: V_BFE_I32 v{{[0-9]+}}, v{{[0-9]+}}, 0, 8
+; SI-NOT: BFE
 ; SI: S_ENDPGM
 define void @bfe_16_bfe_8(i32 addrspace(1)* %out, i32 addrspace(1)* %ptr) nounwind {
   %load = load i32 addrspace(1)* %ptr, align 4
@@ -430,7 +430,7 @@ define void @bfe_16_bfe_8(i32 addrspace(1)* %out, i32 addrspace(1)* %ptr) nounwi
 
 ; Make sure there isn't a redundant BFE
 ; FUNC-LABEL: @sext_in_reg_i8_to_i32_bfe
-; SI: S_BFE_I32 s{{[0-9]+}}, s{{[0-9]+}}, 0x80000
+; SI: S_SEXT_I32_I8 s{{[0-9]+}}, s{{[0-9]+}}
 ; SI-NOT: BFE
 ; SI: S_ENDPGM
 define void @sext_in_reg_i8_to_i32_bfe(i32 addrspace(1)* %out, i32 %a, i32 %b) nounwind {
@@ -478,3 +478,47 @@ define void @sextload_i8_to_i32_bfe_0(i32 addrspace(1)* %out, i8 addrspace(1)* %
   store i32 %ashr, i32 addrspace(1)* %out, align 4
   ret void
 }
+
+; FUNC-LABEL: @sext_in_reg_i1_bfe_offset_0:
+; SI-NOT: SHR
+; SI-NOT: SHL
+; SI: V_BFE_I32 v{{[0-9]+}}, v{{[0-9]+}}, 0, 1
+; SI: S_ENDPGM
+define void @sext_in_reg_i1_bfe_offset_0(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %shl = shl i32 %x, 31
+  %shr = ashr i32 %shl, 31
+  %bfe = call i32 @llvm.AMDGPU.bfe.i32(i32 %shr, i32 0, i32 1)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @sext_in_reg_i1_bfe_offset_1
+; SI: BUFFER_LOAD_DWORD
+; SI-NOT: SHL
+; SI-NOT: SHR
+; SI: V_BFE_I32 v{{[0-9]+}}, v{{[0-9]+}}, 1, 1
+; SI: S_ENDPGM
+define void @sext_in_reg_i1_bfe_offset_1(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %shl = shl i32 %x, 30
+  %shr = ashr i32 %shl, 30
+  %bfe = call i32 @llvm.AMDGPU.bfe.i32(i32 %shr, i32 1, i32 1)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @sext_in_reg_i2_bfe_offset_1:
+; SI: BUFFER_LOAD_DWORD
+; SI: V_LSHLREV_B32_e32 v{{[0-9]+}}, 30, v{{[0-9]+}}
+; SI: V_ASHRREV_I32_e32 v{{[0-9]+}}, 30, v{{[0-9]+}}
+; SI: V_BFE_I32 v{{[0-9]+}}, v{{[0-9]+}}, 1, 2
+; SI: S_ENDPGM
+define void @sext_in_reg_i2_bfe_offset_1(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %shl = shl i32 %x, 30
+  %shr = ashr i32 %shl, 30
+  %bfe = call i32 @llvm.AMDGPU.bfe.i32(i32 %shr, i32 1, i32 2)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
diff --git a/llvm/test/CodeGen/R600/udivrem64.ll b/llvm/test/CodeGen/R600/udivrem64.ll
index b3caebf123d..a71315a12d8 100644
--- a/llvm/test/CodeGen/R600/udivrem64.ll
+++ b/llvm/test/CodeGen/R600/udivrem64.ll
@@ -3,8 +3,7 @@
 
 ;FUNC-LABEL: @test_udiv
 ;EG: RECIP_UINT
-;EG: BFE_UINT
-;EG: BFE_UINT
+;EG: LSHL {{.*}}, 1,
 ;EG: BFE_UINT
 ;EG: BFE_UINT
 ;EG: BFE_UINT
@@ -74,8 +73,7 @@ define void @test_udiv(i64 addrspace(1)* %out, i64 %x, i64 %y) {
 ;EG: BFE_UINT
 ;EG: BFE_UINT
 ;EG: BFE_UINT
-;EG: BFE_UINT
-;EG: BFE_UINT
+;EG: AND_INT {{.*}}, 1,
 ;SI: S_ENDPGM
 define void @test_urem(i64 addrspace(1)* %out, i64 %x, i64 %y) {
   %result = urem i64 %x, %y