R600/SI: Add patterns for v_cvt_{flr|rpi}_i32_f32

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@226230 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/R600/AMDGPUInstructions.td b/lib/Target/R600/AMDGPUInstructions.td
index 4e536c37b0bd69e9a876309740f3b5406035104b..e42796bd14a50223d3dcd6d2b0fe6113ece87331 100644 (file)
--- a/lib/Target/R600/AMDGPUInstructions.td
+++ b/lib/Target/R600/AMDGPUInstructions.td
@@ -438,6 +438,11 @@ def FP_ONE : PatLeaf <
   [{return N->isExactlyValue(1.0);}]
 >;
 
+def FP_HALF : PatLeaf <
+  (fpimm),
+  [{return N->isExactlyValue(0.5);}]
+>;
+
 let isCodeGenOnly = 1, isPseudo = 1 in {
 
 let usesCustomInserter = 1  in {
@@ -603,6 +608,18 @@ class ROTRPattern <Instruction BIT_ALIGN> : Pat <
 // 24-bit arithmetic patterns
 def umul24 : PatFrag <(ops node:$x, node:$y), (mul node:$x, node:$y)>;
 
+// Special conversion patterns
+
+def cvt_rpi_i32_f32 : PatFrag <
+  (ops node:$src),
+  (fp_to_sint (ffloor (fadd $src, FP_HALF)))
+>;
+
+def cvt_flr_i32_f32 : PatFrag <
+  (ops node:$src),
+  (fp_to_sint (ffloor $src))
+>;
+
 /*
 class UMUL24Pattern <Instruction UMUL24> : Pat <
   (mul U24:$x, U24:$y),

diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
index dfb442252cf8acc6b8a5b2fac6dea7d86584e1e1..c09bed7840fcabd92f60e9085122614cc85080b3 100644 (file)
--- a/lib/Target/R600/SIInstructions.td
+++ b/lib/Target/R600/SIInstructions.td
@@ -1230,8 +1230,10 @@ defm V_CVT_F16_F32 : VOP1Inst <vop1<0xa>, "v_cvt_f16_f32",
 defm V_CVT_F32_F16 : VOP1Inst <vop1<0xb>, "v_cvt_f32_f16",
   VOP_F32_I32, f16_to_fp
 >;
-//defm V_CVT_RPI_I32_F32 : VOP1_32 <0x0000000c, "v_cvt_rpi_i32_f32", []>;
-//defm V_CVT_FLR_I32_F32 : VOP1_32 <0x0000000d, "v_cvt_flr_i32_f32", []>;
+defm V_CVT_RPI_I32_F32 : VOP1Inst <vop1<0xc>, "v_cvt_rpi_i32_f32",
+  VOP_I32_F32, cvt_rpi_i32_f32>;
+defm V_CVT_FLR_I32_F32 : VOP1Inst <vop1<0xd>, "v_cvt_flr_i32_f32",
+  VOP_I32_F32, cvt_flr_i32_f32>;
 //defm V_CVT_OFF_F32_I4 : VOP1_32 <0x0000000e, "v_cvt_off_f32_i4", []>;
 defm V_CVT_F32_F64 : VOP1Inst <vop1<0xf>, "v_cvt_f32_f64",
   VOP_F32_F64, fround

diff --git a/test/CodeGen/R600/cvt_flr_i32_f32.ll b/test/CodeGen/R600/cvt_flr_i32_f32.ll
new file mode 100644 (file)
index 0000000..b997a5f
--- /dev/null
+++ b/test/CodeGen/R600/cvt_flr_i32_f32.ll
@@ -0,0 +1,79 @@
+; RUN: llc -march=amdgcn -mcpu=SI < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
+
+declare float @llvm.fabs.f32(float) #1
+declare float @llvm.floor.f32(float) #1
+
+; FUNC-LABEL: {{^}}cvt_flr_i32_f32_0:
+; SI-NOT: add
+; SI: v_cvt_flr_i32_f32_e32 v{{[0-9]+}}, s{{[0-9]+}}
+; SI: s_endpgm
+define void @cvt_flr_i32_f32_0(i32 addrspace(1)* %out, float %x) #0 {
+  %floor = call float @llvm.floor.f32(float %x) #1
+  %cvt = fptosi float %floor to i32
+  store i32 %cvt, i32 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: {{^}}cvt_flr_i32_f32_1:
+; SI: v_add_f32_e64 [[TMP:v[0-9]+]], 1.0, s{{[0-9]+}}
+; SI: v_cvt_flr_i32_f32_e32 v{{[0-9]+}}, [[TMP]]
+; SI: s_endpgm
+define void @cvt_flr_i32_f32_1(i32 addrspace(1)* %out, float %x) #0 {
+  %fadd = fadd float %x, 1.0
+  %floor = call float @llvm.floor.f32(float %fadd) #1
+  %cvt = fptosi float %floor to i32
+  store i32 %cvt, i32 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: {{^}}cvt_flr_i32_f32_fabs:
+; SI-NOT: add
+; SI: v_cvt_flr_i32_f32_e64 v{{[0-9]+}}, |s{{[0-9]+}}|
+; SI: s_endpgm
+define void @cvt_flr_i32_f32_fabs(i32 addrspace(1)* %out, float %x) #0 {
+  %x.fabs = call float @llvm.fabs.f32(float %x) #1
+  %floor = call float @llvm.floor.f32(float %x.fabs) #1
+  %cvt = fptosi float %floor to i32
+  store i32 %cvt, i32 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: {{^}}cvt_flr_i32_f32_fneg:
+; SI-NOT: add
+; SI: v_cvt_flr_i32_f32_e64 v{{[0-9]+}}, -s{{[0-9]+}}
+; SI: s_endpgm
+define void @cvt_flr_i32_f32_fneg(i32 addrspace(1)* %out, float %x) #0 {
+  %x.fneg = fsub float -0.000000e+00, %x
+  %floor = call float @llvm.floor.f32(float %x.fneg) #1
+  %cvt = fptosi float %floor to i32
+  store i32 %cvt, i32 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: {{^}}cvt_flr_i32_f32_fabs_fneg:
+; SI-NOT: add
+; SI: v_cvt_flr_i32_f32_e64 v{{[0-9]+}}, -|s{{[0-9]+}}|
+; SI: s_endpgm
+define void @cvt_flr_i32_f32_fabs_fneg(i32 addrspace(1)* %out, float %x) #0 {
+  %x.fabs = call float @llvm.fabs.f32(float %x) #1
+  %x.fabs.fneg = fsub float -0.000000e+00, %x.fabs
+  %floor = call float @llvm.floor.f32(float %x.fabs.fneg) #1
+  %cvt = fptosi float %floor to i32
+  store i32 %cvt, i32 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: {{^}}no_cvt_flr_i32_f32_0:
+; SI-NOT: v_cvt_flr_i32_f32
+; SI: v_floor_f32
+; SI: v_cvt_u32_f32_e32
+; SI: s_endpgm
+define void @no_cvt_flr_i32_f32_0(i32 addrspace(1)* %out, float %x) #0 {
+  %floor = call float @llvm.floor.f32(float %x) #1
+  %cvt = fptoui float %floor to i32
+  store i32 %cvt, i32 addrspace(1)* %out
+  ret void
+}
+
+attributes #0 = { nounwind }
+attributes #1 = { nounwind readnone }

diff --git a/test/CodeGen/R600/cvt_rpi_i32_f32.ll b/test/CodeGen/R600/cvt_rpi_i32_f32.ll
new file mode 100644 (file)
index 0000000..c273ae9
--- /dev/null
+++ b/test/CodeGen/R600/cvt_rpi_i32_f32.ll
@@ -0,0 +1,76 @@
+; RUN: llc -march=amdgcn -mcpu=SI < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
+
+declare float @llvm.fabs.f32(float) #1
+declare float @llvm.floor.f32(float) #1
+
+; FUNC-LABEL: {{^}}cvt_rpi_i32_f32:
+; SI: v_cvt_rpi_i32_f32_e32 v{{[0-9]+}}, s{{[0-9]+}}
+; SI: s_endpgm
+define void @cvt_rpi_i32_f32(i32 addrspace(1)* %out, float %x) #0 {
+  %fadd = fadd float %x, 0.5
+  %floor = call float @llvm.floor.f32(float %fadd) #1
+  %cvt = fptosi float %floor to i32
+  store i32 %cvt, i32 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: {{^}}cvt_rpi_i32_f32_fabs:
+; SI: v_cvt_rpi_i32_f32_e64 v{{[0-9]+}}, |s{{[0-9]+}}|{{$}}
+; SI: s_endpgm
+define void @cvt_rpi_i32_f32_fabs(i32 addrspace(1)* %out, float %x) #0 {
+  %x.fabs = call float @llvm.fabs.f32(float %x) #1
+  %fadd = fadd float %x.fabs, 0.5
+  %floor = call float @llvm.floor.f32(float %fadd) #1
+  %cvt = fptosi float %floor to i32
+  store i32 %cvt, i32 addrspace(1)* %out
+  ret void
+}
+
+; FIXME: This doesn't work because it forms fsub 0.5, x
+; FUNC-LABEL: {{^}}cvt_rpi_i32_f32_fneg:
+; XSI: v_cvt_rpi_i32_f32_e64 v{{[0-9]+}}, -s{{[0-9]+}}
+; SI: v_sub_f32_e64 [[TMP:v[0-9]+]], 0.5, s{{[0-9]+}}
+; SI: v_cvt_flr_i32_f32_e32 {{v[0-9]+}}, [[TMP]]
+; SI: s_endpgm
+define void @cvt_rpi_i32_f32_fneg(i32 addrspace(1)* %out, float %x) #0 {
+  %x.fneg = fsub float -0.000000e+00, %x
+  %fadd = fadd float %x.fneg, 0.5
+  %floor = call float @llvm.floor.f32(float %fadd) #1
+  %cvt = fptosi float %floor to i32
+  store i32 %cvt, i32 addrspace(1)* %out
+  ret void
+}
+
+; FIXME: This doesn't work for same reason as above
+; FUNC-LABEL: {{^}}cvt_rpi_i32_f32_fabs_fneg:
+; XSI: v_cvt_rpi_i32_f32_e64 v{{[0-9]+}}, -|s{{[0-9]+}}|
+
+; SI: v_sub_f32_e64 [[TMP:v[0-9]+]], 0.5, |s{{[0-9]+}}|
+; SI: v_cvt_flr_i32_f32_e32 {{v[0-9]+}}, [[TMP]]
+; SI: s_endpgm
+define void @cvt_rpi_i32_f32_fabs_fneg(i32 addrspace(1)* %out, float %x) #0 {
+  %x.fabs = call float @llvm.fabs.f32(float %x) #1
+  %x.fabs.fneg = fsub float -0.000000e+00, %x.fabs
+  %fadd = fadd float %x.fabs.fneg, 0.5
+  %floor = call float @llvm.floor.f32(float %fadd) #1
+  %cvt = fptosi float %floor to i32
+  store i32 %cvt, i32 addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: {{^}}no_cvt_rpi_i32_f32_0:
+; SI-NOT: v_cvt_rpi_i32_f32
+; SI: v_add_f32
+; SI: v_floor_f32
+; SI: v_cvt_u32_f32
+; SI: s_endpgm
+define void @no_cvt_rpi_i32_f32_0(i32 addrspace(1)* %out, float %x) #0 {
+  %fadd = fadd float %x, 0.5
+  %floor = call float @llvm.floor.f32(float %fadd) #1
+  %cvt = fptoui float %floor to i32
+  store i32 %cvt, i32 addrspace(1)* %out
+  ret void
+}
+
+attributes #0 = { nounwind }
+attributes #1 = { nounwind readnone }