[Mesa-dev] [PATCH v3] R600/SI: Add pattern for AMDGPUurecip

[Michel Dänzer]

From: Michel Dänzer <michel.daenzer at amd.com>

21 more little piglits with radeonsi.

Signed-off-by: Michel Dänzer <michel.daenzer at amd.com>
---

v3: Use constant for and add comments about scaling multiplications

 lib/Target/R600/AMDGPUInstructions.td |  1 +
 lib/Target/R600/R600Instructions.td   |  3 ++-
 lib/Target/R600/SIInstructions.td     | 12 ++++++++++--
 test/CodeGen/R600/urecip.ll           | 12 ++++++++++++
 4 files changed, 25 insertions(+), 3 deletions(-)
 create mode 100644 test/CodeGen/R600/urecip.ll

diff --git a/lib/Target/R600/AMDGPUInstructions.td b/lib/Target/R600/AMDGPUInstructions.td
index e740348..fa890c1 100644
--- a/lib/Target/R600/AMDGPUInstructions.td
+++ b/lib/Target/R600/AMDGPUInstructions.td
@@ -94,6 +94,7 @@ class Constants {
 int TWO_PI = 0x40c90fdb;
 int PI = 0x40490fdb;
 int TWO_PI_INV = 0x3e22f983;
+int FP_UINT_MAX_PLUS_1 = 0x4f800000;	// 1 << 32 in floating point encoding
 }
 def CONST : Constants;
 
diff --git a/lib/Target/R600/R600Instructions.td b/lib/Target/R600/R600Instructions.td
index b4c45e1..8ede6cc 100644
--- a/lib/Target/R600/R600Instructions.td
+++ b/lib/Target/R600/R600Instructions.td
@@ -1923,10 +1923,11 @@ def : COS_PAT <COS_cm>;
 defm DIV_cm : DIV_Common<RECIP_IEEE_cm>;
 
 // RECIP_UINT emulation for Cayman
+// The multiplication scales from [0,1] to the unsigned integer range
 def : Pat <
   (AMDGPUurecip R600_Reg32:$src0),
   (FLT_TO_UINT_eg (MUL_IEEE (RECIP_IEEE_cm (UINT_TO_FLT_eg R600_Reg32:$src0)),
-                            (MOV_IMM_I32 0x4f800000)))
+                            (MOV_IMM_I32 CONST.FP_UINT_MAX_PLUS_1)))
 >;
 
 
diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
index e2a08fc..0226d5a 100644
--- a/lib/Target/R600/SIInstructions.td
+++ b/lib/Target/R600/SIInstructions.td
@@ -602,8 +602,8 @@ defm V_READFIRSTLANE_B32 : VOP1_32 <0x00000002, "V_READFIRSTLANE_B32", []>;
 defm V_CVT_F32_I32 : VOP1_32 <0x00000005, "V_CVT_F32_I32",
   [(set VReg_32:$dst, (sint_to_fp VSrc_32:$src0))]
 >;
-//defm V_CVT_F32_U32 : VOP1_32 <0x00000006, "V_CVT_F32_U32", []>;
-//defm V_CVT_U32_F32 : VOP1_32 <0x00000007, "V_CVT_U32_F32", []>;
+defm V_CVT_F32_U32 : VOP1_32 <0x00000006, "V_CVT_F32_U32", []>;
+defm V_CVT_U32_F32 : VOP1_32 <0x00000007, "V_CVT_U32_F32", []>;
 defm V_CVT_I32_F32 : VOP1_32 <0x00000008, "V_CVT_I32_F32",
   [(set (i32 VReg_32:$dst), (fp_to_sint VSrc_32:$src0))]
 >;
@@ -1514,6 +1514,14 @@ def : Pat <
   (BUFFER_LOAD_DWORD 0, 1, 0, 0, 0, 0, VReg_32:$voff, SReg_128:$sbase, 0, 0, 0)
 >;
 
+// The multiplication scales from [0,1] to the unsigned integer range
+def : Pat <
+  (AMDGPUurecip VSrc_32:$src0),
+  (V_CVT_U32_F32_e32
+    (V_MUL_F32_e32 CONST.FP_UINT_MAX_PLUS_1,
+                   (V_RCP_IFLAG_F32_e32 (V_CVT_F32_U32_e32 VSrc_32:$src0))))
+>;
+
 /********** ================== **********/
 /**********   VOP3 Patterns    **********/
 /********** ================== **********/
diff --git a/test/CodeGen/R600/urecip.ll b/test/CodeGen/R600/urecip.ll
new file mode 100644
index 0000000..dad02dd
--- /dev/null
+++ b/test/CodeGen/R600/urecip.ll
@@ -0,0 +1,12 @@
+;RUN: llc < %s -march=r600 -mcpu=verde | FileCheck %s
+
+;CHECK: V_RCP_IFLAG_F32_e32
+
+define void @test(i32 %p, i32 %q) {
+   %i = udiv i32 %p, %q
+   %r = bitcast i32 %i to float
+   call void @llvm.SI.export(i32 15, i32 0, i32 1, i32 12, i32 0, float %r, float %r, float %r, float %r)
+   ret void
+}
+
+declare void @llvm.SI.export(i32, i32, i32, i32, i32, float, float, float, float)
-- 
1.8.2