AMDGPU: Select BFI patterns with 64-bit ints

llvm-svn: 324431

3 files changed, 46 insertions, 6 deletions

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td b/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td
index 76d35469027..4f28d6fa430 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td
@@ -570,6 +570,18 @@ multiclass BFIPatterns <Instruction BFI_INT,
     (BFI_INT $x, $y, $z)
   >;
 
+  // 64-bit version
+  def : AMDGPUPat <
+    (or (and i64:$y, i64:$x), (and i64:$z, (not i64:$x))),
+    (REG_SEQUENCE RC64,
+      (BFI_INT (i32 (EXTRACT_SUBREG $x, sub0)),
+               (i32 (EXTRACT_SUBREG $y, sub0)),
+               (i32 (EXTRACT_SUBREG $z, sub0))), sub0,
+      (BFI_INT (i32 (EXTRACT_SUBREG $x, sub1)),
+               (i32 (EXTRACT_SUBREG $y, sub1)),
+               (i32 (EXTRACT_SUBREG $z, sub1))), sub1)
+  >;
+
   // SHA-256 Ch function
   // z ^ (x & (y ^ z))
   def : AMDGPUPat <
@@ -577,6 +589,18 @@ multiclass BFIPatterns <Instruction BFI_INT,
     (BFI_INT $x, $y, $z)
   >;
 
+  // 64-bit version
+  def : AMDGPUPat <
+    (xor i64:$z, (and i64:$x, (xor i64:$y, i64:$z))),
+    (REG_SEQUENCE RC64,
+      (BFI_INT (i32 (EXTRACT_SUBREG $x, sub0)),
+               (i32 (EXTRACT_SUBREG $y, sub0)),
+               (i32 (EXTRACT_SUBREG $z, sub0))), sub0,
+      (BFI_INT (i32 (EXTRACT_SUBREG $x, sub1)),
+               (i32 (EXTRACT_SUBREG $y, sub1)),
+               (i32 (EXTRACT_SUBREG $z, sub1))), sub1)
+  >;
+
   def : AMDGPUPat <
     (fcopysign f32:$src0, f32:$src1),
     (BFI_INT (LoadImm32 (i32 0x7fffffff)), $src0, $src1)
@@ -610,10 +634,25 @@ multiclass BFIPatterns <Instruction BFI_INT,
 // SHA-256 Ma patterns
 
 // ((x & z) | (y & (x | z))) -> BFI_INT (XOR x, y), z, y
-class SHA256MaPattern <Instruction BFI_INT, Instruction XOR> : AMDGPUPat <
-  (or (and i32:$x, i32:$z), (and i32:$y, (or i32:$x, i32:$z))),
-  (BFI_INT (XOR i32:$x, i32:$y), i32:$z, i32:$y)
->;
+multiclass SHA256MaPattern <Instruction BFI_INT, Instruction XOR, RegisterClass RC64> {
+  def : AMDGPUPat <
+    (or (and i32:$x, i32:$z), (and i32:$y, (or i32:$x, i32:$z))),
+    (BFI_INT (XOR i32:$x, i32:$y), i32:$z, i32:$y)
+  >;
+
+  def : AMDGPUPat <
+    (or (and i64:$x, i64:$z), (and i64:$y, (or i64:$x, i64:$z))),
+    (REG_SEQUENCE RC64,
+      (BFI_INT (XOR (i32 (EXTRACT_SUBREG $x, sub0)),
+                    (i32 (EXTRACT_SUBREG $y, sub0))),
+               (i32 (EXTRACT_SUBREG $z, sub0)),
+               (i32 (EXTRACT_SUBREG $y, sub0))), sub0,
+      (BFI_INT (XOR (i32 (EXTRACT_SUBREG $x, sub1)),
+                    (i32 (EXTRACT_SUBREG $y, sub1))),
+               (i32 (EXTRACT_SUBREG $z, sub1)),
+               (i32 (EXTRACT_SUBREG $y, sub1))), sub1)
+  >;
+}
 
 // Bitfield extract patterns
 
diff --git a/llvm/lib/Target/AMDGPU/EvergreenInstructions.td b/llvm/lib/Target/AMDGPU/EvergreenInstructions.td
index 5e26f97b0c8..148b45ba6bb 100644
--- a/llvm/lib/Target/AMDGPU/EvergreenInstructions.td
+++ b/llvm/lib/Target/AMDGPU/EvergreenInstructions.td
@@ -693,7 +693,7 @@ def : EGOrCaymanPat<(fp_to_sint f32:$src0), (FLT_TO_INT_eg (TRUNC $src0))>;
 def : EGOrCaymanPat<(fp_to_uint f32:$src0), (FLT_TO_UINT_eg (TRUNC $src0))>;
 
 // SHA-256 Patterns
-def : SHA256MaPattern <BFI_INT_eg, XOR_INT>;
+defm : SHA256MaPattern <BFI_INT_eg, XOR_INT, R600_Reg64>;
 
 def EG_ExportSwz : ExportSwzInst {
   let Word1{19-16} = 0; // BURST_COUNT
diff --git a/llvm/lib/Target/AMDGPU/SIInstructions.td b/llvm/lib/Target/AMDGPU/SIInstructions.td
index f87b4f4fda6..2a77bfed8dc 100644
--- a/llvm/lib/Target/AMDGPU/SIInstructions.td
+++ b/llvm/lib/Target/AMDGPU/SIInstructions.td
@@ -1098,6 +1098,7 @@ let SubtargetPredicate = isGCN in {
 def : IMad24Pat<V_MAD_I32_I24, 1>;
 def : UMad24Pat<V_MAD_U32_U24, 1>;
 
+// FIXME: This should only be done for VALU inputs
 defm : BFIPatterns <V_BFI_B32, S_MOV_B32, SReg_64>;
 def : ROTRPattern <V_ALIGNBIT_B32>;
 
@@ -1487,7 +1488,7 @@ defm : BFMPatterns <i32, S_BFM_B32, S_MOV_B32>;
 // FIXME: defm : BFMPatterns <i64, S_BFM_B64, S_MOV_B64>;
 
 defm : BFEPattern <V_BFE_U32, V_BFE_I32, S_MOV_B32>;
-def : SHA256MaPattern <V_BFI_B32, V_XOR_B32_e64>;
+defm : SHA256MaPattern <V_BFI_B32, V_XOR_B32_e64, SReg_64>;
 
 def : IntMed3Pat<V_MED3_I32, smax, smax_oneuse, smin_oneuse>;
 def : IntMed3Pat<V_MED3_U32, umax, umax_oneuse, umin_oneuse>;