AMDGPU/GlobalISel: Fix import of integer med3

This isn't too useful now, since nothing is currently trying to form
min/max from cmp+select.

2 files changed, 30 insertions, 32 deletions

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td b/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td
index 6541470f06a..7e71dbdd124 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td
@@ -737,30 +737,6 @@ class ROTRPattern <Instruction BIT_ALIGN> : AMDGPUPat <
   (BIT_ALIGN $src0, $src0, $src1)
 >;
 
-multiclass IntMed3Pat<Instruction med3Inst,
-                 SDPatternOperator min,
-                 SDPatternOperator max,
-                 SDPatternOperator min_oneuse,
-                 SDPatternOperator max_oneuse,
-                 ValueType vt = i32> {
-
-  // This matches 16 permutations of
-  // min(max(a, b), max(min(a, b), c))
-  def : AMDGPUPat <
-  (min (max_oneuse vt:$src0, vt:$src1),
-       (max_oneuse (min_oneuse vt:$src0, vt:$src1), vt:$src2)),
-  (med3Inst vt:$src0, vt:$src1, vt:$src2)
->;
-
-  // This matches 16 permutations of
-  // max(min(x, y), min(max(x, y), z))
-  def : AMDGPUPat <
-  (max (min_oneuse vt:$src0, vt:$src1),
-       (min_oneuse (max_oneuse vt:$src0, vt:$src1), vt:$src2)),
-  (med3Inst $src0, $src1, $src2)
->;
-}
-
 // Special conversion patterns
 
 def cvt_rpi_i32_f32 : PatFrag <
diff --git a/llvm/lib/Target/AMDGPU/SIInstructions.td b/llvm/lib/Target/AMDGPU/SIInstructions.td
index 4c8197975ce..d84720f820e 100644
--- a/llvm/lib/Target/AMDGPU/SIInstructions.td
+++ b/llvm/lib/Target/AMDGPU/SIInstructions.td
@@ -1972,6 +1972,29 @@ defm : BFMPatterns <i32, S_BFM_B32, S_MOV_B32>;
 defm : BFEPattern <V_BFE_U32, V_BFE_I32, S_MOV_B32>;
 defm : SHA256MaPattern <V_BFI_B32, V_XOR_B32_e64, SReg_64>;
 
+multiclass IntMed3Pat<Instruction med3Inst,
+                 SDPatternOperator min,
+                 SDPatternOperator max,
+                 SDPatternOperator min_oneuse,
+                 SDPatternOperator max_oneuse> {
+
+  // This matches 16 permutations of
+  // min(max(a, b), max(min(a, b), c))
+  def : AMDGPUPat <
+  (min (max_oneuse i32:$src0, i32:$src1),
+       (max_oneuse (min_oneuse i32:$src0, i32:$src1), i32:$src2)),
+  (med3Inst VSrc_b32:$src0, VSrc_b32:$src1, VSrc_b32:$src2)
+>;
+
+  // This matches 16 permutations of
+  // max(min(x, y), min(max(x, y), z))
+  def : AMDGPUPat <
+  (max (min_oneuse i32:$src0, i32:$src1),
+       (min_oneuse (max_oneuse i32:$src0, i32:$src1), i32:$src2)),
+  (med3Inst VSrc_b32:$src0, VSrc_b32:$src1, VSrc_b32:$src2)
+>;
+}
+
 defm : IntMed3Pat<V_MED3_I32, smin, smax, smin_oneuse, smax_oneuse>;
 defm : IntMed3Pat<V_MED3_U32, umin, umax, umin_oneuse, umax_oneuse>;
 
@@ -2002,22 +2025,21 @@ multiclass Int16Med3Pat<Instruction med3Inst,
                    SDPatternOperator min,
                    SDPatternOperator max,
                    SDPatternOperator max_oneuse,
-                   SDPatternOperator min_oneuse,
-                   ValueType vt = i16> {
+                   SDPatternOperator min_oneuse> {
   // This matches 16 permutations of
   // max(min(x, y), min(max(x, y), z))
   def : GCNPat <
-  (max (min_oneuse vt:$src0, vt:$src1),
-       (min_oneuse (max_oneuse vt:$src0, vt:$src1), vt:$src2)),
-  (med3Inst SRCMODS.NONE, $src0, SRCMODS.NONE, $src1, SRCMODS.NONE, $src2, DSTCLAMP.NONE)
+  (max (min_oneuse i16:$src0, i16:$src1),
+       (min_oneuse (max_oneuse i16:$src0, i16:$src1), i16:$src2)),
+  (med3Inst SRCMODS.NONE, VSrc_b16:$src0, SRCMODS.NONE, VSrc_b16:$src1, SRCMODS.NONE, VSrc_b16:$src2, DSTCLAMP.NONE)
 >;
 
   // This matches 16 permutations of
   // min(max(a, b), max(min(a, b), c))
   def : GCNPat <
-  (min (max_oneuse vt:$src0, vt:$src1),
-      (max_oneuse (min_oneuse vt:$src0, vt:$src1), vt:$src2)),
-  (med3Inst SRCMODS.NONE, $src0, SRCMODS.NONE, $src1, SRCMODS.NONE, $src2, DSTCLAMP.NONE)
+  (min (max_oneuse i16:$src0, i16:$src1),
+      (max_oneuse (min_oneuse i16:$src0, i16:$src1), i16:$src2)),
+  (med3Inst SRCMODS.NONE, VSrc_b16:$src0, SRCMODS.NONE, VSrc_b16:$src1, SRCMODS.NONE, VSrc_b16:$src2, DSTCLAMP.NONE)
 >;
 }