[PATCH] D98872: [amdgpu] Update med3 combine to skip i64

[Jon Chesterfield via Phabricator via llvm-commits]

JonChesterfield created this revision.
JonChesterfield added reviewers: arsenm, kzhuravl, rampitec, tstellar, tpr.
Herald added subscribers: kerbowa, hiraditya, t-tye, dstuttard, yaxunl, nhaehnle, jvesely.
JonChesterfield requested review of this revision.
Herald added subscribers: llvm-commits, sstefan1, wdng.
Herald added a reviewer: jdoerfert.
Herald added a project: LLVM.

[amdgpu] Update med3 combine to skip i64

Fixes an assumption that a type which is not i32 will be i16. This asserts
when trying to sign/zero extend an i64 to i32.

Test case was cut down from an openmp application. Variations on it are hit by
other combines before reaching the problematic one, e.g. replacing the
immediate values with other function arguments changes the codegen path and
misses this combine.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D98872

Files:
  llvm/lib/Target/AMDGPU/SIISelLowering.cpp
  llvm/test/CodeGen/AMDGPU/smed3.ll


Index: llvm/test/CodeGen/AMDGPU/smed3.ll
===================================================================
--- llvm/test/CodeGen/AMDGPU/smed3.ll
+++ llvm/test/CodeGen/AMDGPU/smed3.ll
@@ -80,6 +80,19 @@
   ret void
 }
 
+; Regression test for performIntMed3ImmCombine extending arguments to 32 bit
+; which failed for 64 bit arguments. Previously asserted / crashed.
+; GCN-LABEL: {{^}}test_intMed3ImmCombine_no_32bit_extend:
+; GCN: v_cmp_lt_i64
+; GCN: v_cmp_gt_i64
+define i64 @test_intMed3ImmCombine_no_32bit_extend(i64 %x) {
+  %smax = call i64 @llvm.smax.i64(i64 %x, i64 -2)
+  %smin = call i64 @llvm.smin.i64(i64 %smax, i64 2)
+  ret i64 %smin
+}
+declare i64 @llvm.smax.i64(i64, i64)
+declare i64 @llvm.smin.i64(i64, i64)
+
 ; GCN-LABEL: {{^}}v_test_smed3_r_i_i_i16:
 ; SICIVI: v_med3_i32 v{{[0-9]+}}, v{{[0-9]+}}, 12, 17
 ; GFX9: v_med3_i16 v{{[0-9]+}}, v{{[0-9]+}}, 12, 17
Index: llvm/lib/Target/AMDGPU/SIISelLowering.cpp
===================================================================
--- llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -9772,15 +9772,19 @@
   }
 
   // If there isn't a 16-bit med3 operation, convert to 32-bit.
-  MVT NVT = MVT::i32;
-  unsigned ExtOp = Signed ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
+  if (VT == MVT::i16) {
+    MVT NVT = MVT::i32;
+    unsigned ExtOp = Signed ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
 
-  SDValue Tmp1 = DAG.getNode(ExtOp, SL, NVT, Op0->getOperand(0));
-  SDValue Tmp2 = DAG.getNode(ExtOp, SL, NVT, Op0->getOperand(1));
-  SDValue Tmp3 = DAG.getNode(ExtOp, SL, NVT, Op1);
+    SDValue Tmp1 = DAG.getNode(ExtOp, SL, NVT, Op0->getOperand(0));
+    SDValue Tmp2 = DAG.getNode(ExtOp, SL, NVT, Op0->getOperand(1));
+    SDValue Tmp3 = DAG.getNode(ExtOp, SL, NVT, Op1);
 
-  SDValue Med3 = DAG.getNode(Med3Opc, SL, NVT, Tmp1, Tmp2, Tmp3);
-  return DAG.getNode(ISD::TRUNCATE, SL, VT, Med3);
+    SDValue Med3 = DAG.getNode(Med3Opc, SL, NVT, Tmp1, Tmp2, Tmp3);
+    return DAG.getNode(ISD::TRUNCATE, SL, VT, Med3);
+  }
+
+  return SDValue();
 }
 
 static ConstantFPSDNode *getSplatConstantFP(SDValue Op) {


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D98872.331572.patch
Type: text/x-patch
Size: 2112 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20210318/bb1f0228/attachment.bin>