Skip to content

[GlobalISel][AMDGPU] Fix handling of v2i128 type for AND, OR, XOR #138574

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Merged
merged 5 commits into from
May 8, 2025
Merged

[GlobalISel][AMDGPU] Fix handling of v2i128 type for AND, OR, XOR #138574

merged 5 commits into from
May 8, 2025

Conversation

chinmaydd
Copy link
Contributor

Current behavior crashes the compiler.

This bug was found using the AMDGPU Fuzzing project.

Fixes SWDEV-508816.

@chinmaydd chinmaydd requested review from arsenm and shiltian May 5, 2025 20:16
@llvmbot
Copy link
Member

llvmbot commented May 5, 2025 •
edited
Loading

@llvm/pr-subscribers-backend-amdgpu

@llvm/pr-subscribers-llvm-globalisel

Author: Chinmay Deshpande (chinmaydd)

Changes

Current behavior crashes the compiler.

This bug was found using the AMDGPU Fuzzing project.

Fixes SWDEV-508816.

Full diff: https://github.com/llvm/llvm-project/pull/138574.diff

4 Files Affected:

  • (modified) llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp (+14-1)
  • (added) llvm/test/CodeGen/AMDGPU/GlobalISel/and.v2i128.ll (+117)
  • (added) llvm/test/CodeGen/AMDGPU/GlobalISel/or.v2i128.ll (+117)
  • (added) llvm/test/CodeGen/AMDGPU/GlobalISel/xor.v2i128.ll (+117)
diff --git a/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
index ff8658ed82a72..e8063d54ac65a 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
@@ -119,6 +119,18 @@ static LegalizeMutation fewerEltsToSize64Vector(unsigned TypeIdx) {
   };
 }
 
+static LegalizeMutation breakCurrentEltsToSize32Or64(unsigned TypeIdx) {
+  return [=](const LegalityQuery &Query) {
+    const LLT Ty = Query.Types[TypeIdx];
+    const LLT EltTy = Ty.getElementType();
+    const int Size = Ty.getSizeInBits();
+    const int EltSize = EltTy.getSizeInBits();
+    const unsigned TargetEltSize = EltSize % 64 == 0 ? 64 : 32;
+    const unsigned NewNumElts = (Size + (TargetEltSize - 1)) / TargetEltSize;
+    return std::pair(TypeIdx, LLT::fixed_vector(NewNumElts, TargetEltSize));
+  };
+}
+
 // Increase the number of vector elements to reach the next multiple of 32-bit
 // type.
 static LegalizeMutation moreEltsToNext32Bit(unsigned TypeIdx) {
@@ -875,7 +887,8 @@ AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
     .legalFor({S32, S1, S64, V2S32, S16, V2S16, V4S16})
     .clampScalar(0, S32, S64)
     .moreElementsIf(isSmallOddVector(0), oneMoreElement(0))
-    .fewerElementsIf(vectorWiderThan(0, 64), fewerEltsToSize64Vector(0))
+    .fewerElementsIf(all(vectorWiderThan(0, 64), scalarOrEltNarrowerThan(0, 64)), fewerEltsToSize64Vector(0))
+    .bitcastIf(all(vectorWiderThan(0, 64), scalarOrEltWiderThan(0, 64)), breakCurrentEltsToSize32Or64(0))
     .widenScalarToNextPow2(0)
     .scalarize(0);
 
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/and.v2i128.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/and.v2i128.ll
new file mode 100644
index 0000000000000..532a797094d14
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/and.v2i128.ll
@@ -0,0 +1,117 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -global-isel=true -mtriple=amdgcn -mcpu=hawaii < %s | FileCheck -check-prefix=GFX7 %s
+; RUN: llc -global-isel=true -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefix=GFX9 %s
+; RUN: llc -global-isel=true -mtriple=amdgcn -mcpu=fiji < %s | FileCheck -check-prefix=GFX8 %s
+; RUN: llc -global-isel=true -mtriple=amdgcn -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s
+; RUN: llc -global-isel=true -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck -check-prefix=GFX10 %s
+
+define <2 x i128> @v_and_v2i128(<2 x i128> %a, <2 x i128> %b) {
+; GFX7-LABEL: v_and_v2i128:
+; GFX7:       ; %bb.0:
+; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    v_and_b32_e32 v0, v0, v8
+; GFX7-NEXT:    v_and_b32_e32 v1, v1, v9
+; GFX7-NEXT:    v_and_b32_e32 v2, v2, v10
+; GFX7-NEXT:    v_and_b32_e32 v3, v3, v11
+; GFX7-NEXT:    v_and_b32_e32 v4, v4, v12
+; GFX7-NEXT:    v_and_b32_e32 v5, v5, v13
+; GFX7-NEXT:    v_and_b32_e32 v6, v6, v14
+; GFX7-NEXT:    v_and_b32_e32 v7, v7, v15
+; GFX7-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: v_and_v2i128:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_and_b32_e32 v0, v0, v8
+; GFX9-NEXT:    v_and_b32_e32 v1, v1, v9
+; GFX9-NEXT:    v_and_b32_e32 v2, v2, v10
+; GFX9-NEXT:    v_and_b32_e32 v3, v3, v11
+; GFX9-NEXT:    v_and_b32_e32 v4, v4, v12
+; GFX9-NEXT:    v_and_b32_e32 v5, v5, v13
+; GFX9-NEXT:    v_and_b32_e32 v6, v6, v14
+; GFX9-NEXT:    v_and_b32_e32 v7, v7, v15
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: v_and_v2i128:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    v_and_b32_e32 v0, v0, v8
+; GFX8-NEXT:    v_and_b32_e32 v1, v1, v9
+; GFX8-NEXT:    v_and_b32_e32 v2, v2, v10
+; GFX8-NEXT:    v_and_b32_e32 v3, v3, v11
+; GFX8-NEXT:    v_and_b32_e32 v4, v4, v12
+; GFX8-NEXT:    v_and_b32_e32 v5, v5, v13
+; GFX8-NEXT:    v_and_b32_e32 v6, v6, v14
+; GFX8-NEXT:    v_and_b32_e32 v7, v7, v15
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: v_and_v2i128:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    v_and_b32_e32 v0, v0, v8
+; GFX10-NEXT:    v_and_b32_e32 v1, v1, v9
+; GFX10-NEXT:    v_and_b32_e32 v2, v2, v10
+; GFX10-NEXT:    v_and_b32_e32 v3, v3, v11
+; GFX10-NEXT:    v_and_b32_e32 v4, v4, v12
+; GFX10-NEXT:    v_and_b32_e32 v5, v5, v13
+; GFX10-NEXT:    v_and_b32_e32 v6, v6, v14
+; GFX10-NEXT:    v_and_b32_e32 v7, v7, v15
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %and = and <2 x i128> %a, %b
+  ret <2 x i128> %and
+}
+
+define <2 x i128> @v_and_v2i128_inline_imm(<2 x i128> %a) {
+; GFX7-LABEL: v_and_v2i128_inline_imm:
+; GFX7:       ; %bb.0:
+; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b64 s[4:5], 64
+; GFX7-NEXT:    s_mov_b64 s[6:7], 0
+; GFX7-NEXT:    s_mov_b64 s[4:5], s[4:5]
+; GFX7-NEXT:    s_mov_b64 s[6:7], s[6:7]
+; GFX7-NEXT:    v_and_b32_e32 v0, s4, v0
+; GFX7-NEXT:    v_and_b32_e32 v1, s5, v1
+; GFX7-NEXT:    v_and_b32_e32 v2, s6, v2
+; GFX7-NEXT:    v_and_b32_e32 v3, s7, v3
+; GFX7-NEXT:    v_and_b32_e32 v4, s4, v4
+; GFX7-NEXT:    v_and_b32_e32 v5, s5, v5
+; GFX7-NEXT:    v_and_b32_e32 v6, s6, v6
+; GFX7-NEXT:    v_and_b32_e32 v7, s7, v7
+; GFX7-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: v_and_v2i128_inline_imm:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_mov_b64 s[4:5], 64
+; GFX9-NEXT:    s_mov_b64 s[6:7], 0
+; GFX9-NEXT:    s_mov_b64 s[4:5], s[4:5]
+; GFX9-NEXT:    s_mov_b64 s[6:7], s[6:7]
+; GFX9-NEXT:    v_and_b32_e32 v0, s4, v0
+; GFX9-NEXT:    v_and_b32_e32 v1, s5, v1
+; GFX9-NEXT:    v_and_b32_e32 v2, s6, v2
+; GFX9-NEXT:    v_and_b32_e32 v3, s7, v3
+; GFX9-NEXT:    v_and_b32_e32 v4, s4, v4
+; GFX9-NEXT:    v_and_b32_e32 v5, s5, v5
+; GFX9-NEXT:    v_and_b32_e32 v6, s6, v6
+; GFX9-NEXT:    v_and_b32_e32 v7, s7, v7
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: v_and_v2i128_inline_imm:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    s_mov_b64 s[4:5], 64
+; GFX8-NEXT:    s_mov_b64 s[6:7], 0
+; GFX8-NEXT:    s_mov_b64 s[4:5], s[4:5]
+; GFX8-NEXT:    s_mov_b64 s[6:7], s[6:7]
+; GFX8-NEXT:    v_and_b32_e32 v0, s4, v0
+; GFX8-NEXT:    v_and_b32_e32 v1, s5, v1
+; GFX8-NEXT:    v_and_b32_e32 v2, s6, v2
+; GFX8-NEXT:    v_and_b32_e32 v3, s7, v3
+; GFX8-NEXT:    v_and_b32_e32 v4, s4, v4
+; GFX8-NEXT:    v_and_b32_e32 v5, s5, v5
+; GFX8-NEXT:    v_and_b32_e32 v6, s6, v6
+; GFX8-NEXT:    v_and_b32_e32 v7, s7, v7
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+  %and = and <2 x i128> %a, <i128 64, i128 64>
+  ret <2 x i128> %and
+}
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/or.v2i128.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/or.v2i128.ll
new file mode 100644
index 0000000000000..eaba0500dc1f3
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/or.v2i128.ll
@@ -0,0 +1,117 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -global-isel=true -mtriple=amdgcn -mcpu=hawaii < %s | FileCheck -check-prefix=GFX7 %s
+; RUN: llc -global-isel=true -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefix=GFX9 %s
+; RUN: llc -global-isel=true -mtriple=amdgcn -mcpu=fiji < %s | FileCheck -check-prefix=GFX8 %s
+; RUN: llc -global-isel=true -mtriple=amdgcn -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s
+; RUN: llc -global-isel=true -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck -check-prefix=GFX10 %s
+
+define <2 x i128> @v_or_v2i128(<2 x i128> %a, <2 x i128> %b) {
+; GFX7-LABEL: v_or_v2i128:
+; GFX7:       ; %bb.0:
+; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    v_or_b32_e32 v0, v0, v8
+; GFX7-NEXT:    v_or_b32_e32 v1, v1, v9
+; GFX7-NEXT:    v_or_b32_e32 v2, v2, v10
+; GFX7-NEXT:    v_or_b32_e32 v3, v3, v11
+; GFX7-NEXT:    v_or_b32_e32 v4, v4, v12
+; GFX7-NEXT:    v_or_b32_e32 v5, v5, v13
+; GFX7-NEXT:    v_or_b32_e32 v6, v6, v14
+; GFX7-NEXT:    v_or_b32_e32 v7, v7, v15
+; GFX7-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: v_or_v2i128:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_or_b32_e32 v0, v0, v8
+; GFX9-NEXT:    v_or_b32_e32 v1, v1, v9
+; GFX9-NEXT:    v_or_b32_e32 v2, v2, v10
+; GFX9-NEXT:    v_or_b32_e32 v3, v3, v11
+; GFX9-NEXT:    v_or_b32_e32 v4, v4, v12
+; GFX9-NEXT:    v_or_b32_e32 v5, v5, v13
+; GFX9-NEXT:    v_or_b32_e32 v6, v6, v14
+; GFX9-NEXT:    v_or_b32_e32 v7, v7, v15
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: v_or_v2i128:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    v_or_b32_e32 v0, v0, v8
+; GFX8-NEXT:    v_or_b32_e32 v1, v1, v9
+; GFX8-NEXT:    v_or_b32_e32 v2, v2, v10
+; GFX8-NEXT:    v_or_b32_e32 v3, v3, v11
+; GFX8-NEXT:    v_or_b32_e32 v4, v4, v12
+; GFX8-NEXT:    v_or_b32_e32 v5, v5, v13
+; GFX8-NEXT:    v_or_b32_e32 v6, v6, v14
+; GFX8-NEXT:    v_or_b32_e32 v7, v7, v15
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: v_or_v2i128:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    v_or_b32_e32 v0, v0, v8
+; GFX10-NEXT:    v_or_b32_e32 v1, v1, v9
+; GFX10-NEXT:    v_or_b32_e32 v2, v2, v10
+; GFX10-NEXT:    v_or_b32_e32 v3, v3, v11
+; GFX10-NEXT:    v_or_b32_e32 v4, v4, v12
+; GFX10-NEXT:    v_or_b32_e32 v5, v5, v13
+; GFX10-NEXT:    v_or_b32_e32 v6, v6, v14
+; GFX10-NEXT:    v_or_b32_e32 v7, v7, v15
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %or = or <2 x i128> %a, %b
+  ret <2 x i128> %or
+}
+
+define <2 x i128> @v_or_v2i128_inline_imm(<2 x i128> %a) {
+; GFX7-LABEL: v_or_v2i128_inline_imm:
+; GFX7:       ; %bb.0:
+; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b64 s[4:5], 64
+; GFX7-NEXT:    s_mov_b64 s[6:7], 0
+; GFX7-NEXT:    s_mov_b64 s[4:5], s[4:5]
+; GFX7-NEXT:    s_mov_b64 s[6:7], s[6:7]
+; GFX7-NEXT:    v_or_b32_e32 v0, s4, v0
+; GFX7-NEXT:    v_or_b32_e32 v1, s5, v1
+; GFX7-NEXT:    v_or_b32_e32 v2, s6, v2
+; GFX7-NEXT:    v_or_b32_e32 v3, s7, v3
+; GFX7-NEXT:    v_or_b32_e32 v4, s4, v4
+; GFX7-NEXT:    v_or_b32_e32 v5, s5, v5
+; GFX7-NEXT:    v_or_b32_e32 v6, s6, v6
+; GFX7-NEXT:    v_or_b32_e32 v7, s7, v7
+; GFX7-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: v_or_v2i128_inline_imm:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_mov_b64 s[4:5], 64
+; GFX9-NEXT:    s_mov_b64 s[6:7], 0
+; GFX9-NEXT:    s_mov_b64 s[4:5], s[4:5]
+; GFX9-NEXT:    s_mov_b64 s[6:7], s[6:7]
+; GFX9-NEXT:    v_or_b32_e32 v0, s4, v0
+; GFX9-NEXT:    v_or_b32_e32 v1, s5, v1
+; GFX9-NEXT:    v_or_b32_e32 v2, s6, v2
+; GFX9-NEXT:    v_or_b32_e32 v3, s7, v3
+; GFX9-NEXT:    v_or_b32_e32 v4, s4, v4
+; GFX9-NEXT:    v_or_b32_e32 v5, s5, v5
+; GFX9-NEXT:    v_or_b32_e32 v6, s6, v6
+; GFX9-NEXT:    v_or_b32_e32 v7, s7, v7
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: v_or_v2i128_inline_imm:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    s_mov_b64 s[4:5], 64
+; GFX8-NEXT:    s_mov_b64 s[6:7], 0
+; GFX8-NEXT:    s_mov_b64 s[4:5], s[4:5]
+; GFX8-NEXT:    s_mov_b64 s[6:7], s[6:7]
+; GFX8-NEXT:    v_or_b32_e32 v0, s4, v0
+; GFX8-NEXT:    v_or_b32_e32 v1, s5, v1
+; GFX8-NEXT:    v_or_b32_e32 v2, s6, v2
+; GFX8-NEXT:    v_or_b32_e32 v3, s7, v3
+; GFX8-NEXT:    v_or_b32_e32 v4, s4, v4
+; GFX8-NEXT:    v_or_b32_e32 v5, s5, v5
+; GFX8-NEXT:    v_or_b32_e32 v6, s6, v6
+; GFX8-NEXT:    v_or_b32_e32 v7, s7, v7
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+  %or = or <2 x i128> %a, <i128 64, i128 64>
+  ret <2 x i128> %or
+}
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/xor.v2i128.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/xor.v2i128.ll
new file mode 100644
index 0000000000000..291d27b0cf527
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/xor.v2i128.ll
@@ -0,0 +1,117 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -global-isel=true -mtriple=amdgcn -mcpu=hawaii < %s | FileCheck -check-prefix=GFX7 %s
+; RUN: llc -global-isel=true -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefix=GFX9 %s
+; RUN: llc -global-isel=true -mtriple=amdgcn -mcpu=fiji < %s | FileCheck -check-prefix=GFX8 %s
+; RUN: llc -global-isel=true -mtriple=amdgcn -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s
+; RUN: llc -global-isel=true -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck -check-prefix=GFX10 %s
+
+define <2 x i128> @v_xor_v2i128(<2 x i128> %a, <2 x i128> %b) {
+; GFX7-LABEL: v_xor_v2i128:
+; GFX7:       ; %bb.0:
+; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    v_xor_b32_e32 v0, v0, v8
+; GFX7-NEXT:    v_xor_b32_e32 v1, v1, v9
+; GFX7-NEXT:    v_xor_b32_e32 v2, v2, v10
+; GFX7-NEXT:    v_xor_b32_e32 v3, v3, v11
+; GFX7-NEXT:    v_xor_b32_e32 v4, v4, v12
+; GFX7-NEXT:    v_xor_b32_e32 v5, v5, v13
+; GFX7-NEXT:    v_xor_b32_e32 v6, v6, v14
+; GFX7-NEXT:    v_xor_b32_e32 v7, v7, v15
+; GFX7-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: v_xor_v2i128:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_xor_b32_e32 v0, v0, v8
+; GFX9-NEXT:    v_xor_b32_e32 v1, v1, v9
+; GFX9-NEXT:    v_xor_b32_e32 v2, v2, v10
+; GFX9-NEXT:    v_xor_b32_e32 v3, v3, v11
+; GFX9-NEXT:    v_xor_b32_e32 v4, v4, v12
+; GFX9-NEXT:    v_xor_b32_e32 v5, v5, v13
+; GFX9-NEXT:    v_xor_b32_e32 v6, v6, v14
+; GFX9-NEXT:    v_xor_b32_e32 v7, v7, v15
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: v_xor_v2i128:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    v_xor_b32_e32 v0, v0, v8
+; GFX8-NEXT:    v_xor_b32_e32 v1, v1, v9
+; GFX8-NEXT:    v_xor_b32_e32 v2, v2, v10
+; GFX8-NEXT:    v_xor_b32_e32 v3, v3, v11
+; GFX8-NEXT:    v_xor_b32_e32 v4, v4, v12
+; GFX8-NEXT:    v_xor_b32_e32 v5, v5, v13
+; GFX8-NEXT:    v_xor_b32_e32 v6, v6, v14
+; GFX8-NEXT:    v_xor_b32_e32 v7, v7, v15
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX10-LABEL: v_xor_v2i128:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    v_xor_b32_e32 v0, v0, v8
+; GFX10-NEXT:    v_xor_b32_e32 v1, v1, v9
+; GFX10-NEXT:    v_xor_b32_e32 v2, v2, v10
+; GFX10-NEXT:    v_xor_b32_e32 v3, v3, v11
+; GFX10-NEXT:    v_xor_b32_e32 v4, v4, v12
+; GFX10-NEXT:    v_xor_b32_e32 v5, v5, v13
+; GFX10-NEXT:    v_xor_b32_e32 v6, v6, v14
+; GFX10-NEXT:    v_xor_b32_e32 v7, v7, v15
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %xor = xor <2 x i128> %a, %b
+  ret <2 x i128> %xor
+}
+
+define <2 x i128> @v_xor_v2i128_inline_imm(<2 x i128> %a) {
+; GFX7-LABEL: v_xor_v2i128_inline_imm:
+; GFX7:       ; %bb.0:
+; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b64 s[4:5], 64
+; GFX7-NEXT:    s_mov_b64 s[6:7], 0
+; GFX7-NEXT:    s_mov_b64 s[4:5], s[4:5]
+; GFX7-NEXT:    s_mov_b64 s[6:7], s[6:7]
+; GFX7-NEXT:    v_xor_b32_e32 v0, s4, v0
+; GFX7-NEXT:    v_xor_b32_e32 v1, s5, v1
+; GFX7-NEXT:    v_xor_b32_e32 v2, s6, v2
+; GFX7-NEXT:    v_xor_b32_e32 v3, s7, v3
+; GFX7-NEXT:    v_xor_b32_e32 v4, s4, v4
+; GFX7-NEXT:    v_xor_b32_e32 v5, s5, v5
+; GFX7-NEXT:    v_xor_b32_e32 v6, s6, v6
+; GFX7-NEXT:    v_xor_b32_e32 v7, s7, v7
+; GFX7-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: v_xor_v2i128_inline_imm:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_mov_b64 s[4:5], 64
+; GFX9-NEXT:    s_mov_b64 s[6:7], 0
+; GFX9-NEXT:    s_mov_b64 s[4:5], s[4:5]
+; GFX9-NEXT:    s_mov_b64 s[6:7], s[6:7]
+; GFX9-NEXT:    v_xor_b32_e32 v0, s4, v0
+; GFX9-NEXT:    v_xor_b32_e32 v1, s5, v1
+; GFX9-NEXT:    v_xor_b32_e32 v2, s6, v2
+; GFX9-NEXT:    v_xor_b32_e32 v3, s7, v3
+; GFX9-NEXT:    v_xor_b32_e32 v4, s4, v4
+; GFX9-NEXT:    v_xor_b32_e32 v5, s5, v5
+; GFX9-NEXT:    v_xor_b32_e32 v6, s6, v6
+; GFX9-NEXT:    v_xor_b32_e32 v7, s7, v7
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: v_xor_v2i128_inline_imm:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    s_mov_b64 s[4:5], 64
+; GFX8-NEXT:    s_mov_b64 s[6:7], 0
+; GFX8-NEXT:    s_mov_b64 s[4:5], s[4:5]
+; GFX8-NEXT:    s_mov_b64 s[6:7], s[6:7]
+; GFX8-NEXT:    v_xor_b32_e32 v0, s4, v0
+; GFX8-NEXT:    v_xor_b32_e32 v1, s5, v1
+; GFX8-NEXT:    v_xor_b32_e32 v2, s6, v2
+; GFX8-NEXT:    v_xor_b32_e32 v3, s7, v3
+; GFX8-NEXT:    v_xor_b32_e32 v4, s4, v4
+; GFX8-NEXT:    v_xor_b32_e32 v5, s5, v5
+; GFX8-NEXT:    v_xor_b32_e32 v6, s6, v6
+; GFX8-NEXT:    v_xor_b32_e32 v7, s7, v7
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+  %xor = xor <2 x i128> %a, <i128 64, i128 64>
+  ret <2 x i128> %xor
+}

@github-actions GitHub Actions
Copy link

github-actions bot commented May 5, 2025 •
edited
Loading

✅ With the latest revision this PR passed the C/C++ code formatter.

chinmaydd added 5 commits May 6, 2025 16:36
@chinmaydd
[GlobalISel][AMDGPU] Fix handling of v2i128 type for AND, OR, XOR
af2fdcc 
Change-Id: I709d434e111f61e867c4fc284f1f4e768a083015
@chinmaydd
[GlobalISel][AMDGPU] Fix formatting
ca16db8 
Change-Id: If980ba1599f9eb805ae4ba0566d1e0c5459ff8ff
@chinmaydd
[GlobalISel][AMDGPU] Address comments
4027444 
Change-Id: Ia82e785f936dae63180b62a297b5cd2a1d1b8bf3
@chinmaydd
[GlobalISel][AMDGPU] Update tests
475de1d 
Change-Id: I28f0273da90b5caa9133762ab272ede95a7dd82e
@chinmaydd
[GlobalISel][AMDGPU] Update tests
73137e3 
Change-Id: I25afdc40b28be4201473923ff527062e9c76f19f
@chinmaydd chinmaydd force-pushed the chdeshpa/gisel-fix branch from a10d36d to 73137e3 Compare May 6, 2025 20:37
arsenm
arsenm reviewed May 7, 2025
View reviewed changes
llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
.legalFor({S32, S1, S64, V2S32, S16, V2S16, V4S16})
.clampScalar(0, S32, S64)
.moreElementsIf(isSmallOddVector(0), oneMoreElement(0))
.fewerElementsIf(vectorWiderThan(0, 64), fewerEltsToSize64Vector(0))
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't see why this was wrong. Why was the rule wrong? Is this just papering over a bug in LegalizerHelper

Copy link
Contributor Author

@chinmaydd chinmaydd May 7, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I dont believe so.

A 2xi128 vector is wider than 64 bits which satisfies the predicate here. Once the fewerElements action is chosen, the mutation at fewerEltsToSize64Vector chooses a vector of type 0xs128 to break it down. This eventually causes an FPE in LegalizerHelper::getNarrowTypeBreakdown.

I think fewerEltsToSize64Vector was not written with vectors of element types wider than 64 bits in mind.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@arsenm sorry for the ping. Thoughts ?

arsenm
arsenm approved these changes May 8, 2025
View reviewed changes
Copy link
Contributor

@arsenm arsenm left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We should do something with these clumsy rules at some point

@arsenm arsenm merged commit 3a5af23 into llvm:main May 8, 2025
11 checks passed
@chinmaydd chinmaydd deleted the chdeshpa/gisel-fix branch May 8, 2025 17:34
lenary added a commit to lenary/llvm-project that referenced this pull request May 9, 2025
@lenary
Merge branch 'main' into pr/riscv-tablegen-sdnodes
cc084d9 
* main: (420 commits)
  [AArch64] Merge scaled and unscaled narrow zero stores (llvm#136705)
  [RISCV] One last migration to getInsertSubvector [nfc]
  [flang][OpenMP] Update `do concurrent` mapping pass to use `fir.do_concurrent` op (llvm#138489)
  [MLIR][LLVM] Fix llvm.mlir.global mismatching print and parser order (llvm#138986)
  [lld][NFC] Fix minor typo in docs (llvm#138898)
  [RISCV] Migrate getConstant indexed insert/extract subvector to new API (llvm#139111)
  GlobalISel: Translate minimumnum and maximumnum (llvm#139106)
  [MemProf] Simplify unittest save and restore of options (llvm#139117)
  [BOLT][AArch64] Patch functions targeted by optional relocs (llvm#138750)
  [Coverage] Support -fprofile-list for cold function coverage (llvm#136333)
  Remove unused forward decl (llvm#139108)
  [AMDGPU][NFC] Get rid of OPW constants. (llvm#139074)
  [CIR] Upstream extract op for VectorType (llvm#138413)
  [mlir][xegpu] Handle scalar uniform ops in SIMT distribution.  (llvm#138593)
  [GlobalISel][AMDGPU] Fix handling of v2i128 type for AND, OR, XOR (llvm#138574)
  AMDGPU][True16][CodeGen] FP_Round f64 to f16 in true16 (llvm#128911)
  Reland [Clang] Deprecate `__is_trivially_relocatable` (llvm#139061)
  [HLSL][NFC] Stricter Overload Tests (clamp,max,min,pow) (llvm#138993)
  [MLIR] Fixing the memref linearization size computation for non-packed memref (llvm#138922)
  [TableGen][NFC] Use early exit to simplify large block in emitAction. (llvm#138220)
  ...
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@shiltian shiltian shiltian left review comments

@arsenm arsenm arsenm approved these changes

Assignees
No one assigned
Labels
backend:AMDGPU llvm:globalisel
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
4 participants
@chinmaydd @llvmbot @arsenm @shiltian