[all-commits] [llvm/llvm-project] 50b859: [AMDGPU] Stop coercing structs with FP and int fie...

[Addmisol via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 50b859cca1ccf7d174ee61a8a130ae14220209e4
      https://github.com/llvm/llvm-project/commit/50b859cca1ccf7d174ee61a8a130ae14220209e4
  Author: Addmisol <addmisol9 at gmail.com>
  Date:   2026-04-17 (Fri, 17 Apr 2026)

  Changed paths:
    M clang/lib/CodeGen/Targets/AMDGPU.cpp
    A clang/test/CodeGen/amdgpu-abi-struct-coerce.c
    M clang/test/CodeGen/amdgpu-variadic-call.c
    M clang/test/CodeGenOpenCL/amdgpu-abi-struct-coerce.cl
    M clang/test/Headers/amdgcn-openmp-device-math-complex.c

  Log Message:
  -----------
  [AMDGPU] Stop coercing structs with FP and int fields to integer arrays (#185083)

Fixes #184150

This PR fixes the ABI lowering code for small aggregates (≤64 bits) on
AMDGPU targets to selectively coerce based on element types:

- Structs containing only sub-32-bit integers (char, short): Continue to
coerce to i16/i32/[2 x i32] for efficient register packing
- Structs containing floats or full-sized integers (i32, i64, float,
double): Preserve original types using ABIArgInfo::getDirect() without
coercion

Previously, ALL small aggregates were unconditionally coerced to integer
types. A struct like { float, int } would be lowered to [2 x i32],
losing the floating-point type information. This prevented attaching
FP-specific attributes like nofpclass to the float
  component.

  Changes

- clang/lib/CodeGen/Targets/AMDGPU.cpp: Added
containsOnlyPackableIntegerTypes() helper function that recursively
checks if an aggregate contains only sub-32-bit integer types. Updated
classifyReturnType and classifyArgumentType to use this helper - only
coercing
aggregates that contain exclusively small integers, while preserving
types for aggregates containing floats or full-sized integers.
- clang/test/CodeGenOpenCL/amdgpu-abi-struct-coerce.cl: Updated expected
output to reflect that char-only structs are still coerced (e.g.,
struct_char_x8 -> [2 x i32]) while preserving correct behavior.
- clang/test/CodeGen/amdgpu-abi-struct-coerce.c: Added test coverage for
various struct types including mixed float/int fields, demonstrating the
selective coercion behavior.

  Before/After

  // Struct with float - NOW preserves types
  typedef struct { float f; int i; } fp_int_pair;

  Before: define [2 x i32] @ foo([2 x i32] %x.coerce)
After: define %struct.fp_int_pair @ foo(float %x.coerce0, i32
%x.coerce1)

  // Struct with only small integers - STILL coerced for efficiency
  typedef struct { char a, b, c, d, e, f, g, h; } eight_chars;

  Before: define [2 x i32] @ bar([2 x i32] %x.coerce)
  After:  define [2 x i32] @ bar([2 x i32] %x.coerce)  // Unchanged

  Test Plan

  - Updated existing ABI tests in amdgpu-abi-struct-coerce.cl
  - Added new test amdgpu-abi-struct-coerce.c for mixed FP/int structs
  - Updated affected OpenMP complex math header tests



To unsubscribe from these emails, change your notification settings at https://github.com/llvm/llvm-project/settings/notifications