[llvm] [AArch64][GlobalISel] BF16 libcalls operations. (PR #200740)
via llvm-commits
llvm-commits at lists.llvm.org
Mon Jun 1 00:09:30 PDT 2026
llvmorg-github-actions[bot] wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-llvm-globalisel
@llvm/pr-subscribers-backend-aarch64
Author: David Green (davemgreen)
<details>
<summary>Changes</summary>
This fills in a number of bf16 instructions that lower through f32 libcall.
---
Patch is 931.77 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/200740.diff
5 Files Affected:
- (modified) llvm/include/llvm/CodeGen/GlobalISel/LegalizerInfo.h (+6)
- (modified) llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp (+7-14)
- (modified) llvm/test/CodeGen/AArch64/bf16-instructions.ll (+788-394)
- (modified) llvm/test/CodeGen/AArch64/bf16-v4-instructions.ll (+3959-1922)
- (modified) llvm/test/CodeGen/AArch64/bf16-v8-instructions.ll (+8946-3683)
``````````diff
diff --git a/llvm/include/llvm/CodeGen/GlobalISel/LegalizerInfo.h b/llvm/include/llvm/CodeGen/GlobalISel/LegalizerInfo.h
index 1d6a6cfb15a50..92ecd74f68482 100644
--- a/llvm/include/llvm/CodeGen/GlobalISel/LegalizerInfo.h
+++ b/llvm/include/llvm/CodeGen/GlobalISel/LegalizerInfo.h
@@ -854,6 +854,12 @@ class LegalizeRuleSet {
markAllIdxsAsCovered();
return actionIf(LegalizeAction::WidenScalar, Predicate, Mutation);
}
+ /// Widen the scalar, specified in mutation, when type index 0 is any type in
+ /// the given list.
+ LegalizeRuleSet &widenScalarFor(std::initializer_list<LLT> Types,
+ LegalizeMutation Mutation) {
+ return actionFor(LegalizeAction::WidenScalar, Types, Mutation);
+ }
/// Widen the scalar, specified in mutation, when type indexes 0 and 1 is any
/// type pair in the given list.
LegalizeRuleSet &
diff --git a/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp b/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp
index 6092f63fcde72..517efec6f83e3 100644
--- a/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp
+++ b/llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp
@@ -477,24 +477,17 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
.moreElementsToNextPow2(0)
.lowerFor({f16, bf16, v4f16, v4bf16, v8f16, v8bf16});
- getActionDefinitionsBuilder(G_FREM)
- .libcallFor({f32, f64, f128})
- .minScalar(0, f32)
- .scalarize(0);
-
- getActionDefinitionsBuilder({G_FCOS, G_FSIN, G_FPOW, G_FLOG, G_FLOG2,
+ getActionDefinitionsBuilder({G_FREM, G_FCOS, G_FSIN, G_FPOW, G_FLOG, G_FLOG2,
G_FLOG10, G_FTAN, G_FEXP, G_FEXP2, G_FEXP10,
G_FACOS, G_FASIN, G_FATAN, G_FATAN2, G_FCOSH,
G_FSINH, G_FTANH, G_FMODF})
- // We need a call for these, so we always need to scalarize.
- .scalarize(0)
- // Regardless of FP16 support, widen 16-bit elements to 32-bits.
- .minScalar(0, f32)
- .libcallFor({f32, f64, f128});
+ .libcallFor({f32, f64, f128})
+ .widenScalarFor({f16, bf16}, changeElementTo(0, f32))
+ .scalarize(0);
getActionDefinitionsBuilder({G_FPOWI, G_FLDEXP})
- .scalarize(0)
- .minScalar(0, f32)
- .libcallFor({{f32, i32}, {f64, i32}, {f128, i32}});
+ .libcallFor({{f32, i32}, {f64, i32}, {f128, i32}})
+ .widenScalarFor({f16, bf16}, changeElementTo(0, f32))
+ .scalarize(0);
getActionDefinitionsBuilder({G_LROUND, G_INTRINSIC_LRINT})
.legalFor({{i32, f32}, {i32, f64}, {i64, f32}, {i64, f64}})
diff --git a/llvm/test/CodeGen/AArch64/bf16-instructions.ll b/llvm/test/CodeGen/AArch64/bf16-instructions.ll
index 7e982d1142e8f..2ddfdeed09f85 100644
--- a/llvm/test/CodeGen/AArch64/bf16-instructions.ll
+++ b/llvm/test/CodeGen/AArch64/bf16-instructions.ll
@@ -4,8 +4,7 @@
; RUN: llc < %s -mtriple aarch64 -mattr=-bf16 -global-isel -global-isel-abort=2 2>&1 | FileCheck %s --check-prefixes=CHECK,CHECK-CVT,CHECK-CVT-GI
; RUN: llc < %s -mtriple aarch64 -mattr=+bf16 -global-isel -global-isel-abort=2 2>&1 | FileCheck %s --check-prefixes=CHECK,CHECK-BF16,CHECK-BF16-GI
-; CHECK-CVT-GI: warning: Instruction selection used fallback path for test_frem
-; CHECK-CVT-GI-NEXT: warning: Instruction selection used fallback path for test_fptosi_i32
+; CHECK-CVT-GI: warning: Instruction selection used fallback path for test_fptosi_i32
; CHECK-CVT-GI-NEXT: warning: Instruction selection used fallback path for test_fptosi_i64
; CHECK-CVT-GI-NEXT: warning: Instruction selection used fallback path for test_fptoui_i32
; CHECK-CVT-GI-NEXT: warning: Instruction selection used fallback path for test_fptoui_i64
@@ -17,26 +16,8 @@
; CHECK-CVT-GI-NEXT: warning: Instruction selection used fallback path for test_sitofp_i32_fadd
; CHECK-CVT-GI-NEXT: warning: Instruction selection used fallback path for test_fpext_double
; CHECK-CVT-GI-NEXT: warning: Instruction selection used fallback path for test_fpext_double_fast
-; CHECK-CVT-GI-NEXT: warning: Instruction selection used fallback path for test_powi
-; CHECK-CVT-GI-NEXT: warning: Instruction selection used fallback path for test_sin
-; CHECK-CVT-GI-NEXT: warning: Instruction selection used fallback path for test_cos
-; CHECK-CVT-GI-NEXT: warning: Instruction selection used fallback path for test_tan
-; CHECK-CVT-GI-NEXT: warning: Instruction selection used fallback path for test_acos
-; CHECK-CVT-GI-NEXT: warning: Instruction selection used fallback path for test_asin
-; CHECK-CVT-GI-NEXT: warning: Instruction selection used fallback path for test_atan
-; CHECK-CVT-GI-NEXT: warning: Instruction selection used fallback path for test_atan2
-; CHECK-CVT-GI-NEXT: warning: Instruction selection used fallback path for test_cosh
-; CHECK-CVT-GI-NEXT: warning: Instruction selection used fallback path for test_sinh
-; CHECK-CVT-GI-NEXT: warning: Instruction selection used fallback path for test_tanh
-; CHECK-CVT-GI-NEXT: warning: Instruction selection used fallback path for test_pow
-; CHECK-CVT-GI-NEXT: warning: Instruction selection used fallback path for test_exp
-; CHECK-CVT-GI-NEXT: warning: Instruction selection used fallback path for test_exp2
-; CHECK-CVT-GI-NEXT: warning: Instruction selection used fallback path for test_log
-; CHECK-CVT-GI-NEXT: warning: Instruction selection used fallback path for test_log10
-; CHECK-CVT-GI-NEXT: warning: Instruction selection used fallback path for test_log2
-;
-; CHECK-BF16-GI: warning: Instruction selection used fallback path for test_frem
-; CHECK-BF16-GI-NEXT: warning: Instruction selection used fallback path for test_fptosi_i32
+;
+; CHECK-BF16-GI: warning: Instruction selection used fallback path for test_fptosi_i32
; CHECK-BF16-GI-NEXT: warning: Instruction selection used fallback path for test_fptosi_i64
; CHECK-BF16-GI-NEXT: warning: Instruction selection used fallback path for test_fptoui_i32
; CHECK-BF16-GI-NEXT: warning: Instruction selection used fallback path for test_fptoui_i64
@@ -48,23 +29,7 @@
; CHECK-BF16-GI-NEXT: warning: Instruction selection used fallback path for test_sitofp_i32_fadd
; CHECK-BF16-GI-NEXT: warning: Instruction selection used fallback path for test_fpext_double
; CHECK-BF16-GI-NEXT: warning: Instruction selection used fallback path for test_fpext_double_fast
-; CHECK-BF16-GI-NEXT: warning: Instruction selection used fallback path for test_powi
-; CHECK-BF16-GI-NEXT: warning: Instruction selection used fallback path for test_sin
-; CHECK-BF16-GI-NEXT: warning: Instruction selection used fallback path for test_cos
-; CHECK-BF16-GI-NEXT: warning: Instruction selection used fallback path for test_tan
-; CHECK-BF16-GI-NEXT: warning: Instruction selection used fallback path for test_acos
-; CHECK-BF16-GI-NEXT: warning: Instruction selection used fallback path for test_asin
-; CHECK-BF16-GI-NEXT: warning: Instruction selection used fallback path for test_atan
-; CHECK-BF16-GI-NEXT: warning: Instruction selection used fallback path for test_atan2
-; CHECK-BF16-GI-NEXT: warning: Instruction selection used fallback path for test_cosh
-; CHECK-BF16-GI-NEXT: warning: Instruction selection used fallback path for test_sinh
-; CHECK-BF16-GI-NEXT: warning: Instruction selection used fallback path for test_tanh
-; CHECK-BF16-GI-NEXT: warning: Instruction selection used fallback path for test_pow
-; CHECK-BF16-GI-NEXT: warning: Instruction selection used fallback path for test_exp
-; CHECK-BF16-GI-NEXT: warning: Instruction selection used fallback path for test_exp2
-; CHECK-BF16-GI-NEXT: warning: Instruction selection used fallback path for test_log
-; CHECK-BF16-GI-NEXT: warning: Instruction selection used fallback path for test_log10
-; CHECK-BF16-GI-NEXT: warning: Instruction selection used fallback path for test_log2
+;
define bfloat @test_fadd(bfloat %a, bfloat %b) #0 {
; CHECK-CVT-SD-LABEL: test_fadd:
@@ -464,40 +429,78 @@ define bfloat @test_fdiv(bfloat %a, bfloat %b) #0 {
}
define bfloat @test_frem(bfloat %a, bfloat %b) #0 {
-; CHECK-CVT-LABEL: test_frem:
-; CHECK-CVT: // %bb.0:
-; CHECK-CVT-NEXT: str x30, [sp, #-16]! // 8-byte Folded Spill
-; CHECK-CVT-NEXT: // kill: def $h1 killed $h1 def $d1
-; CHECK-CVT-NEXT: // kill: def $h0 killed $h0 def $d0
-; CHECK-CVT-NEXT: shll v0.4s, v0.4h, #16
-; CHECK-CVT-NEXT: shll v1.4s, v1.4h, #16
-; CHECK-CVT-NEXT: // kill: def $s0 killed $s0 killed $q0
-; CHECK-CVT-NEXT: // kill: def $s1 killed $s1 killed $q1
-; CHECK-CVT-NEXT: bl fmodf
-; CHECK-CVT-NEXT: fmov w9, s0
-; CHECK-CVT-NEXT: mov w8, #32767 // =0x7fff
-; CHECK-CVT-NEXT: ubfx w10, w9, #16, #1
-; CHECK-CVT-NEXT: add w8, w9, w8
-; CHECK-CVT-NEXT: add w8, w10, w8
-; CHECK-CVT-NEXT: lsr w8, w8, #16
-; CHECK-CVT-NEXT: fmov s0, w8
-; CHECK-CVT-NEXT: // kill: def $h0 killed $h0 killed $s0
-; CHECK-CVT-NEXT: ldr x30, [sp], #16 // 8-byte Folded Reload
-; CHECK-CVT-NEXT: ret
+; CHECK-CVT-SD-LABEL: test_frem:
+; CHECK-CVT-SD: // %bb.0:
+; CHECK-CVT-SD-NEXT: str x30, [sp, #-16]! // 8-byte Folded Spill
+; CHECK-CVT-SD-NEXT: // kill: def $h1 killed $h1 def $d1
+; CHECK-CVT-SD-NEXT: // kill: def $h0 killed $h0 def $d0
+; CHECK-CVT-SD-NEXT: shll v0.4s, v0.4h, #16
+; CHECK-CVT-SD-NEXT: shll v1.4s, v1.4h, #16
+; CHECK-CVT-SD-NEXT: // kill: def $s0 killed $s0 killed $q0
+; CHECK-CVT-SD-NEXT: // kill: def $s1 killed $s1 killed $q1
+; CHECK-CVT-SD-NEXT: bl fmodf
+; CHECK-CVT-SD-NEXT: fmov w9, s0
+; CHECK-CVT-SD-NEXT: mov w8, #32767 // =0x7fff
+; CHECK-CVT-SD-NEXT: ubfx w10, w9, #16, #1
+; CHECK-CVT-SD-NEXT: add w8, w9, w8
+; CHECK-CVT-SD-NEXT: add w8, w10, w8
+; CHECK-CVT-SD-NEXT: lsr w8, w8, #16
+; CHECK-CVT-SD-NEXT: fmov s0, w8
+; CHECK-CVT-SD-NEXT: // kill: def $h0 killed $h0 killed $s0
+; CHECK-CVT-SD-NEXT: ldr x30, [sp], #16 // 8-byte Folded Reload
+; CHECK-CVT-SD-NEXT: ret
;
-; CHECK-BF16-LABEL: test_frem:
-; CHECK-BF16: // %bb.0:
-; CHECK-BF16-NEXT: str x30, [sp, #-16]! // 8-byte Folded Spill
-; CHECK-BF16-NEXT: // kill: def $h1 killed $h1 def $d1
-; CHECK-BF16-NEXT: // kill: def $h0 killed $h0 def $d0
-; CHECK-BF16-NEXT: shll v0.4s, v0.4h, #16
-; CHECK-BF16-NEXT: shll v1.4s, v1.4h, #16
-; CHECK-BF16-NEXT: // kill: def $s0 killed $s0 killed $q0
-; CHECK-BF16-NEXT: // kill: def $s1 killed $s1 killed $q1
-; CHECK-BF16-NEXT: bl fmodf
-; CHECK-BF16-NEXT: bfcvt h0, s0
-; CHECK-BF16-NEXT: ldr x30, [sp], #16 // 8-byte Folded Reload
-; CHECK-BF16-NEXT: ret
+; CHECK-BF16-SD-LABEL: test_frem:
+; CHECK-BF16-SD: // %bb.0:
+; CHECK-BF16-SD-NEXT: str x30, [sp, #-16]! // 8-byte Folded Spill
+; CHECK-BF16-SD-NEXT: // kill: def $h1 killed $h1 def $d1
+; CHECK-BF16-SD-NEXT: // kill: def $h0 killed $h0 def $d0
+; CHECK-BF16-SD-NEXT: shll v0.4s, v0.4h, #16
+; CHECK-BF16-SD-NEXT: shll v1.4s, v1.4h, #16
+; CHECK-BF16-SD-NEXT: // kill: def $s0 killed $s0 killed $q0
+; CHECK-BF16-SD-NEXT: // kill: def $s1 killed $s1 killed $q1
+; CHECK-BF16-SD-NEXT: bl fmodf
+; CHECK-BF16-SD-NEXT: bfcvt h0, s0
+; CHECK-BF16-SD-NEXT: ldr x30, [sp], #16 // 8-byte Folded Reload
+; CHECK-BF16-SD-NEXT: ret
+;
+; CHECK-CVT-GI-LABEL: test_frem:
+; CHECK-CVT-GI: // %bb.0:
+; CHECK-CVT-GI-NEXT: str x30, [sp, #-16]! // 8-byte Folded Spill
+; CHECK-CVT-GI-NEXT: // kill: def $h0 killed $h0 def $d0
+; CHECK-CVT-GI-NEXT: // kill: def $h1 killed $h1 def $d1
+; CHECK-CVT-GI-NEXT: shll v0.4s, v0.4h, #16
+; CHECK-CVT-GI-NEXT: shll v1.4s, v1.4h, #16
+; CHECK-CVT-GI-NEXT: // kill: def $s0 killed $s0 killed $q0
+; CHECK-CVT-GI-NEXT: // kill: def $s1 killed $s1 killed $q1
+; CHECK-CVT-GI-NEXT: bl fmodf
+; CHECK-CVT-GI-NEXT: fmov w9, s0
+; CHECK-CVT-GI-NEXT: mov w8, #32767 // =0x7fff
+; CHECK-CVT-GI-NEXT: fcmp s0, #0.0
+; CHECK-CVT-GI-NEXT: ubfx w10, w9, #16, #1
+; CHECK-CVT-GI-NEXT: add w8, w9, w8
+; CHECK-CVT-GI-NEXT: orr w9, w9, #0x400000
+; CHECK-CVT-GI-NEXT: add w8, w8, w10
+; CHECK-CVT-GI-NEXT: csel w8, w9, w8, vs
+; CHECK-CVT-GI-NEXT: lsr w8, w8, #16
+; CHECK-CVT-GI-NEXT: fmov s0, w8
+; CHECK-CVT-GI-NEXT: // kill: def $h0 killed $h0 killed $s0
+; CHECK-CVT-GI-NEXT: ldr x30, [sp], #16 // 8-byte Folded Reload
+; CHECK-CVT-GI-NEXT: ret
+;
+; CHECK-BF16-GI-LABEL: test_frem:
+; CHECK-BF16-GI: // %bb.0:
+; CHECK-BF16-GI-NEXT: str x30, [sp, #-16]! // 8-byte Folded Spill
+; CHECK-BF16-GI-NEXT: // kill: def $h0 killed $h0 def $d0
+; CHECK-BF16-GI-NEXT: // kill: def $h1 killed $h1 def $d1
+; CHECK-BF16-GI-NEXT: shll v0.4s, v0.4h, #16
+; CHECK-BF16-GI-NEXT: shll v1.4s, v1.4h, #16
+; CHECK-BF16-GI-NEXT: // kill: def $s0 killed $s0 killed $q0
+; CHECK-BF16-GI-NEXT: // kill: def $s1 killed $s1 killed $q1
+; CHECK-BF16-GI-NEXT: bl fmodf
+; CHECK-BF16-GI-NEXT: bfcvt h0, s0
+; CHECK-BF16-GI-NEXT: ldr x30, [sp], #16 // 8-byte Folded Reload
+; CHECK-BF16-GI-NEXT: ret
%r = frem bfloat %a, %b
ret bfloat %r
}
@@ -2118,23 +2121,23 @@ define bfloat @test_sqrt(bfloat %a) #0 {
}
define bfloat @test_powi(bfloat %a, i32 %b) #0 {
-; CHECK-CVT-LABEL: test_powi:
-; CHECK-CVT: // %bb.0:
-; CHECK-CVT-NEXT: str x30, [sp, #-16]! // 8-byte Folded Spill
-; CHECK-CVT-NEXT: // kill: def $h0 killed $h0 def $d0
-; CHECK-CVT-NEXT: shll v0.4s, v0.4h, #16
-; CHECK-CVT-NEXT: // kill: def $s0 killed $s0 killed $q0
-; CHECK-CVT-NEXT: bl __powisf2
-; CHECK-CVT-NEXT: fmov w9, s0
-; CHECK-CVT-NEXT: mov w8, #32767 // =0x7fff
-; CHECK-CVT-NEXT: ubfx w10, w9, #16, #1
-; CHECK-CVT-NEXT: add w8, w9, w8
-; CHECK-CVT-NEXT: add w8, w10, w8
-; CHECK-CVT-NEXT: lsr w8, w8, #16
-; CHECK-CVT-NEXT: fmov s0, w8
-; CHECK-CVT-NEXT: // kill: def $h0 killed $h0 killed $s0
-; CHECK-CVT-NEXT: ldr x30, [sp], #16 // 8-byte Folded Reload
-; CHECK-CVT-NEXT: ret
+; CHECK-CVT-SD-LABEL: test_powi:
+; CHECK-CVT-SD: // %bb.0:
+; CHECK-CVT-SD-NEXT: str x30, [sp, #-16]! // 8-byte Folded Spill
+; CHECK-CVT-SD-NEXT: // kill: def $h0 killed $h0 def $d0
+; CHECK-CVT-SD-NEXT: shll v0.4s, v0.4h, #16
+; CHECK-CVT-SD-NEXT: // kill: def $s0 killed $s0 killed $q0
+; CHECK-CVT-SD-NEXT: bl __powisf2
+; CHECK-CVT-SD-NEXT: fmov w9, s0
+; CHECK-CVT-SD-NEXT: mov w8, #32767 // =0x7fff
+; CHECK-CVT-SD-NEXT: ubfx w10, w9, #16, #1
+; CHECK-CVT-SD-NEXT: add w8, w9, w8
+; CHECK-CVT-SD-NEXT: add w8, w10, w8
+; CHECK-CVT-SD-NEXT: lsr w8, w8, #16
+; CHECK-CVT-SD-NEXT: fmov s0, w8
+; CHECK-CVT-SD-NEXT: // kill: def $h0 killed $h0 killed $s0
+; CHECK-CVT-SD-NEXT: ldr x30, [sp], #16 // 8-byte Folded Reload
+; CHECK-CVT-SD-NEXT: ret
;
; CHECK-BF16-LABEL: test_powi:
; CHECK-BF16: // %bb.0:
@@ -2146,28 +2149,49 @@ define bfloat @test_powi(bfloat %a, i32 %b) #0 {
; CHECK-BF16-NEXT: bfcvt h0, s0
; CHECK-BF16-NEXT: ldr x30, [sp], #16 // 8-byte Folded Reload
; CHECK-BF16-NEXT: ret
+;
+; CHECK-CVT-GI-LABEL: test_powi:
+; CHECK-CVT-GI: // %bb.0:
+; CHECK-CVT-GI-NEXT: str x30, [sp, #-16]! // 8-byte Folded Spill
+; CHECK-CVT-GI-NEXT: // kill: def $h0 killed $h0 def $d0
+; CHECK-CVT-GI-NEXT: shll v0.4s, v0.4h, #16
+; CHECK-CVT-GI-NEXT: // kill: def $s0 killed $s0 killed $q0
+; CHECK-CVT-GI-NEXT: bl __powisf2
+; CHECK-CVT-GI-NEXT: fmov w9, s0
+; CHECK-CVT-GI-NEXT: mov w8, #32767 // =0x7fff
+; CHECK-CVT-GI-NEXT: fcmp s0, #0.0
+; CHECK-CVT-GI-NEXT: ubfx w10, w9, #16, #1
+; CHECK-CVT-GI-NEXT: add w8, w9, w8
+; CHECK-CVT-GI-NEXT: orr w9, w9, #0x400000
+; CHECK-CVT-GI-NEXT: add w8, w8, w10
+; CHECK-CVT-GI-NEXT: csel w8, w9, w8, vs
+; CHECK-CVT-GI-NEXT: lsr w8, w8, #16
+; CHECK-CVT-GI-NEXT: fmov s0, w8
+; CHECK-CVT-GI-NEXT: // kill: def $h0 killed $h0 killed $s0
+; CHECK-CVT-GI-NEXT: ldr x30, [sp], #16 // 8-byte Folded Reload
+; CHECK-CVT-GI-NEXT: ret
%r = call bfloat @llvm.powi.bf16.i32(bfloat %a, i32 %b)
ret bfloat %r
}
define bfloat @test_sin(bfloat %a) #0 {
-; CHECK-CVT-LABEL: test_sin:
-; CHECK-CVT: // %bb.0:
-; CHECK-CVT-NEXT: str x30, [sp, #-16]! // 8-byte Folded Spill
-; CHECK-CVT-NEXT: // kill: def $h0 killed $h0 def $d0
-; CHECK-CVT-NEXT: shll v0.4s, v0.4h, #16
-; CHECK-CVT-NEXT: // kill: def $s0 killed $s0 killed $q0
-; CHECK-CVT-NEXT: bl sinf
-; CHECK-CVT-NEXT: fmov w9, s0
-; CHECK-CVT-NEXT: mov w8, #32767 // =0x7fff
-; CHECK-CVT-NEXT: ubfx w10, w9, #16, #1
-; CHECK-CVT-NEXT: add w8, w9, w8
-; CHECK-CVT-NEXT: add w8, w10, w8
-; CHECK-CVT-NEXT: lsr w8, w8, #16
-; CHECK-CVT-NEXT: fmov s0, w8
-; CHECK-CVT-NEXT: // kill: def $h0 killed $h0 killed $s0
-; CHECK-CVT-NEXT: ldr x30, [sp], #16 // 8-byte Folded Reload
-; CHECK-CVT-NEXT: ret
+; CHECK-CVT-SD-LABEL: test_sin:
+; CHECK-CVT-SD: // %bb.0:
+; CHECK-CVT-SD-NEXT: str x30, [sp, #-16]! // 8-byte Folded Spill
+; CHECK-CVT-SD-NEXT: // kill: def $h0 killed $h0 def $d0
+; CHECK-CVT-SD-NEXT: shll v0.4s, v0.4h, #16
+; CHECK-CVT-SD-NEXT: // kill: def $s0 killed $s0 killed $q0
+; CHECK-CVT-SD-NEXT: bl sinf
+; CHECK-CVT-SD-NEXT: fmov w9, s0
+; CHECK-CVT-SD-NEXT: mov w8, #32767 // =0x7fff
+; CHECK-CVT-SD-NEXT: ubfx w10, w9, #16, #1
+; CHECK-CVT-SD-NEXT: add w8, w9, w8
+; CHECK-CVT-SD-NEXT: add w8, w10, w8
+; CHECK-CVT-SD-NEXT: lsr w8, w8, #16
+; CHECK-CVT-SD-NEXT: fmov s0, w8
+; CHECK-CVT-SD-NEXT: // kill: def $h0 killed $h0 killed $s0
+; CHECK-CVT-SD-NEXT: ldr x30, [sp], #16 // 8-byte Folded Reload
+; CHECK-CVT-SD-NEXT: ret
;
; CHECK-BF16-LABEL: test_sin:
; CHECK-BF16: // %bb.0:
@@ -2179,28 +2203,49 @@ define bfloat @test_sin(bfloat %a) #0 {
; CHECK-BF16-NEXT: bfcvt h0, s0
; CHECK-BF16-NEXT: ldr x30, [sp], #16 // 8-byte Folded Reload
; CHECK-BF16-NEXT: ret
+;
+; CHECK-CVT-GI-LABEL: test_sin:
+; CHECK-CVT-GI: // %bb.0:
+; CHECK-CVT-GI-NEXT: str x30, [sp, #-16]! // 8-byte Folded Spill
+; CHECK-CVT-GI-NEXT: // kill: def $h0 killed $h0 def $d0
+; CHECK-CVT-GI-NEXT: shll v0.4s, v0.4h, #16
+; CHECK-CVT-GI-NEXT: // kill: def $s0 killed $s0 killed $q0
+; CHECK-CVT-GI-NEXT: bl sinf
+; CHECK-CVT-GI-NEXT: fmov w9, s0
+; CHECK-CVT-GI-NEXT: mov w8, #32767 // =0x7fff
+; CHECK-CVT-GI-NEXT: fcmp s0, #0.0
+; CHECK-CVT-GI-NEXT: ubfx w10, w9, #16, #1
+; CHECK-CVT-GI-NEXT: add w8, w9, w8
+; CHECK-CVT-GI-NEXT: orr w9, w9, #0x400000
+; CHECK-CVT-GI-NEXT: add w8, w8, w10
+; CHECK-CVT-GI-NEXT: csel w8, w9, w8, vs
+; CHECK-CVT-GI-NEXT: lsr w8, w8, #16
+; CHECK-CVT-GI-NEXT: fmov s0, w8
+; CHECK-CVT-GI-NEXT: // kill: def $h0 killed $h0 killed $s0
+; CHECK-CVT-GI-NEXT: ldr x30, [sp], #16 // 8-byte Folded Reload
+; CHECK-CVT-GI-NEXT: ret
%r = call bfloat @llvm.sin.bf16(bfloat %a)
ret bfloat %r
}
define bfloat @test_cos(bfloat %a) #0 {
-; CHECK-CVT-LABEL: test_cos:
-; CHECK-CVT: // %bb.0:
-; CHECK-CVT-NEXT: str x30, [sp, #-16]! // 8-byte Folded Spill
-; CHECK-CVT-NEXT: // kill: def $h0 killed $h0 def $d0
-; CHECK-CVT-NEXT: shll v0.4s, v0.4h, #16
-; CHECK-CVT-NEXT: // kill: def $s0 killed $s0 killed $q0
-; CHECK-CVT-NEXT: bl cosf
-; CHECK-CVT-NEXT: fmov w9, s0
-; CHECK-CVT-NEXT: mov w8, #32767 // =0x7fff
-; CHECK-CVT-NEXT: ubfx w10, w9, #16, #1
-; CHECK-CVT-NEXT: add w8, w9, w8
-; CHECK-CVT-NEXT: add w8, w10, w8
-; CHECK-CVT-NEXT: lsr w8, w8, #16
-; CHECK-CVT-NEXT: fmov s0, w8
-; CHECK-CVT-NEXT: // kill: def $h0 killed $h0 killed $s0
-; CHECK-CVT-NEXT: ldr x30, [sp], #16 // 8-byte Folded Reload
-; CHECK-CVT-NEXT: ret
+; CHECK-CVT-SD-LABEL: test_cos:
+; CHECK-CVT-SD: // %bb.0:
+; CHECK-CVT-SD-NEXT: str x30, [sp, #-16]! // 8-byte Folded Spill
+; CHECK-CVT-SD-NEXT: // kill: def $h0 killed $h0 def $d0
+; CHECK-CVT-SD-NEXT: shll v0.4s, v0.4h, #16
+; CHECK-CVT-SD-NEXT: // kill: def $s0 killed $s0 killed $q0
+; CHECK-CVT-SD-NEXT: bl cosf
+; CHECK-CVT-SD-NEXT: fmov w9, s0
+; CHECK-CVT-SD-NEXT: mov w8, #32767 // =0x7fff
+; CHECK-CVT-SD-NEXT: ubfx w10, w9, #16, #1
+; CHECK-CVT-SD-NEXT: add w8, w9, w8
+; CHECK-CVT-SD-NEXT: add w8, w10, w8
+; CHECK-CVT-SD-NEXT: lsr w8, w8, #16
+; CHECK-CVT-SD-NEXT: fmov s0, w8
+; CHECK-CVT-SD-NEXT: // kill: def $h0 killed $h0 killed $s0
+; CHECK-CVT-SD-NEXT: ldr x30, [sp], #16 // 8-byte Folded Reload
+; ...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/200740
More information about the llvm-commits
mailing list