[llvm] 6ddb3a6 - [AMDGPU] Add another test showing unwanted VALU codegen (#145062)

[via llvm-commits]

Author: Jay Foad
Date: 2025-06-20T17:54:44+01:00
New Revision: 6ddb3a69c15963b7cf1cbb0323dd3e467b71cefc

URL: https://github.com/llvm/llvm-project/commit/6ddb3a69c15963b7cf1cbb0323dd3e467b71cefc
DIFF: https://github.com/llvm/llvm-project/commit/6ddb3a69c15963b7cf1cbb0323dd3e467b71cefc.diff

LOG: [AMDGPU] Add another test showing unwanted VALU codegen (#145062)

Added: 
    

Modified: 
    llvm/test/CodeGen/AMDGPU/pseudo-scalar-transcendental.ll

Removed: 
    


################################################################################
diff  --git a/llvm/test/CodeGen/AMDGPU/pseudo-scalar-transcendental.ll b/llvm/test/CodeGen/AMDGPU/pseudo-scalar-transcendental.ll
index a2e7f2e62f5dd..d957aaa8a48ad 100644
--- a/llvm/test/CodeGen/AMDGPU/pseudo-scalar-transcendental.ll
+++ b/llvm/test/CodeGen/AMDGPU/pseudo-scalar-transcendental.ll
@@ -410,19 +410,40 @@ define amdgpu_cs half @srcmods_neg_f16(half inreg %src) {
   ret half %result
 }
 
-declare half @llvm.exp2.f16(half)
-declare float @llvm.exp2.f32(float)
-declare half @llvm.amdgcn.exp2.f16(half)
-declare float @llvm.amdgcn.exp2.f32(float)
-declare half @llvm.log2.f16(half)
-declare float @llvm.log2.f32(float)
-declare half @llvm.amdgcn.log.f16(half)
-declare float @llvm.amdgcn.log.f32(float)
-declare half @llvm.amdgcn.rcp.f16(half)
-declare float @llvm.amdgcn.rcp.f32(float)
-declare half @llvm.sqrt.f16(half)
-declare float @llvm.sqrt.f32(float)
-declare half @llvm.amdgcn.sqrt.f16(half)
-declare float @llvm.amdgcn.sqrt.f32(float)
-declare half @llvm.fabs.f16(half)
-declare float @llvm.fabs.f32(float)
+; TODO: SelectionDAG should avoid generating v_rcp_iflag_f32.
+define amdgpu_cs float @fdiv_f32_i32(float inreg %a, i32 inreg %b) {
+; GFX12-SDAG-LABEL: fdiv_f32_i32:
+; GFX12-SDAG:       ; %bb.0:
+; GFX12-SDAG-NEXT:    s_cvt_f32_u32 s1, s1
+; GFX12-SDAG-NEXT:    s_delay_alu instid0(SALU_CYCLE_3) | instskip(NEXT) | instid1(TRANS32_DEP_1)
+; GFX12-SDAG-NEXT:    v_rcp_iflag_f32_e32 v0, s1
+; GFX12-SDAG-NEXT:    v_mul_f32_e32 v0, s0, v0
+; GFX12-SDAG-NEXT:    ; return to shader part epilog
+;
+; GFX12-GISEL-LABEL: fdiv_f32_i32:
+; GFX12-GISEL:       ; %bb.0:
+; GFX12-GISEL-NEXT:    s_cvt_f32_u32 s1, s1
+; GFX12-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_3) | instskip(NEXT) | instid1(TRANS32_DEP_1)
+; GFX12-GISEL-NEXT:    v_s_rcp_f32 s1, s1
+; GFX12-GISEL-NEXT:    s_mul_f32 s0, s0, s1
+; GFX12-GISEL-NEXT:    s_wait_alu 0xfffe
+; GFX12-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_2)
+; GFX12-GISEL-NEXT:    v_mov_b32_e32 v0, s0
+; GFX12-GISEL-NEXT:    ; return to shader part epilog
+  %uint = uitofp i32 %b to float
+  %result = fdiv afn float %a, %uint
+  ret float %result
+}
+
+define amdgpu_cs half @fdiv_f16_i16(half inreg %a, i16 inreg %b) {
+; GFX12-LABEL: fdiv_f16_i16:
+; GFX12:       ; %bb.0:
+; GFX12-NEXT:    v_cvt_f16_u16_e32 v0, s1
+; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(TRANS32_DEP_1)
+; GFX12-NEXT:    v_rcp_f16_e32 v0, v0
+; GFX12-NEXT:    v_mul_f16_e32 v0, s0, v0
+; GFX12-NEXT:    ; return to shader part epilog
+  %uint = uitofp i16 %b to half
+  %result = fdiv afn half %a, %uint
+  ret half %result
+}