[llvm] [AMDGPU] Insert readfirstlane in the function returns in sgpr. (PR #135326)

[Matt Arsenault via llvm-commits]

================
@@ -0,0 +1,142 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 | FileCheck %s -check-prefixes=GFX11
+
+define amdgpu_ps i32 @uniform_v_to_s_i32(float inreg %a, float inreg %b) {
+; GFX11-LABEL: uniform_v_to_s_i32:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    v_max_f32_e64 v0, s0, s1
+; GFX11-NEXT:    v_cmp_o_f32_e64 vcc_lo, s0, s1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, 0x7fc00000, v0, vcc_lo
+; GFX11-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX11-NEXT:    ; return to shader part epilog
+  %max0 = call float @llvm.maximum.f32(float %a, float %b)
+  %cast = bitcast float %max0 to i32
+  ret i32 %cast
+}
+
+define amdgpu_ps i64 @uniform_v_to_s_i64(double inreg %a, double inreg %b) {
+; GFX11-LABEL: uniform_v_to_s_i64:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    v_max_f64 v[0:1], s[0:1], s[2:3]
+; GFX11-NEXT:    v_cmp_u_f64_e64 s0, s[0:1], s[2:3]
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_cndmask_b32_e64 v1, v1, 0x7ff80000, s0
+; GFX11-NEXT:    v_cndmask_b32_e64 v0, v0, 0, s0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_readfirstlane_b32 s1, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX11-NEXT:    ; return to shader part epilog
+  %max0 = call double @llvm.maximum.f64(double %a, double %b)
+  %cast = bitcast double %max0 to i64
+  ret i64 %cast
+}
+
+define amdgpu_ps <2 x i32> @uniform_v_to_s_2_i32(<2 x float> inreg %a, <2 x float> inreg %b) {
+; GFX11-LABEL: uniform_v_to_s_2_i32:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    v_max_f32_e64 v0, s0, s2
+; GFX11-NEXT:    v_cmp_o_f32_e64 vcc_lo, s0, s2
+; GFX11-NEXT:    v_max_f32_e64 v1, s1, s3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, 0x7fc00000, v0, vcc_lo
+; GFX11-NEXT:    v_cmp_o_f32_e64 vcc_lo, s1, s3
+; GFX11-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, 0x7fc00000, v1, vcc_lo
+; GFX11-NEXT:    v_readfirstlane_b32 s1, v1
+; GFX11-NEXT:    ; return to shader part epilog
+  %max0 = call <2 x float> @llvm.maximum.f32(<2 x float> %a, <2 x float> %b)
+  %cast = bitcast <2 x float> %max0 to <2 x i32>
+  ret <2 x i32> %cast
+}
+
+define amdgpu_ps ptr @uniform_v_to_s_ptr(ptr inreg %x) {
+; GFX11-LABEL: uniform_v_to_s_ptr:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    s_mov_b32 s1, 0
+; GFX11-NEXT:    flat_load_b32 v0, v[0:1]
+; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    v_max_f32_e32 v1, 1.0, v0
+; GFX11-NEXT:    v_cmp_o_f32_e32 vcc_lo, v0, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, 0x7fc00000, v1, vcc_lo
+; GFX11-NEXT:    v_cvt_u32_f32_e32 v0, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX11-NEXT:    ; return to shader part epilog
+  %val = load float, ptr %x, align 4
+  %max = call float @llvm.maximum.f32(float %val, float 1.0)
+  %int = fptoui float %max to i32
+  %ptr = inttoptr i32 %int to ptr
+  ret ptr %ptr
+}
+
+define amdgpu_ps double @uniform_v_to_s_double(double inreg %a, double inreg %b) {
+; GFX11-LABEL: uniform_v_to_s_double:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    v_max_f64 v[0:1], s[0:1], s[2:3]
+; GFX11-NEXT:    v_cmp_u_f64_e64 s0, s[0:1], s[2:3]
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_cndmask_b32_e64 v1, v1, 0x7ff80000, s0
+; GFX11-NEXT:    v_cndmask_b32_e64 v0, v0, 0, s0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    v_readfirstlane_b32 s1, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s0, v0
----------------
arsenm wrote:

Yes, but double should not be treated as a type returned in SGPRs

https://github.com/llvm/llvm-project/pull/135326