[llvm] 45ef27a - [AMDGPU] Handle True16 high-half 16-bit packing idiom with uniform operand (#211102)

via llvm-commits llvm-commits at lists.llvm.org
Wed Jul 29 08:42:51 PDT 2026


Author: Zach Goldthorpe
Date: 2026-07-29T09:42:46-06:00
New Revision: 45ef27a69895fcb756fec0d7a16d39706e5d9c06

URL: https://github.com/llvm/llvm-project/commit/45ef27a69895fcb756fec0d7a16d39706e5d9c06
DIFF: https://github.com/llvm/llvm-project/commit/45ef27a69895fcb756fec0d7a16d39706e5d9c06.diff

LOG: [AMDGPU] Handle True16 high-half 16-bit packing idiom with uniform operand (#211102)

With True16 instructions, the high-half packing idiom from #206058 tries
to `OR` a 16-bit VGPR operand into the high half of the other operand,
but this causes a crash if the other operand is uniform, as this
generates an invalid `$sgpr = COPY $sgpr_hi16`.

Added: 
    

Modified: 
    llvm/lib/Target/AMDGPU/SIInstructions.td
    llvm/test/CodeGen/AMDGPU/pack-half-to-lshl-or.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/AMDGPU/SIInstructions.td b/llvm/lib/Target/AMDGPU/SIInstructions.td
index 39df5dfd4d01d..d42e123e61eba 100644
--- a/llvm/lib/Target/AMDGPU/SIInstructions.td
+++ b/llvm/lib/Target/AMDGPU/SIInstructions.td
@@ -4100,9 +4100,9 @@ def : GCNPat <
     (i32 (bitconvert (v2i16 (build_vector (i16 0), (i16 VGPR_16:$src_hi))))),
     i32:$src2)),
   (REG_SEQUENCE VGPR_32,
-    (i16 (EXTRACT_SUBREG $src2, lo16)), lo16,
+    (i16 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS $src2, VGPR_32)), lo16)), lo16,
     (V_OR_B16_t16_e64 (i32 0), VGPR_16:$src_hi,
-                      (i32 0), (i16 (EXTRACT_SUBREG $src2, hi16))), hi16)
+                      (i32 0), (i16 (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS $src2, VGPR_32)), hi16))), hi16)
 >;
 
 // With multiple uses of the shift, this will duplicate the shift and

diff  --git a/llvm/test/CodeGen/AMDGPU/pack-half-to-lshl-or.ll b/llvm/test/CodeGen/AMDGPU/pack-half-to-lshl-or.ll
index 446c598d746c8..a57a42f56c117 100644
--- a/llvm/test/CodeGen/AMDGPU/pack-half-to-lshl-or.ll
+++ b/llvm/test/CodeGen/AMDGPU/pack-half-to-lshl-or.ll
@@ -120,3 +120,176 @@ define void @pack_high_half_to_store(i16 %hi, i32 %lo, ptr addrspace(1) %out) {
   store i32 %or, ptr addrspace(1) %out
   ret void
 }
+
+define i32 @pack_high_half_sgpr(i16 %hi, i32 inreg %lo) {
+; GFX8-LABEL: pack_high_half_sgpr:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GFX8-NEXT:    v_or_b32_e32 v0, s16, v0
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: pack_high_half_sgpr:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_lshl_or_b32 v0, v0, 16, s16
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX11-FAKE16-LABEL: pack_high_half_sgpr:
+; GFX11-FAKE16:       ; %bb.0:
+; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v0, 16, s0
+; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX11-TRUE16-LABEL: pack_high_half_sgpr:
+; GFX11-TRUE16:       ; %bb.0:
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_or_b16 v1.h, v0.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v1
+; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX12-FAKE16-LABEL: pack_high_half_sgpr:
+; GFX12-FAKE16:       ; %bb.0:
+; GFX12-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-FAKE16-NEXT:    s_wait_expcnt 0x0
+; GFX12-FAKE16-NEXT:    s_wait_samplecnt 0x0
+; GFX12-FAKE16-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
+; GFX12-FAKE16-NEXT:    v_lshl_or_b32 v0, v0, 16, s0
+; GFX12-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX12-TRUE16-LABEL: pack_high_half_sgpr:
+; GFX12-TRUE16:       ; %bb.0:
+; GFX12-TRUE16-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-TRUE16-NEXT:    s_wait_expcnt 0x0
+; GFX12-TRUE16-NEXT:    s_wait_samplecnt 0x0
+; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
+; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_or_b16 v1.h, v0.l, v1.h
+; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v0, v1
+; GFX12-TRUE16-NEXT:    s_setpc_b64 s[30:31]
+  %ext = zext i16 %hi to i32
+  %shl = shl i32 %ext, 16
+  %or = or i32 %shl, %lo
+  ret i32 %or
+}
+
+define i32 @pack_high_half_sgpr.half(half %hi, i32 inreg %lo) {
+; GFX8-LABEL: pack_high_half_sgpr.half:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GFX8-NEXT:    v_or_b32_e32 v0, s16, v0
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: pack_high_half_sgpr.half:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_lshl_or_b32 v0, v0, 16, s16
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX11-FAKE16-LABEL: pack_high_half_sgpr.half:
+; GFX11-FAKE16:       ; %bb.0:
+; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v0, 16, s0
+; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX11-TRUE16-LABEL: pack_high_half_sgpr.half:
+; GFX11-TRUE16:       ; %bb.0:
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_or_b16 v1.h, v0.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v1
+; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX12-FAKE16-LABEL: pack_high_half_sgpr.half:
+; GFX12-FAKE16:       ; %bb.0:
+; GFX12-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-FAKE16-NEXT:    s_wait_expcnt 0x0
+; GFX12-FAKE16-NEXT:    s_wait_samplecnt 0x0
+; GFX12-FAKE16-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
+; GFX12-FAKE16-NEXT:    v_lshl_or_b32 v0, v0, 16, s0
+; GFX12-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX12-TRUE16-LABEL: pack_high_half_sgpr.half:
+; GFX12-TRUE16:       ; %bb.0:
+; GFX12-TRUE16-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-TRUE16-NEXT:    s_wait_expcnt 0x0
+; GFX12-TRUE16-NEXT:    s_wait_samplecnt 0x0
+; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
+; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_or_b16 v1.h, v0.l, v1.h
+; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v0, v1
+; GFX12-TRUE16-NEXT:    s_setpc_b64 s[30:31]
+  %bc = bitcast half %hi to i16
+  %ext = zext i16 %bc to i32
+  %shl = shl i32 %ext, 16
+  %or = or i32 %shl, %lo
+  ret i32 %or
+}
+
+define i32 @pack_high_half_sgpr.bfloat(bfloat %hi, i32 inreg %lo) {
+; GFX8-LABEL: pack_high_half_sgpr.bfloat:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
+; GFX8-NEXT:    v_or_b32_e32 v0, s16, v0
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: pack_high_half_sgpr.bfloat:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_lshl_or_b32 v0, v0, 16, s16
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX11-FAKE16-LABEL: pack_high_half_sgpr.bfloat:
+; GFX11-FAKE16:       ; %bb.0:
+; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v0, 16, s0
+; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX11-TRUE16-LABEL: pack_high_half_sgpr.bfloat:
+; GFX11-TRUE16:       ; %bb.0:
+; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT:    v_or_b16 v1.h, v0.l, v1.h
+; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v0, v1
+; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX12-FAKE16-LABEL: pack_high_half_sgpr.bfloat:
+; GFX12-FAKE16:       ; %bb.0:
+; GFX12-FAKE16-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-FAKE16-NEXT:    s_wait_expcnt 0x0
+; GFX12-FAKE16-NEXT:    s_wait_samplecnt 0x0
+; GFX12-FAKE16-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-FAKE16-NEXT:    s_wait_kmcnt 0x0
+; GFX12-FAKE16-NEXT:    v_lshl_or_b32 v0, v0, 16, s0
+; GFX12-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX12-TRUE16-LABEL: pack_high_half_sgpr.bfloat:
+; GFX12-TRUE16:       ; %bb.0:
+; GFX12-TRUE16-NEXT:    s_wait_loadcnt_dscnt 0x0
+; GFX12-TRUE16-NEXT:    s_wait_expcnt 0x0
+; GFX12-TRUE16-NEXT:    s_wait_samplecnt 0x0
+; GFX12-TRUE16-NEXT:    s_wait_bvhcnt 0x0
+; GFX12-TRUE16-NEXT:    s_wait_kmcnt 0x0
+; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v1, s0
+; GFX12-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT:    v_or_b16 v1.h, v0.l, v1.h
+; GFX12-TRUE16-NEXT:    v_mov_b32_e32 v0, v1
+; GFX12-TRUE16-NEXT:    s_setpc_b64 s[30:31]
+  %bc = bitcast bfloat %hi to i16
+  %ext = zext i16 %bc to i32
+  %shl = shl i32 %ext, 16
+  %or = or i32 %shl, %lo
+  ret i32 %or
+}


        


More information about the llvm-commits mailing list